Aplicación de imágenes de satélites y datos LiDAR en la modelización de recursos forestales

TESIS DOCTORAL Aplicación de imágenes de satélites y datos LiDAR en la modelización de recursos forestales Forest resource modelling combining satellite imagery and LiDAR data Autora Ángela Blázquez Casado Directores Dr. José Miguel Olano Mendoza Dr. José Ramón González Olabarría Dr. Francisco Rodríguez Puerta Enero 2019 PROGRAMA DE DOCTORADO EN CONSERVACIÓN Y USO SOTENIBLE DE SISTEMAS FORESTALES TESIS DOCTORAL Aplicación de imágenes de satélites y datos LiDAR en la modelización de recursos forestales Forest resource modelling combining satellite imagery and LiDAR data Presentada por Ángela Blázquez Casado para optar al grado de Doctora con Mención Doctorado Industrial por la Universidad de Valladolid Dirigida por: Dr. José Miguel Olano Mendoza Dr. José Ramón González Olabarría Dr. Francisco Rodríguez Puerta “La educación es el arma más poderosa que puedes usar para cambiar el mundo” Nelson Mandela A mi familia 6 INDICE Nota a los lectores .......................................................................................................................................................................... 9 Resumen .....................................................................................................................................................................................................11 Abstract ..................................................................................................................................................................................................... 13 Capítulo 1. Introducción ......................................................................................................................................................... 18 1.1 La importancia de los recursos forestales .......................................................................................... 18 1.2 Uso de imágenes de satélite en la evaluación de recursos forestales ......... 20 1.3 La tecnología LiDAR en la evaluación de recursos forestales ................................... 24 1.4 Integración de imágenes de satélite y datos LiDAR como herramientas en la evaluación de recursos forestales ............................................................................................................. 27 1.5 La modelización forestal integrando información procedente de sensores remotos .............................................................................................................................................................................................. 31 1.6 Aproximación a distintas escalas ............................................................................................................. 32 1.7 Objetivos y alcance .................................................................................................................................................... 34 1.8 Marco de estudio .......................................................................................................................................................... 34 Chapter 2. Combining low density LiDAR and satellite images to discriminate species in mixed Mediterranean forest ............................................................................................................ 42 Abstract ........................................................................................................................................................................................... 42 2.1 Introduction.......................................................................................................................................................................... 43 2.2 Material and methods ........................................................................................................................................... 46 2.2.1 Study area ................................................................................................................................................................ 46 2.2.2 Field data ................................................................................................................................................................. 48 2.2.3 Remote sensing data ................................................................................................................................ 49 2.2.4 Remotely sensed data processing ............................................................................................ 49 2.2.4.1 LiDAR data processing .................................................................................................................. 49 2.2.4.2 Tree crown selection ..................................................................................................................... 53 2.2.4.3 Plèiades data processing ....................................................................................................... 53 2.2.5 Classification technique ........................................................................................................................ 54 2.3 Results ....................................................................................................................................................................................... 55 2.3.1 Training model selection ......................................................................................................................... 55 2.3.2 Validation model ............................................................................................................................................ 57 2.4 Discussion ............................................................................................................................................................................ 58 2.5 Conclusion ........................................................................................................................................................................... 62 2.6 References ........................................................................................................................................................................... 63 7 Chapter 3. Stand types discrimination comparing machine learning algorithms in monteverde, Canary Islands ................................................................................................. 72 Abstract ............................................................................................................................................................................................ 72 3.1 Introduction .......................................................................................................................................................................... 73 3.2 Material and methods ............................................................................................................................................74 3.2.1 Study area .................................................................................................................................................................74 3.2.2 Field data ..................................................................................................................................................................74 3.2.3 LiDAR data ................................................................................................................................................................ 75 3.2.4 Multispectral imagery data source ........................................................................................... 75 3.2.5 Segmentation process .............................................................................................................................76 3.2.6 Data analysis .......................................................................................................................................................76 3.3 Results and discussion ........................................................................................................................................... 77 3.4 References ............................................................................................................................................................................ 81 Chapter 4. Fire and burn severity assessment: calibration of Relative Differenced Normalized Burn Ratio (RdNBR) with field data................................................... 86 Abstract ........................................................................................................................................................................................... 86 4.1 Introduction ......................................................................................................................................................................... 87 4.2 Material and Methods ............................................................................................................................................ 89 4.2.1 Data Sources ........................................................................................................................................................ 89 4.2.1.1 Studied fires ................................................................................................................................................ 89 4.2.1.2 Field data ..................................................................................................................................................... 90 4.2.2 Satellite image collation and processing and RdNBR calculation ........ 93 4.2.3 Methodological approaches............................................................................................................ 93 4.3 Results ...................................................................................................................................................................................... 95 4.4 Discussion .......................................................................................................................................................................... 100 4.5 References ........................................................................................................................................................................ 104 Capítulo 5. Discusión general, Futuras líneas de trabajo y Conclusiones ............ 112 5.1 Discusión general ......................................................................................................................................................... 112 5.2 Futuras líneas de trabajo .................................................................................................................................... 117 5.3 Conclusiones .................................................................................................................................................................... 119 Agradecimientos ........................................................................................................................................................................124 Referencias ........................................................................................................................................................................................ 128 14 models show an accuracy around 90%, and are very similar, so none of them is postulated as clearly superior to the others, although small differences between them might reveal relevant when applied over large surfaces. In chapter 4, post-fire severity was assessed by calibrating the spectral index RdNBR (Relative Differenced Normalized Burn Ratio) with variables measured in 28 large fires across three southern European countries. The index provides the ability to assess, through variables that reflect intensity (mean bark char height) and severity (loss of trees and shrub cover), the impact induced by fire on forest ecosystems. The results can be incorporated into biomass loss estimations when predictive models, with 67% of accuracy, are applied across large and heterogeneous regions. The importance of the time between the fire and its assessment is reflected in the model formulations, due to vegetation response variations after the fire. Mixed models highlight significant variations between fires due to specific contexts that modify these relationships. The use of BRT (Boosted Regression Trees) shows that relationships between severity measures and spectral information may not be linear. CAPÍTULO 1 Introducción “Caminante, son tus huellas el camino y nada más; Caminante, no hay camino, se hace camino al andar” Antonio Machado Capítulo 1 18 Capítulo 1. Introducción 1.1 La importancia de los recursos forestales A lo largo de la historia del ser humano, los bosques han tenido una enorme importancia. Tanto la madera, como muchos otros productos forestales han supuesto la base de la economía y el bienestar de nuestros antepasados. Actualmente, la actividad forestal es fuente de alimentación, medicinas y combustible para más de un billón de personas, contribuyendo al desarrollo económico en las zonas rurales, especialmente en las zonas más desfavorecidas (FAO 2018). En este sentido, los bosques y el sector forestal juegan cada vez un papel más importante en la transición hacia una economía innovadora y eficiente en el consumo de recursos (Wong and Prokofieva 2014). Esto es, una economía basada en el conocimiento y la utilización de recursos ecológicos, procesos y métodos biológicos para proporcionar bienes y servicios de forma sostenible en todos los sectores económicos (bioeconomy) (Carus 2017). El concepto de servicios ecosistémicos se refiere a los bienes y servicios que los ecosistemas proveen a los seres humanos. Estos servicios son múltiples. Por ejemplo, los servicios de apoyo como el reciclado de los nutrientes; los servicios de aprovisionamiento referidos a la cantidad de bienes o materias primas que un ecosistema ofrece como la madera, el agua o los alimentos; los servicios de regulación, como la protección frente a inundaciones, la depuración de las aguas, la polinización de los cultivos y los servicios culturales que proporcionan beneficios recreativos, estéticos y espirituales (Hassan et al. 2005). Tradicionalmente son los servicios de aprovisionamiento los que se han cuantificado a través de un beneficio económico directo, por ejemplo, la madera u otros recursos no maderables como el piñón o la resina (Calama et al. 2010). Sin embargo, en los últimos tiempos ha aumentado la conciencia de otros servicios como los servicios culturales y de soporte como la biodiversidad (Rego et al. 2014). La persistencia de estos servicios está amenazada por los cambios de usos, el cambio climático y, en definitiva, los diferentes componentes del cambio global (Thom and Seidl 2016). Introducción 19 De acuerdo con el último informe del estado de los bosques en Europa (FAO 2015), el valor de mercado total de los productos forestales no maderables se estima en 2277 millones de euros para toda la región europea, de los cuales el 73.0% fueron generados por productos derivados directamente de las plantas, aunque los datos disponibles son difícilmente comparables debido a la falta de homogeneidad en la recolección de estos por los propios países y a la cantidad variable de países que los recogen estos datos anualmente. En el anterior informe publicado en 2011 (FAO 2011) participaron cinco países más en la recolección de dichos datos y se estimó que el valor de mercado total de los productos forestales no maderables fue durante el anterior periodo de 2763 millones de euros, un 17.6% menos que en el informe actual. Sin embargo, el porcentaje de productos derivados directamente de plantas fue del 76.6%, solamente un 3.6% más que en el mercado actual. Por tanto, puede considerarse que el mercado de los productos forestales no maderables ha sido estable en el tiempo a pesar de la crisis sufrida en los mercados europeos. En zonas donde el crecimiento de los árboles está más limitado por factores climáticos como, por ejemplo, en la Cuenca del Mediterráneo los aprovechamientos de los productos forestales no maderables (corcho, setas, trufas, piñón, etc.) tienen un valor superior al de la madera (Vázquez-Piqué and Pereira 2008; Palahí et al. 2009; Calama et al. 2010). De hecho, en Europa el sector de los productos forestales alimenticios (setas, castaña, piñón, etc.) está liderado por Italia, España y Turquía (44600, 42100 y 22700 toneladas respectivamente). Si bien, en términos económicos el patrón es ligeramente diferente, encabezando los datos de valor económico España, Italia y Portugal con 196, 88 y 55 millones de euros respectivamente (FAO 2015). En los ecosistemas mediterráneos, los incendios son una de las principales amenazas que se ciernen sobre sus servicios ecosistémicos (Seidl et al. 2015). La superficie forestal media afectada por incendios en los países miembros de la Unión Europea fue de 410776 ha en los últimos 5 años (2013-2017), un 0.23% de la superficie forestal total. El impacto de los incendios se concentra de modo desproporcionado en los países del Suroeste de Europa (España, Italia y Portugal), que con un 23% de la superficie forestal, tuvieron el 83% de la superficie quemada (JRC-EFFIS 2017). Los datos de superficie quemada pueden ser engañosos, ya que la cantidad de superficie afectada por un incendio no indica, por sí sola, ni la gravedad del daño, ni las pérdidas económicas asociadas. En Capítulo 1 20 este sentido, la inclusión de la severidad de un incendio permite estimar de un modo más certero la pérdida de recursos. Esto es muy relevante, ya que la cuantificación de los daños reales es clave para evaluar adecuadamente las medidas políticas y prácticas de manejo forestal sostenible relacionadas con la prevención y lucha contra el fuego. La gestión de los montes o, en su caso, la ausencia de gestión tiene un impacto crítico sobre la disponibilidad de los servicios ecosistémicos (Asta et al. 2002; Arozena et al. 2015). Sin embargo, asegurar la persistencia de la masa es siempre el objetivo principal en la gestión forestal, ejecutando aprovechamientos que serán la base de la financiación del coste de las mejoras necesarias para mantener la multifuncionalidad de nuestros bosques. Así, la gestión forestal actual y futura debe de ir encaminada hacia una optimización sostenible y multifuncional de los recursos y servicios ecosistémicos forestales de los bosques (Wong and Prokofieva 2014). 1.2 Uso de imágenes de satélite en la evaluación de recursos forestales La teledetección se define como una técnica de adquisición de datos de la superficie terrestre desde la distancia usando la señal reflejada desde la superficie terrestre en una o más regiones del espectro electromagnético (Campbell and Wynne 2011). Los inicios de esta disciplina se remontan al año 1858, cuando se tomaron las primeras imágenes desde el aire mediante globo aerostático en Francia (Campbell and Wynne 2011). Ya en 1909 se tomaron las primeras imágenes desde avión, para poco después convertirse en una herramienta de uso habitual para la observación de la superficie durante la Primera Guerra Mundial (Fig. 1). Introducción 21 Fig. 1: Fotograma tomado sobre la ciudad de Soria el 22/08/1946. Escala del vuelo: 40,890. Fuente: Fototeca Digital CNIG – Vuelo AMS-46/47. Ministerio de Defensa (CECAF) procedente de los archivos del Ejército del Aire. Nombre del fichero: H0350_294_050.ecw. Los satélites son un paso más allá en la observación de la superficie terrestre. El primer satélite, Sputnik I, fue lanzado por la Unión Soviética en 1957 y fue el detonante de una carrera tecnológica de naturaleza política, militar, tecnológica y científica, marcando el inicio de la era espacial (Garber 2007). Esta tecnología, que en su origen era de naturaleza militar y de uso muy restringido, ha ido popularizándose para usos civiles. De hecho, en la actualidad la teledetección se usa de forma habitual en muy diversas disciplinas, siendo un instrumento indispensable para la observación y el estudio de los ecosistemas y sus recursos, así como los cambios que éstos están sufriendo en el nuevo contexto de cambio global (Tielidze 2015; Urban et al. 2018). A lo largo de las últimas décadas la disponibilidad de imágenes satelitales ha aumentado notablemente. Las imágenes procedentes de los satélites de las diversas misiones Landsat han sido las más utilizadas, a pesar de su resolución espacial media (30 metros en el multiespectral), por la disponibilidad de su amplio repositorio (lanzado en 1972), buena resolución espectral y su alta resolución temporal (16 días de revista). Un ejemplo de ello es la creación de mapas de cobertura forestal en Canadá (Wulder et al. 2008) o la evaluación de ¯ Capítulo 1 22 la dinámica que experimentan los bosques europeos durante los más de 40 años que lleva este satélite registrando imágenes (Potapov et al. 2015). Otras misiones posteriores, como SPOT o Plèiades fueron lanzados siendo uno de sus objetivos la disminución de la resolución espacial, llegando hasta 2.5 y 0.5 metros respectivamente (Fig. 2). Un ejemplo de su uso lo encontramos en el trabajo de Meng et al. (2016) en el que se estudia la diversidad estructural de los bosques de la región autónoma de Guangxi Zhuang en China mediante información procedente del satélite SPOT 5. Sus resultados indicaron que variables como el área basimétrica pueden predecirse de manera fiable utilizando variables espectrales o de textura procedentes de estas imágenes. Otro ejemplo es el trabajo de Akbari and Kalbi (2017), donde se ha examinado la capacidad real de las imágenes de muy alta resolución Plèiades para determinar diversidad de especies en bosques mixtos del norte de Irán. Si bien es indudable que estos satélites con muy alta resolución espacial aportan gran precisión, su elevado coste y su baja resolución espectral limitan su utilización. Fig. 2: Imagen multiespectral de muy alta resolución tomada por el satélite Plèiades. En este contexto, los satélites Sentinel 2A y 2B, en órbita desde junio de 2015 y marzo de 2017, respectivamente, han sido diseñados expresamente para monitorizar la superficie de la tierra y sus cambios. Estos satélites presentan alta resolución espectral (13 bandas) y resolución espacial media (hasta 10 metros). 0 0.8 1.6 2.4 km ¯ Introducción 23 Sin embargo, el potencial de esta misión se basa en su trabajo en conjunto para proveer de una cobertura global de la superficie terrestre cada 5 días. Actualmente, se está evaluando la potencialidad de Sentinel para el monitoreo de la cobertura terrestre. Por ejemplo, en el ámbito agrícola, Immitzer et al. (2016) generaron una clasificación para el mapeo de seis tipos de cultivos de verano a partir de imágenes de Sentinel. Li et al. (2018) construyeron índices de vegetación basados en las bandas del visible, borde del rojo e infrarrojo cercano de Sentinel para evaluar el contenido de clorofila en las copas de los árboles de cultivos agroforestales con el objetivo de detectar la capacidad fotosintética y estadios de estrés en el desarrollo de las plantas. Otro de los usos potenciales en el ámbito forestal es el estudio de perturbaciones (Hirschmugl et al. 2017). Concretamente, los daños causados por incendios (Fig. 3) y por tormentas de viento y nieve son algunos de los más habituales en los bosques templados de Europa (FAO 2015). Mejorar el conocimiento del estado actual de los bosques y, en concreto, de estas masas donde existe riesgo o donde ya han sufrido una perturbación pudiéndose cuantificar la pérdida de recursos forestales tanto maderables como no maderables, es objeto de evaluación. Fig. 3: Progresión en áreas forestales quemadas en la República del Congo con imágenes de Sentinel (Verhegghen et al. 2016). De cara a la gestión forestal, el volumen de existencias es una de las variables biofísicas más comunes en las que se basa la planificación de una masa forestal. La exploración de las posibilidades que ofrecen, en este sentido, las imágenes multiespectrales procedentes de satélites también han sido Capítulo 1 30 ecosistema ante el efecto del cambio climático y de las actividades antropogénicas mediante el uso de imágenes Landsat multitemporales y datos LiDAR. Sin embargo, también existen algunos trabajos donde se emplean datos multitemporales tanto LiDAR como satelitales para evaluar dinámicas o cambios sufridos en un periodo de tiempo determinado. Por ejemplo, Wulder et al. (2009) demostraron que la estructura forestal después de un incendio y los cambios absolutos y relativos sufridos en la estructura del bosque están fuertemente relacionados, siendo los cambios absolutos y relativos en la estructura de la vegetación y en la cobertura los más relacionados con los distintos índices espectrales de severidad de incendio. Así, conociendo las condiciones de desarrollo de las masas forestales, es posible predecir diferentes escenarios futuros, lo que podría servir como base de las directrices de gestión. Fig. 8: Ejemplo de cambio en una masa forestal con información LiDAR multitemporal. Fuentes de datos: Plano Nacional de Observación del Territorio (PNOT), Plan Nacional de Ortofotografía Aérea (PNOA). Introducción 31 1.5 La modelización forestal integrando información procedente de sensores remotos La modelización es la abstracción o simplificación de fenómenos que se dan en la naturaleza con el objetivo de predecir y/o explicar dicho fenómeno (Vanclay 1994). La clave de la correcta gestión sobre las masas forestales radica en el conocimiento expreso del desarrollo de las especies forestales y de su comportamiento frente a posibles cambios (Diéguez-Aranda et al. 2009). En este sentido, la generación de modelos de crecimiento y producción de los sistemas forestales ha sido una de las ramas más desarrollados en investigación forestal desde finales del siglo XVIII, aunque como destaca Tesch (1980), el estudio del crecimiento forestal se remonta a la época de Aristóteles. Este tipo de modelos han evolucionado desde tablas que predicen variables tradicionales como la altura (Assman 1970), a modelos cuyo objetivo es la predicción de la fijación del carbono, tanto a nivel de árbol como de masa. Sin embargo, la creciente necesidad de caracterizar cambios en el espacio a gran escala ha requerido un cambio en el paradigma de la modelización, siendo posible el análisis a gran escala de estos procesos utilizando modelos controlados por sensores remotos. Existen numerosos métodos de modelización capaces de estimar variables forestales de masa como la altura, el área basimétrica, el volumen y la biomasa en pie a partir de datos obtenidos por sensores remotos (García et al. 2018; Pearse et al. 2018). Hasta ahora, los modelos de regresión paramétricos han dominado la modelización forestal. Los modelos de regresión lineal múltiple, no lineales, lineales generalizados o mixtos son algunos de los más utilizados (Hudak et al. 2006; Popescu 2007; Sullivan et al. 2017). Sin embargo, el hecho de que las relaciones entre las variables predictoras y variable respuesta no sean siempre lineales, ni sigan una distribución normal, y la toma de datos masiva que actualmente se da mediante los sensores remotos, está haciendo que las técnicas de modelización basados en aprendizaje automático se presenten como la alternativa más eficiente en la modelización actual. El objetivo de las técnicas de modelización no paramétricas es imitar, mediante una estructura matemática, el funcionamiento de un cerebro biológico. Son estructuras capaces de aprender a resolver problemas a partir del conocimiento extraído de una muestra de entrenamiento. Los modelos no parmétricos más usados son los vecinos más cercanos (k-NN: K-Nearest Capítulo 1 32 Neightbour), los árboles aleatorios (RF: Random Forest) y la máquina de vector soporte (SVM: Support Vector Machine) (Xu et al. 2018). Además, existen otros modelos como las redes neuronales (ANN - Artificial Neural Networks) o los árboles de regresión iterados (BRT – Boosted Regression Trees), también aplicados en trabajos relacionados con la modelización en el sector medioambiental (Nitze et al. 2012; Youssef et al. 2016). Los modelos paramétricos siempre parten del supuesto de normalidad, aunque muchas de las variables implicadas en la modelización forestal no los cumplan. Pero la gran ventaja de los modelos paramétricos es que la relación entre los predictores y la variable respuesta se fija a través de los coeficientes, por lo que se asume robustez en la predicción cuando son validados. Esto implica que los modelos resultantes pueden ser aplicados sobre otros contextos posteriores cuando se cumplan características similares. En cambio, los modelos no paramétricos no están sujetos al supuesto de normalidad, ni tampoco se ven afectados por problemas de colinealidad. Sin embargo, el enfoque de caja negra sobre las relaciones entre las variables hace que este tipo de modelos sean difíciles de interpretar. Por tanto, ninguna técnica se revela como claramente superior a las demás para predecir atributos de los inventarios forestales y la mejor aproximación depende de los objetivos (Brosofske et al. 2014). 1.6 Aproximación a distintas escalas La recogida de información sobre la estructura forestal ha dependido históricamente del trabajo de campo, de alcance limitado cuando se trata de estudiar zonas extensas y/o remotas (Bergen and Dronova 2007). La integración del trabajo en campo con la información procedente de sensores remotos implica el cambio de escala, desde escala de monte o región a global. Para hacer efectivo este cambio de escala, existen varios sensores que aportan información espacial completa de toda la superficie global, pero con muy diversas resoluciones espaciales. La unidad mínima de trabajo es un factor a tener en cuenta. Estudios a nivel de árbol individual o a nivel de parcela o segmento, requieren información con distinta resolución espacial. Xu et al. (2017) haciendo una estimación de la Introducción 33 biomasa a nivel de árbol, encontraron que la varianza residual de las ecuaciones alométricas fueron las responsables de la mayor parte de la incertidumbre, mientras que la varianza de la altura LiDAR tuvo poca contribución en el error. Posteriormente, al integrar el error a nivel de árbol individual en los análisis a escala de parcela, se encontró que los errores se reducían aún más al incrementar la escala, estando concentrados en la estimación de biomasa. La relación a nivel de píxel o de parcela es conceptualmente similar a la relación entre árboles. La principal diferencia es que la relación se establece entre parámetros biofísicos y la variable objetivo de todos los árboles dentro del área de interés (Rodriguez-Veiga et al. 2017). Clark et al. (2005) observaron una disminución de la precisión en sus clasificaciones cuando aumentaba el tamaño del píxel de sus datos. En cambio, Ghosh et al. (2014) encontraron que sus clasificaciones globales a una resolución intermedia presentan mayor precisión que a una resolución menor o mayor, y que este efecto podría deberse a que el tamaño del píxel más pequeño recoge demasiada variabilidad espectral intraespecífica haciendo que no exista apenas separabilidad espectral entre especies. Y, por el contrario, en el caso del tamaño del píxel más grande, recoge escasa variabilidad haciendo, de igual manera, que no exista separabilidad espectral entre especies. Es importante que exista equilibrio entre la resolución espacial de la información utilizada y la unidad mínima de trabajo. También existe relación entre la precisión de los modelos y la escala de trabajo. Kimberley et al. (2017) estudiaron la caracterización del error de predicción en función de la escala en la productividad del Pinus radiata D. Don. en Nueva Zelanda y concluyeron que la precisión mejora gradualmente a medida que aumenta la escala. La modelización en los trabajos a pequeña escala generalmente aporta buena precisión. Del mismo modo, trabajos a gran escala, como la regional, aportan resultados de gran calidad, ya que sus errores se ven compensados. Capítulo 1 34 1.7 Objetivos y alcance Objetivo global En la presente tesis se trata de aportar nuevos conocimientos de la potencialidad de la teledetección combinando información continua procedente de imágenes satelitales y de datos LiDAR, para desarrollar herramientas capaces de evaluar distintos recursos forestales a gran escala. Objetivos específicos i. Combinación de datos LiDAR de baja densidad e imágenes de satélite de alta resolución Plèiades para discriminar especies pie a pie en masas mixtas de bosques mediterráneos. ii. Comparación de diferentes metodologías de modelización de aprendizaje automático RF, SVML, SVMR y ANN (Random Forest, Linear Support Vector Machine, Radial Support Vector Machine y Artificial Neural Networks) para la clasificación de las distintas tipologías de masas forestales presentes en los bosques de monteverde de la isla de Tenerife en Islas Canarias (España). iii. Evaluación de la severidad post-incendio a través del calibrado del índice espectral RdNBR (Relative Differenced Normalized Burn Ratio) con datos de campo con el objeto de ser capaces de cuantificar la pérdida del recurso tras una perturbación. 1.8 Marco de estudio El cuerpo de esta tesis se estructura en tres capítulos bien diferenciados (Fig. 9), en los que se aborda la evaluación de recursos forestales o, en su caso, la pérdida de estos a distintas escalas: (i) Clasificación de especies a nivel de árbol individual para la cuantificación de la producción de piñón y resina en masas mixtas de Pinus pinea L. y Pinus pinaster Ait.; (ii) Clasificación de las diferentes tipologías de masa a escala regional en bosques multifuncionales de monteverde; y (iii) Calibración de la severidad post-incendio en diferentes sucesos repartidos en el sur de Europa. En todos los casos, la información de Introducción 35 partida procede de diferentes sensores remotos, tanto a nivel estático como multitemporal. Además, también son utilizados diferentes métodos estadísticos para la evaluación de los recursos en cada uno de los trabajos. En el capítulo primero, se realizó un estudio para caracterizar con precisión, a nivel de árbol individual, el número de pies de cada especie presentes en bosques mixtos de Pinus pinea L. y Pinus pinaster Ait. de la meseta norte española. Los modelos fueron entrenados con masas puras, con el objetivo de cuantificar, con la mayor precisión posible, la producción de piña y resina en este tipo de masas mixtas. Para ello, se hizo una clasificación basada en datos LiDAR, a nivel de árbol individual, procedentes del vuelo PNOA 2010 de baja densidad e imágenes Plèiades de alta resolución espacial tomadas en el año 2014. En el segundo capítulo, se identificaron las diferentes tipologías de masa presentes en los bosques de monteverde con cobertura completa, elevada densidad y continuidad vertical de la isla de Tenerife. El objetivo fue caracterizar la diversidad de estos mediante una comparación de cuatro métodos de clasificación para identificar las diferentes tipologías de masa. Para ello, se tomó como base de datos de partida la información multitemporal, tanto LiDAR a nivel de masa (dos vuelos en 2009 y 2016) como espectral mediante imágenes de Sentinel (dos imágenes en los años 2015 y 2017) con resolución espacial media. En el tercer capítulo, se evaluó la capacidad de determinar la severidad tras el paso de un incendio. El estudio ha sido desarrollado a partir de datos tomados en 28 grandes incendios repartidos por tres países del sur de Europa con el objetivo de considerar la pérdida de biomasa. Para ello, se calibró el índice espectral RdNBR a través de datos de afectación tomados en campo mediante dos métodos estadísticos distintos, tomando como base de datos el repositorio multitemporal de imágenes Landsat con baja resolución espacial. Además, en esta tesis se han utilizado distintos métodos estadísticos, tanto modelos clásicos paramétricos como no paramétricos, para estimar las relaciones existentes entre los datos tomados en campo y las diferentes variables procedentes de los sensores remotos. En el capítulo primero fue Capítulo 1 36 utilizado Random Forest para realizar la clasificación de las dos especies objeto de estudio. En el segundo capítulo, se enfrentaron cuatro modelos no paramétricos como son Random Forest, Linear Support Vector Machine, Radial Support Vector Machine y Artificial Neural Networks para la clasificación de las diferentes tipologías de masa presentes en los bosques de monteverde. Por último, en el capítulo tercero se evaluó la capacidad de predecir la severidad de un incendio y, por tanto, el nivel de afectación en el recurso de biomasa mediante dos métodos estadísticos, uno lineal (modelos mixtos) y otro no lineal (Boosted Regression Trees, BRT). Introducción 37 RF –Random forest Capítulo 3 Capítulo 2 Capítulo 4 Satélite -Pléiades Datos de campo Resolución Espacial Alta Sensores Remotos Satélite –Sentinel 2 Fotografía aérea Datos de campo Datos de campo Resolución Espacial Media SVML –Linear Support Vector Machine SVMR –Radial Support Vector Machine Clasificación de Especies forestales Producción de piña y resina Análisis Estadísticos Clasificación de Tipologías de Masa Diversidad forestal Calibración de la severidad post-incendio Pérdida de los recursos disponibles Modelos Mixtos Recurso Forestal LiDAR PNOA Información multitemporal Información estática Satélite –Landsat Resolución Espacial Baja BRT-Boosted RegressionTrees RF –Random forest ANN –Artificial Neural Networks Información multitemporal Árbol Individual Método de Masa Fig. 9: Esquema general del cuerpo de la tesis. Chapter 2 46 L. forests, based on an automatic algorithm for crown delineation. The practical motivation is that by distinguishing these two species at a tree scale will permit us to apply the pre-existing models for different non-wood forest products (resin or pine nuts) which act at a tree level scale (e.g. Calama et al. 2016; Nanos et al. 2004). For this purpose, (i) in a first phase, a Random Forest model was trained with a dataset collected in pure stands of both species, to determine which LiDAR and satellite variables allow us to obtain the best discrimination between groups; and (ii) the Random Forest constructed was then validated by classifying individuals into an independent set of mixed and pure stands. Our hypotheses are: (i) combining LiDAR and Pleiades data to discriminate species in mixed Mediterranean conifer forests should be more efficient than the use of each technique separately; (ii) this combination should provide an efficient tool to discriminate pine species in mixed Mediterranean forests at individual tree level, and (iii) the results for Pinus pinea L. should be more accurate than those for Pinus pinaster Ait. due to the fact that the former usually grows in open stands. 2.2 Material and methods 2.2.1 Study area Forested land accounts for almost a third of the total land area of Spain (18.27 million ha). Pure forests of Pinus pinea L. cover 401,701 ha while Pinus pinaster Ait. forests make up 1,131,901 ha (MAGRAMA 2012). The study area is situated in the Northern Plateau of the Iberian Peninsula (Fig.1), a region with 65,275 ha of pure Pinus pinea L. forests, 271,000 ha of pure Pinus pinaster Ait. forests and 53,000 ha of mixed Pinus pinea L. and Pinus pinaster Ait. forests (López et al. 2009). The Duero river basin is on the Northern Plateau, 700-800 m a.s.l., with a homogenous Mediterranean continental climate, characterized by low average annual rainfall (450 mm), strong summer drought (<50 mm) and low average annual temperatures (12 ºC). Soils mainly present a very high sand content (> 90%) and very low water holding capacity (WHC < 100 mm), except in the upper areas where limestone soils show a percentage of clays and limes over 40-50%, reaching WHC values > 250 mm. The plots within the study area were clustered in three different forest types: pure stands of Pinus pinea L., pure stands of Pinus pinaster Ait., and mixed stands, where the least represented species should account for at least 35% of the total Species classification using low density LiDAR and satellite images 47 number of trees. All considered stands were at mature development stage, where regeneration is quite poor. Fig.1 Study area. Locations of training and validation sets. Chapter 2 48 2.2.2 Field data In this study we used a subset from the network of permanent plots installed by the Forest Research Centre of the Spanish National Institute for Agricultural Research (INIA-CIFOR) in the province of Valladolid. Mixed and pure Pinus pinaster Ait. plots were installed between summer of 2015 and spring of 2016. Pure even aged Pinus pinea L. plots were installed in 1996 in cooperation with the regional forest service and have been regularly monitored since then (the most recent inventory was carried out in 2016). Plots were selected in order to cover the whole range of site conditions, stand stocking and age identified within the region. The plots were all located in public mature forests, and at installation, the stand should have not been altered (thinnings, final cuttings or pruning) during at least the previous five years. The plots were circular in shape, with variable radius. Pure plots included at least 20 trees and mixed plots included at least 10 trees of the least represented species. Plot centre coordinates were recorded using a high precision submeter positioning GPS, with an accuracy below 1 m. Polar coordinates (x,y) of each tree within the plot were computed after measuring their distance and heading from the centre of the plot using a VERTEX IV distance meter and a compass. Trees were then numbered and measured. Tree measurements included species, diameter at breast height (1.3 m) (DBH, cm), total height (Height, m), crown base height (Live_Branch, m) and crown radius (Crown_Radio, m). Afterwards, each tree measured during the field work was plotted on a map. From the whole available network of plots, the less represented conditions of composition were both pure Pinus pinaster Ait. plots and mixed P. pinaster – P. pinea ones. Thus, all the available plots from pure Pinus pinaster Ait. stands and mixed stands were included in the sample. In a second phase, we selected all the available pure Pinus pinea L. plots located in the vicinity of the pure P. pinaster and mixed plots. At the end, twenty-seven plots were selected for the present study (Fig.1): nine pure Pinus pinaster Ait. plots, six mixed plots and twelve pure Pinus pinea L. plots. Of these, seven plots in pure Pinus pinea L. stands and five plots in pure Pinus pinaster Ait. stands were randomly selected and used for training the model. The remaining fifteen plots – five in pure Pinus pinea L. stands, four in pure Pinus pinaster Ait. stands and six in mixed stands– were therefore used to validate the model. All individual trees measured inside each selected plot were enclosed in the sample. Species classification using low density LiDAR and satellite images 49 2.2.3 Remote sensing data Low-density LiDAR data for the study area was provided by the Spanish Geographic Institute (PNOA 2018). Point clouds were captured in 2010 using a laser scanner, Leica ALS60 sensor, with a mean density of 0.5 points per m2 and a root mean square error of 20 cm in Z altimetric accuracy. Digital files, classified and coloured, were downloaded in .laz format (2x2 km). Pleiades is a constellation of the National Centre for Space Studies (CNES, French Government) which provide very-high-resolution images. It was designed by the Optical and Radar Federated Earth Observation program (ORFEO 2018) and was launched in October 2003. Pleiades 1A and Pleiades 1B operate as a constellation in the same orbit, phased 180º apart. The identical twin satellites capture images which are transferred, orthorectified and geometrically corrected by the Spanish Remote Sensing National Plan (PNT 2018). Images were captured in June-July 2014 and each one covers an area of 400 km2. The Pleiades images comprise four multispectral bands with a spatial resolution of 2 m. The spectral ranges of the four bands are 430-550 nm (Blue), 490-610 nm (Green), 600-720 nm (Red) and 750-950 nm (Infrared). Additionally, there is a panchromatic band (480-830 nm) with a spatial resolution of 0.5 m. The time delay in the field calibration of the remote sensing data do not increase the level of uncertainty since in these mature stands large changes in such short lapse of time are not expected, particularly if there is not any disturbance. Low productivity in these inner Mediterranean stands (mean annual volume increment about 1 m3 ha-1) confirms this issue. In addition, as field data came from a net of permanent plots, we can confirm little change on time in the structure (number of stems per ha, stand composition) in the last years. 2.2.4 Remotely sensed data processing 2.2.4.1 LiDAR data processing LiDAR point files (in .laz format), which include the field plots, were compiled and normalized, subtracting their corresponding ground class points. Points below 3 m were eliminated as it was considered that this was the lowest level at which a distinction between bushes and trees could be drawn. Points over 50 m were considered ‘mistakes’, as there are no trees above this height in the study area, Chapter 2 50 hence, they were also removed. The algorithm for tree delineation is the one by Valbuena et al. (2016) written as a function in a SQL language for use in PostgreSQL 9.2 with the POSTGIS 2.0 add-on, which gave the database manager geometrical and other GIS functions. It was applied to each normalised point cloud in .shp format. The algorithm is based on cloud point LiDAR analysis, searching directly for relative maximums. Once a relative maximum has been located, the next lower LiDAR return is found, and the distance between them is calculated. The algorithm determines whether this distance implies that the point is a new apex or whether it forms part of the first crown found, thus allowing individual tree crowns to be delineated. To reach each crown, the algorithm divides the point cloud into layers half a metre deep and analyses the distance from each point to the crown polygons already delineated in each layer, in a downward direction (Valbuena et al. 2016a). This process delineates the horizontal projection of each crown as a polygon that takes in all the points which impact on each tree, identifying individual tree crowns drawing them as vector format (Fig.2). The result takes the form of a .shp layer that contains each crown polygon for each tree in graphic form and an associated table with the following data for each crown: maximum height of tree crown (Max_Height, m), minimum height of tree crown (Min_Height, m), crown area of individual tree (Crown_Area, m2), crown height of individual tree (Crown_Height, m), understood as the difference between Max_Height and Min_Height. Density (Density, trees per ha) was calculated from the centroid of the polygon using the 6-tree sampling method (Prodan 1968). Species classification using low density LiDAR and satellite images 51 Fig.2 Crown delineation in a mixed stand using the Valbuena et al. (2016) algorithm. In order to characterize the structural differences between Pinus pinea L. and Pinus pinaster Ait., three relations were derived from the previous variables: Heights relation (HR), calculated by the relationship between maximum height and minimum height ((Max_Height – Min_Height) / Max_Height); Area and height relation (AHR), estimated as the quotient between crown area and maximum tree height (Crown_Area / Max_Height), it reveals the relationship between Pinus pinea L. and Pinus pinaster Ait. outline crowns, something each species is usually characterized by; and Crown and Height relation (CHR), calculated as the relationship between maximum height and crown base height ((Max_Height – Crown_height) / Max_Height), it expresses the relationship between height tree and natural pruning height, which is generally higher in Pinus pinea L. than in Pinus pinaster Ait. Thus, six LiDAR variables got into modelling: Crown_Area, Crown_Height, Density, HR, AHR, and CHR (Table 1 and Table 2). Chapter 2 52 Table 1. Descriptive statistics of training dataset. Pp_p: Pinus pinea L. individual trees in pure stand plots, Pt_p: Pinus pinaster Ait. individual trees in pure stand plots. Minimum Maximum Mean Standard Desviation Pp_p Pt_p Pp_p Pt_p Pp_p Pt_p Pp_p Pt_p FIELD DATA DBH (cm) 17.7 28.3 63.7 48.0 35.9 36.6 10.7 5.2 Height (m) 6.7 11.7 15.4 18.8 10.7 15.0 2.5 1.4 Live_Branch (m) 2.0 0.0 9.0 12.7 4.9 7.8 2.0 2.5 Crown_Radius (m) 1.7 0.9 5.8 3.9 3.3 2.5 0.9 0.6 LiDAR DATA Max_Height (m) 6.4 9.3 14.4 16.1 9.9 13.5 2.4 1.5 Min_Height (m) 3.0 3.0 11.5 12.0 6.3 8.0 2.1 2.2 Crown_Area (m2) 16.9 19.5 80.1 100.8 34.9 30.0 17.7 12.7 Crown_Height (m) 0.8 0.9 10.2 11.2 3.7 5.5 1.7 2.1 Density (trees/ha) 55.0 63.0 237.0 216.0 114.3 107.5 42.7 24.1 HR 0.12 0.08 0.77 0.78 0.37 0.41 0.13 0.15 AHR 1.69 1.34 7.20 7.04 3.46 2.23 1.34 0.88 CHR -0.86 -0.86 -0.41 -0.26 -0.67 -0.51 0.10 0.13 SATELLITE DATA Blue 96.4 93.2 123.2 129.4 107.6 107.6 7.0 6.9 Green 120.6 116.2 146.0 141.0 130.4 127.2 5.2 5.0 Red 135.4 133.6 155.1 156.7 143.6 143.2 4.0 4.3 Infrared 246.6 220.3 306.4 278.1 269.2 249.1 13.3 13.7 NDVI 0.26 0.21 0.35 0.31 0.30 0.27 0.02 0.02 EVI 0.42 0.34 0.59 0.53 0.51 0.45 0.04 0.05 Table 2. Descriptive statistics of validation dataset. Pp_p: Pinus pinea L. individual trees in pure stand plots, Pt_p: Pinus pinaster Ait. individual trees in pure stand plots, Pp_m: Pinus pinea L. individual trees in mixed stand plots, Pt_m: Pinus pinaster Ait. individual trees in mixed stand plots. Minimum Maximum Mean Standard Deviation Pp_p Pt_p Pp_m Pt_m Pp_p Pt_p Pp_m Pt_m Pp_p Pt_p Pp_m Pt_m Pp_p Pt_p Pp_m Pt_m FIELD DATA DBH (cm) 34.6 27.2 26.9 22.5 71.0 58.6 57.2 52.6 44.6 41.0 40.0 32.7 7.1 6.9 7.5 6.2 Height (m) 12.4 11.5 9.6 11.0 19.9 21.6 16.7 16.9 15.0 16.2 12.6 13.7 1.7 2.8 1.6 1.6 Live_Branch (m) 3.8 2.6 2.4 2.7 9.8 14.2 9.1 11.3 6.4 8.6 5.4 7.9 1.7 3.6 1.5 1.9 Crown_Radius (m) 2.6 1.8 2.1 1.4 6.0 3.7 5.9 3.7 3.9 2.7 3.8 2.5 0.7 0.5 0.8 0.6 LiDAR DATA Max_Height (m) 11.1 7.7 6.8 10.0 18.4 18.0 16.0 15.6 14.0 13.7 12.0 12.7 1.7 2.8 1.6 1.3 Min_Height (m) 2.1 2.2 3.1 4.2 13.3 13.8 12.2 10.8 8.4 7.8 7.4 7.4 2.6 2.5 1.7 1.7 Crown_Area (m2) 23.5 12.5 19.5 19.5 87.7 75.9 105.2 107.0 42.7 26.0 38.6 36.2 16.9 14.9 18.5 17.7 Crown_Height (m) 2.3 1.2 0.5 1.6 11.9 14.9 11.2 10.3 5.5 5.9 4.6 5.3 2.1 3.1 1.7 1.8 Density (trees/ha) 41.0 60.0 48.0 68.0 227.0 179.0 206.0 197.0 106.6 113.8 108.5 122.7 44.6 28.9 36.4 42.0 HR 0.2 0.1 0.08 0.15 0.8 0.9 0.78 0.70 0.4 0.4 0.38 0.41 0.2 0.2 0.13 0.13 AHR 1.7 0.9 1.53 1.53 6.1 5.8 9.01 9.62 3.0 1.9 3.23 2.87 1.0 1.0 1.53 1.51 CHR -5.1 -4.8 0.22 0.30 -0.7 0.1 0.92 0.85 -2.0 -0.9 0.62 0.59 1.0 1.0 0.13 0.13 SATELLITE DATA Blue 65.6 54.5 90.6 92.0 87.9 99.4 120.5 113.6 74.7 72.3 102.2 103.0 5.1 11.6 18.7 6.7 Green 86.3 74.4 113.5 117.0 107.3 115.2 133.5 131.5 96.8 93.4 123.7 124.8 4.7 10.5 17.4 4.1 Red 102.1 96.7 132.1 132.4 121.5 135.5 148.4 147.1 111.2 109.7 139.8 140.3 4.0 9.8 17.0 3.8 Infrared 199.3 136.9 201.1 228.9 270.4 269.9 287.5 282.7 238.1 208.1 253.0 252.5 17.2 34.7 29.6 15.7 NDVI 0.27 0.17 0.20 0.22 0.43 0.38 0.36 0.34 0.36 0.30 0.29 0.28 0.03 0.05 0.04 0.03 EVI 0.45 0.27 0.32 0.36 0.76 0.65 0.61 0.59 0.63 0.51 0.48 0.47 0.06 0.09 0.08 0.06 Species classification using low density LiDAR and satellite images 53 2.2.4.2 Tree crown selection Following the process of individual tree crown delineation, a subset of trees with clearly identified crowns and which had been matched with trees in the field inventory of the plot were selected for the analysis. Selected training data in pure stands is more reliable in relation to mixed ones, due to the extreme importance of tree position’s high precision at individual tree level. Once the delineation algorithm generated the crowns, both shapes were matched one to one automatically though QGIS intersect tool (QGIS Development Team 2017), throwing away crowns when more than one measured tree coincided on one delineated crown to guarantee that each crown has been issued to coincide with one measured tree. 95.5% and 81.0% of trees measured has respectively overlapped with crown delineation data in P. pinea and P. pinaster pure stands. On the other hand, overlapping between field data measured and delineated crowns in mixed stands was 126.0% due to an overestimation of delineated crowns. Thus, the algorithm is properly applied to this kind of stands. This resulted in 42 Pinus pinea L. trees and 76 Pinus pinaster Ait. trees from the pure forest plots to be used for training the model (Table 1), while 82 Pinus pinea L. (36 from pure plots and 46 from mixed plots) and 71 Pinus pinaster Ait. (36 from pure plots and 35 from mixed stands) were selected for the validation process (Table 2). 2.2.4.3 Plèiades data processing Firstly, 0.5 m high resolution multispectral images were created from the pansharpening algorithm of high-resolution panchromatic and lower resolution multispectral imagery. Based on the previous individual crown delineation, variables derived from Pleiades images were estimated for each object. Pleiades reflectance information was assigned to crown delineation with the mean values of all pixels within each crown, estimating Blue band value, Green band value, Red band value, and Infrared band value. Data from each band was assigned to the generated objects from the crown delineation using QGIS Zonal Statistics Plugin. In addition, different image bands were used to calculate two spectral indices: NDVI Normalized Difference Vegetation Index (Rouse et al. 1973) and EVI Enhanced Vegetation Index (Liu and Huete 1995). NDVI was selected as it is the most commonly used vegetation indicator, capable of assessing the quantity, quality, Chapter 2 54 and developmental stage of vegetation. It was calculated using Pleiades bands following the equation: [(Infrared – Red) / (Infrared + Red)], where infrared and red refers to the recorded Pleiades values in the infrared and red bands, respectively. EVI is an optimized vegetation index designed to enhance the vegetation signal with improved sensitivity which reduces the adverse effects of environmental factors such as atmospheric conditions and soil background. This was calculated using the equation: [2.5*(Infrared – Red) / (Infrared + 2.4*Red+1)]. Statistics for the training and validation dataset are also shown in Table 1 and Table 2. The six variables from satellite data that were finally available for modelling were: Blue band, Green band, Red band, Infrared band, NDVI index and EVI index. 2.2.5 Classification technique Random Forest classifier (Breiman et al. 2001) is a machine learning methodology based on independent decision trees providing diverse ways to explore numerically and graphically complex relationships, improving the accuracy of the model prediction (Valbuena et al. 2016b; Vega Isuhuaylas et al. 2018). It averages several decision trees, so, there is a significantly lower risk of overfitting and it also reduces the chance of stumbling across a classifier that does not perform well. A large number of classification trees are produced from a random subset of training data, with permutations introduced at each node, selecting the most common classification result. In this study, Random Forest was used to classify species based on LiDAR and satellite data at individual tree level. R software (R Core Team 2016) was used to generate the model, applying Random Forest library as is common practice in forest science. Three parameters are needed to optimize the model: (i) mtry, number of variables randomly sampled as candidates at each split; (ii) ntree, number of bootstrap replicates; and (iii) nodesize, minimum size of terminal nodes. To identify those parameter values leading to the best species discrimination, the model was optimized based on out-of-bag error estimates (OOB), defined as the rate of classification error estimated for the different subsampling sets from the training set used to train the model classifier. In addition, random forest algorithm can estimate importance of the variable in every model, showing the mean decrease of accuracy, which is defined as the loss of accuracy measured by the OOB-error Species classification using low density LiDAR and satellite images 55 when leaving out a variable (Breiman et al. 2001). Higher values of those statistics indicate higher importance of each variable for the classification. Three training models were developed. First, the ‘LiDAR model’ was constructed only using LiDAR variables; then the ‘Spectral model’ was built including only satellite variables; and finally, all available variables were incorporated into the so-called ‘Complete model’. Each of them was designed using the same approach, attempting to identify the best combination of the three Random Forest parameters and they were evaluated based on out-of-bag error estimation. The training model which provided the greatest predictive capacity was then validated in terms of classification accuracy on the three independent validation datasets: (i) individual trees in pure Pinus pinea L. stand plots, (ii) individual trees in pure Pinus pinaster Ait. stand plots, and (iii) individual trees in Pinus pinea L. and Pinus pinaster Ait. mixed stand plots. Validation was done by: (i) building confusion matrices which show the relationship between the false positive fraction and true negative fraction (or vice versa), (ii) estimating overall accuracy (OA) as ([number of correctly classified Pinus pinea L. + number of correctly classified Pinus pinaster Ait.] / total number of trees used in the validation), and (iii) calculating receiver operating characteristic (ROC) area under the curve (AUC) in order to quantify the uncertainty in models’ prediction (Zipkin et al. 2012). ROC curve is a plot of the sensitivity of a diagnostic test against one minus its specificity, as the cut-off criterion for indicating a positive test is varied (Everitt and Skrondal 2010). As discussed in Yesilnacar (2005), the quantitative-qualitative relationship between the AUC and prediction accuracy can be classified as follows: 50-60% (poor), 60-70% (average), 70-80% (good), 8090% (very good), and 90-100% (excellent). 2.3 Results 2.3.1 Training model selection To achieve the best discrimination between the two pine species, Pinus pinea L. and Pinus pinaster Ait., three training models were developed based on data from individual trees in pure stand plots. The ‘Spectral model’ is less accurate than the ‘LiDAR model’ (Table 3), so species discrimination based on the latter is better. However, the ‘Complete model’, which combines LiDAR and satellite variables provided the lowest out-of-bag error estimates. Chapter 2 62 information, in order to prevent combining information from inside and outside the crown delineation. 2.5 Conclusion The species classification tool developed allows, jointly used with an algorithm for automatic crown delineation, to discriminate two very similar species in mixed Mediterranean conifer forests using continuous spatial information on surface: Pleiades images and open source LiDAR data. The method presented here should be valid for regional or landscape-scale assessment of resources computed at tree-level (such as NWFP), enhancing the economic value of the forest production and providing support for forest management decision making. Species classification using low density LiDAR data and satellite image 63 2.6 References Arias-Rodil M, Diéguez-Aranda U, Álvarez-González JG, et al (2018) Modeling diameter distributions in radiata pine plantations in Spain with existing countrywide LiDAR data. Ann For Sci 75:36. doi: 10.1007/s13595-018-0712-z Ballanti L, Blesius L, Hines E, Kruse B (2016) Tree species classification using hyperspectral imagery: A comparison of two classifiers. Remote Sens 8:1–18. doi: 10.3390/rs8060445 Barba J, Curiel Yuste J, Martínez-Vilalta J, Lloret F (2013) Drought-induced tree species replacement is reflected in the spatial variability of soil respiration in a mixed Mediterranean forest. For Ecol Manage 306:79–87. doi: 10.1016/j.foreco.2013.06.025 Barrett F, McRoberts RE, Tomppo E, et al (2016) A questionnaire-based review of the operational use of remotely sensed data by national forest inventories. Remote Sens Environ 174:279–289. doi: 10.1016/j.rse.2015.08.029 Beguet B, Chehata N, Boukir S, et al (2014) Classification of forest structure using very high resolution Pleiades image texture. Geosci Remote Sens Symp (IGARSS), 2014 IEEE Int 2324–2327. doi: 10.1109/IGARSS.2014.6946936 Bravo-Oviedo A, Gallardo-Andrés C, del Río M, Montero G (2010) Regional changes of Pinus pinaster site index in Spain using a climate-based dominant height model. Can J For Res 40:2036–2048. doi: 10.1139/X10-143 Breiman L, Friedman J, Olshen R, Stone C (2001) Classification and regression trees. Chapman Hall, New York Calama R, Gordo FJ, Mutke S, Montero G (2008) An empirical ecological-type model for predicting stone pine (Pinus pinea L.) cone production in the Northern Plateau (Spain). For Ecol Manage 255:660–673. doi: 10.1016/j.foreco.2007.09.079 Calama R, Gordo J, Madrigal G, et al (2016) Enhanced tools for predicting annual stone pine (Pinus pinea L.) cone production at tree and forest scale in Inner Spain. For Syst 25:. doi: 10.5424/fs/2016253-09671 Castaño-Díaz M, Álvarez-Álvarez P, Tobin B, et al (2017) Evaluation of the use of low-density LiDAR data to estimate structural attributes and biomass yield in a short-rotation willow coppice: an example in a field trial. Ann For Sci 74:69. doi: 10.1007/s13595-017-0665-7 Dalponte M, Bruzzone L, Gianelle D (2012) Tree species classification in the Southern Alps based on the fusion of very high geometrical resolution multispectral/hyperspectral images and LiDAR data. Remote Sens Environ 123:258–270. doi: 10.1016/j.rse.2012.03.013 Dechesne C, Mallet C, Le Bris A, Gouet-Brunet V (2017) Semantic segmentation of forest stands of pure species combining airborne lidar data and very high resolution multispectral imagery. ISPRS J Photogramm Remote Sens 126:129–145. doi: 10.1016/j.isprsjprs.2017.02.011 Deng S, Katoh M, Yu X, et al (2016) Comparison of tree species classifications at the individual tree level by combining ALS data and RGB images using different algorithms. Remote Sens 8:. doi: 10.3390/rs8121034 Chapter 2 64 Dubayah RO, Drake JB (2000) Lidar Remote Sensing for Forestry Applications. J For 98:44–46 European Environmental Agency (2006) European forest types: Categories and types for sustainable forest management reporting policy. 114 Everitt BS, Skrondal A (2010) The Cambridge Dictionary of Statistics Fassnacht FE, Latifi H, Stereńczak K, et al (2016) Review of studies on tree species classification from remotely sensed data. Remote Sens Environ 186:64–87. doi: 10.1016/j.rse.2016.08.013 Forrester DI, Benneter A, Bouriaud O, Bauhus J (2017) Diversity and competition influence tree allometric relationships – developing functions for mixed-species forests. J Ecol 105:761– 774. doi: 10.1111/1365-2745.12704 Gordo FJ, Montero G, Gil L (2012) La problemática de la regeneración natural de los pinares en los arenales de la Meseta Castellana Hollaus M, Dorigo W, Wagner W, et al (2009) Operational wide-area stem volume estimation based on airborne laser scanning and national forest inventory data. Int J Remote Sens 30:5159–5175. doi: 10.1080/01431160903022894 Holmgren J, Persson Å (2004) Identifying species of individual trees using airborne laser scanner. Remote Sens Environ 90:415–423. doi: 10.1016/S0034-4257(03)00140-8 Holmgren J, Persson Å, Söderman U (2008) Species identification of individual trees by combining high resolution LiDAR data with multi-spectral images. Int J Remote Sens 29:1537–1552. doi: 10.1080/01431160701736471 Korpela I, Ole Ørka H, Maltamo M, et al (2010) Tree species classification using airborne LiDAR - effects of stand and tree parameters, downsizing of training set, intensity normalization, and sensor type. Silva Fenn 44:319–339. doi: 10.14214/sf.156 Leckie DG, Tinis S, Nelson T, et al (2005) Issues in species classification of trees in old growth conifer stands. Can J Remote Sens 31:175–190. doi: 10.5589/m05-004 Ledo A, Cañellas I, Barbeito I, et al (2014) Species coexistence in a mixed Mediterranean pine forest: Spatio-temporal variability in trade-offs between facilitation and competition. For Ecol Manage 322:89–97. doi: 10.1016/j.foreco.2014.02.038 Lee J, Cai X, Lellmann J, et al (2016) Individual tree species classification from airborne multisensor imagery using robust PCA. IEEE J 1–15 Lim K, Treitz P, Wulder M, et al (2003) LiDAR remote sensing of forest structure. Prog Phys Geogr 27:88–106. doi: 10.1191/0309133303pp360ra Lin C, Lo CS, Thomson G (2011) A textural modification of the MMAC algorithm for individual tree delineation in forest stand using aerial bitmap images. Proc - 4th Int Congr Image Signal Process CISP 2011 3:1604–1608. doi: 10.1109/CISP.2011.6100501 Liu HQ, Huete A (1995) A feedback based modification of the NDVI to minimize canopy background and atmospheric noise. IEEE Trans Geosci Remote Sens 33:457–465 Species classification using low density LiDAR data and satellite image 65 Lopez-García P (1980) Estudio de semillas prehistóricas en algunos yacimientos españoles. Trab Prehist 37:419–432 López C, Espinosa J, Bengoa J (2009) Mapa de Vegetación de Castilla y León Maack J, Kattenborn T, Fassnacht FE, et al (2015) Modeling forest biomass using very-highresolution data - combining textural, spectral and photogrammetric predictors derived from spaceborne stereo images. Eur J Remote Sens 48:245–261. doi: 10.5721/EuJRS20154814 MAGRAMA (2012) Criteria and indicators for sustainable forest Management, in Spanish forests 2012 Miura N, Jones SD (2010) Characterizing forest ecological structure using pulse types and heights of airborne laser scanning. Remote Sens Environ 114:1069–1076. doi: 10.1016/j.rse.2009.12.017 Moreno-Fernández D, Cañellas I, Calama R, et al (2013) Thinning increases cone production of stone pine (Pinus pinea L.) stands in the Northern Plateau (Spain). Ann For Sci 70:761– 768. doi: 10.1007/s13595-013-0319-3 Næsset E (2004) Effects of different flying altitudes on biophysical stand properties estimated from canopy height and density measured with a small-footprint airborne scanning laser. Remote Sens Environ 91:243–255. doi: 10.1016/j.rse.2004.03.009 Nanos N, Tadesse W, Montero G, et al (2000) Modelling resin production distributions for Pinus Pinaster Ait using two probability functions. Ann For Sci 57:379–377. doi: 10.1051/forest:2000128 Nanos N, Tadesse W, Montero G, et al (2001) Spatial stochastic modeling of resin yield from pine stands. Can J For Res 31:1140–1147. doi: 10.1139/cjfr-31-7-1140 Nanos N, Calama R, Montero G, Gil Luis (2004) Geostatistical prediction of height/diameter models. For Ecol Manage 195: 221-235. doi: 10.1016/j.foreco.2004.02.031 Nguyen HT, Hutyra LR, Hardiman BS, Raciti SM (2016) Characterizing forest structure variations across an intact tropical peat dome using field samplings and airborne LiDAR. Ecol Appl 26:587–601. doi: 10.1890/15-0017 ORFEO (2018) ORFEO Ørka HO, Gobakken T, Næsset E, et al (2012) Simultaneously acquired airborne laser scanning and multispectral imagery for individual tree species identification. Can J Remote Sens 38:125–138. doi: 10.5589/m12-021 PNOA (2018) Plan Nacional de Ortofotografía Aérea PNT (2018) Plan Nacional de Teledeteccíon Popescu SC (2007) Estimating biomass of individual pine trees using airborne lidar. Biomass and Bioenergy 31:646–655. doi: 10.1016/j.biombioe.2007.06.022 Chapter 2 66 Popescu SC, Wynne RH, Nelson RF (2003) Measuring individual tree crown diameter with lidar and assessing its influence on estimating forest volume and biomass. Can J Remote Sens 29:564–577. doi: 10.5589/m03-027 Prieto-Recio C, Martín-García J, Bravo F, Diez JJ (2015) Unravelling the associations between climate, soil properties and forest management in Pinus pinaster decline in the Iberian Peninsula. For Ecol Manage 356:74–83. doi: 10.1016/j.foreco.2015.07.033 Prodan M (1968) Forest bimetrics, Oxford, UK QGIS Development Team (2017) QGIS Geographic Information System. Open Source Geospatial Foundation Project R Development Core Team (2016) R: A Language and Environment for Statistical Computing Riofrío J, del Río M, Pretzsch H, Bravo F (2017) Changes in structural heterogeneity and stand productivity by mixing Scot pine and Maritime pine. For Ecol Manage InProcess: doi: 10.1016/j.ibmb.2009.12.007 Rodríguez-García A, Martín JA, López R, et al (2015) Influence of climate variables on resin yield and secretory structures in tapped Pinus pinaster Ait. in central Spain. Agric For Meteorol 202:83–93. doi: 10.1016/j.agrformet.2014.11.023 Rodríguez F, Lizarralde I (2015) Inventario forestal con tecnología LiDAR, ¡Más barato y más preciso! Forestalis 24:33–34 Rougier S, Anne P (2014) Improvements of Urban vegetation segmentation and classification using multi ‐ temporal Pleiades images. South‐Eastern Eur J Earth Obs Geomatics 3:409–414 Rouse JW, Hass RH, Schell JA, Deering DW (1973) Monitoring vegetation systems in the great plains with ERTS. Third Earth Resour Technol Satell Symp 1:309–317. doi: citeulike-articleid:12009708 Ruiz LÁ, Recio JA, Crespo-Peremarch P, Sapena M (2016) An object-based approach for mapping forest structural types based on low-density LiDAR and multispectral imagery. Geocarto Int 6049:1–15. doi: 10.1080/10106049.2016.1265595 Soliño M, Yu T, Alía R, et al (2018) Resin-tapped pine forests in Spain: Ecological diversity and economic valuation. Sci Total Environ 625:1146–155. doi: 10.1016/j.scitotenv.2018.01.027 Suratno A, Seielstad C, Queen L (2009a) Tree species identification in mixed coniferous forest using airborne laser scanning. ISPRS J Photogramm Remote Sens 64:683–693. doi: 10.1016/j.isprsjprs.2009.07.001 Suratno A, Seielstad C, Queen L (2009b) Mapping Tree Species Using LIDAR in MixedConiferous Forests. Proc Silvilaser 2009 1–10 Tonolli S, Dalponte M, Neteler M, et al (2011) Fusion of airborne LiDAR and satellite multispectral data for the estimation of timber volume in the Southern Alps. Remote Sens Environ 115:2486–2498. doi: 10.1016/j.rse.2011.05.009 Species classification using low density LiDAR data and satellite image 67 Valbuena M, Santamaría J, Sanz F (2016a) Estimation of diameter and height of individual trees for Pinus sylvestris L. based on the individualising of crowns using airborne LiDAR and the National forest inventory data. For Syst 25:1–11. doi: 10.5424/fs/2016251-05790 Valbuena R, Maltamo M, Packalen P (2016b) Classification of forest development stages from national low-density lidar datasets: a comparison of machine learning methods. Rev Teledetección 15–25. doi: 10.4995/raet.2016.4029 Vauhkonen J, Ørka HO, Holmgren J, et al (2014) Tree species recognition based on Airborne Laser Scanning and complementary data source. 27:. doi: 10.1007/978-94-017-8663-8 Vega Isuhuaylas LA, Hirata Y, Santos LCV, Torobeo NS (2018) Natural forest mapping in the Andes (Peru): A comparison of the performance of machine-learning algorithms. Remote Sens 10:. doi: 10.3390/rs10050782 Wang L (2010) A Multi-scale Approach for Delineating Individual Tree Crowns with Very High Resolution Imagery. Photogramm Eng Remote Sens 76:371–378. doi: 10.14358/pers.76.4.371 Xie Y, Sha Z, Yu M (2008) Remote sensing imagery in vegetation mapping: a review. J Plant Ecol 1:9–23. doi: 10.1093/jpe/rtm005 Yesilnacar EK (2005) The Application of Computational Intelligence to Landslide Susceptibility Mapping in Turkey (Ph.D Thesis) Department of Geomatics the University of Melbourne, p. 423. Zhang Z, Kazakova A, Moskal LM, Styers DM (2016) Object-based tree species classification in urban ecosystems using LiDAR and hyperspectral data. Forests 7:1–16. doi: 10.3390/f7060122 Zipkin EF, Grant EHC, Fagan WF (2012) Evaluating the predictive abilities of community occupancy models using AUC while accounting for imperfect detection. Ecol Appl 22, 1962-1972. CHAPTER 3 Stand types discrimination comparing machine learning algorithms in monteverde, Canary Islands “Mira profundamente en la naturaleza y entonces comprenderás todo mejor” Albert Einstein Chapter 3 78 Fig. 2. Distribution of selected features as a result of variable selection applying random forest (VSURF) according to monteverde typologies. Values under normalization from minimum ‘0’ to maximum value ‘1’. DTHM-2016, digital tree height model based on 2016 LiDAR data; P95-2016, height percentile 95 based on 2016 LiDAR data; HEIGHT-2009, maximum height based on 2009 LiDAR data; DTM-2016, digital terrain model based on 2016 LiDAR data; Red-SD, standard deviation of the Red band orthophotograph data; HGR, height growth based on 2009 and 2016 LiDAR data; RATIO R/G, ratio between Red and Green bands orthophotograph data; NDVI705, Normalized Differenced Vegetation Index 705 2015 image data; RED, Red band orthophotograph data; Band04, Sentinel band 04 2015 image data. Stand types discrimination comparing machine learning algorithms 79 Table 1. Accuracy, Accuracy Standard Deviation, Kappa Value and Kappa Value Standard Deviation resulted from each model. ANN, Artificial Neural networks algorithm; SVML, Linear support vector machine algorithm; SVMR, radial support vector machine algorithm; RF, random forest algorithm. Accuracy Accuracy SD Kappa Kappa SD ANN 0.891 0.073 0.817 0.129 SVML 0.904 0.041 0.842 0.070 SVMR 0.902 0.060 0.836 0.106 RF 0.904 0.056 0.841 0.092 Table 2. Importance of variables in the linear support vector machine algorithm (SVML) final model for the monteverde typology classification. Values state the loss of accuracy and/or standard deviation according to the lack of each feature from SVML calculated. DTHM-2016, digital tree height model based on 2016 LiDAR data; P95-2016, height percentile 95 based on 2016 LiDAR data; HEIGHT-2009, maximum height based on 2009 LiDAR data; DTM-2016, digital terrain model based on 2016 LiDAR data; Red-SD, standard deviation of the Red band orthophotograph data; HGR, height growth based on 2009 and 2016 LiDAR data; RATIO R/G, ratio between Red and Green bands orthophotograph data; NDVI705, Normalized Differenced Vegetation Index 705; RED, Red band orthophotograph data; Band04, Sentinel band 04 image data. Source Mean Standard Deviation SVML 0.904 0.041 DTHM-2016 LiDAR 0.892 0.055 P95-2016 LiDAR 0.903 0.052 HEIGHT-2009 LiDAR 0.909 0.045 DTM-2016 LiDAR 0.880 0.049 Red-SD Orthophotography 0.908 0.043 HGR LiDAR 0.886 0.057 RATIO R/G Orthophotography 0.907 0.046 NDVI705 Sentinel 0.900 0.042 RED Orthophotography 0.915 0.044 Band04 Sentinel 0.915 0.049 Obtained results differ with other previous ones in the literature. For instance, Valbuena et al. (2016) found that the best machine learning algorithm to determine Mediterranean forest development stages are RF and ANN, and Vega Chapter 3 80 Isuhuaylas et al. (2018) met with SVM and RF to classify Andes mountain forests and shrubland land cover classes. The high accuracy values confirm the importance of taking into account diverse machine-learning methods for stand classification purposes and different explanatory aspects. Notwithstanding the small deviation between accuracy values, our work proves that SVML is the best algorithm for the monteverde forest classification due to the minimal results’ scattering. Over 90% of cases monteverde stand type (BR, FBA, LA, NMG, or NMF) determined from remote sensing data are correct, though small size of the study area and the only three, plus two, stand types considered here should not be forgotten. Our results confirm machine learning classification is a suitable tool to optimize classification in monteverde forest and, thus, its management. Boost of machine learning algorithms applied to classify forest is broadly enough demonstrated. Although differences between errors in accuracy and scattering may not seem significant at a first glance, such differences are highly important when working at large scales. Errors may be higher when classifying larger areas with more stand types. So, comparisons between algorithms should be considered when stand classification analyses are performed owing to different behaviour of algorithms relying on stand types, features and size of the study area. Stand types discrimination comparing machine learning algorithms 81 3.4 References Arozena ME, Panareda, JM, 2013. Forest transition and biogeographic meaning of the current laurel forest landscape in Canary Islands, Spain. Physical Geography 34: 211-235. Genuer R, Poggi JM, Tuleau-Malot C, 2015. VSURF: An R Package for Variable Selection Using Random Forests. https://journal.r-project.org/archive/2015/RJ-2015-018/RJ-2015-018.pdf [8 June 2018]. Henrich V, Krauss G, Götze C, Sandow C, 2012. IDB - www.indexdatabase.de, Entwicklung einer Datenbank für Fernerkundungsindizes. AK Fernerkundung, Bochum 45: 10. Kuhn M, Jed Wing A, Weston S, Williams A, Keefer C, Engelhardt A, Cooper T, et al., 2018. caret: Classification and Regression TrainingR package version 6.0-78. https://cran.rproject.org/package=caret [8 June 2018]. McGaughey R, 2007. Fusion/LDV: Software for Lidar Data Analysis and Visualization; USDA Forest Service, Pacific Northwest Research Station: Portland, OR, USA. Nitze I, Schulthess U, Asche H, 2012. Comparison of machine learning algorithms Random Forest, Artificial Neural Network and Support Vector Machine to maximum likelihood for supervised crop type classification. Proc GEOBIA, Rio de Janeiro (Brazil), May 7-9, pp: 35. OTB Development Team, 2017. The ORFEO Tool Box Software Guide. Updated for OTB-5.10.0 QGIS Development Team, 2017. QGIS Geographic Information System. Open Source Geospatial Foundation Project. http://qgis.osgeo.org Sinergise, 2018. Sentinel 2 EO products. https://www.sentinelhub.com/develop/documentation/eo_products/Sentinel2EOproducts. [24 May 2018]. Valbuena R, Maltamo M, Packalen P, 2016. Classification of forest development stages from national low-density LiDAR datasets: a comparison of machine learning methods. Revista de Teledetección 45: 15-25. Vega Isuhuaylas LA, Hirata Y, Ventura Santos LC, Serrudo Torobeo N, 2018. Natural forest mapping in the Andes (Peru): a comparison of the performance of machine-learning algorithms. Remote Sens 10: 782. Xu C, Manley B, Morgenroth J, 2018. Evaluation of modelling approaches in predicting forest volume and stand age for small-scale plantation forests in New Zealand with RapidEye and LiDAR. Int J Appl Earth Obs Geoinformation 73: 386-396. Zhu XX, Tuia D, Mou L, Xia GS, Zhang L, Xu F, Fraundorfer F, 2017. Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources. IEEE Geosci Remote Sens Mag 5: 8-36. CHAPTER 4 Fire and burn severity assessment: calibration of Relative Differenced Normalized Burn Ratio (RdNBR) with field data “Salvaje no es quien vive en la naturaleza, salvaje es quien la destruye” Anónimo Chapter 4 86 Chapter 4. Fire and burn severity assessment: calibration of Relative Differenced Normalized Burn Ratio (RdNBR) with field data Abstract The assessment of burn severity is highly important in order to describe and measure the effects of fire on vegetation, wildlife habitat and soils. The estimation of burn severity based on remote sensing is a powerful tool that, to be useful, needs to be related and validated with field data. The present paper explores the relationships between field accessible variables and Relative Differenced Normalized Burn Ratio (RdNBR) index by using linear mixed-effects models and boosted regression trees, based on data from 28 large fires and 668 field measurements across three countries in southern Europe. The RdNBR clearly reflected the mean height of charred stem and loss of ligneous, living shrub and tree cover during the fire. The paper confirms that remote sensing indices provide an acceptable assessment of fire induced impact on forest vegetation but also highlights there are important between-fire variations due to specific contexts that modify these relationships. These variations can be effectively assessed and should be taken into account in future predictive efforts. RdNBR calibration with field data 87 4.1 Introduction Fire is a key ecological process in many forested areas, and its variation in size and severity strongly influences forest structure and wildlife habitats (Herrando and Brotons 2002; Herrando et al., 2003). Large amounts of data are required to define post-fire management plans, either at strategic or tactical level. In this context, remote sensing is a necessary and reliable tool for providing spatially continuous information about the state of forest ecosystems, associated resources and for assessing their evolution overtime (Franklin, 2001; Jones and Vaughan, 2010; Lefsky et al., 2002; Xie et al., 2008). Remote sensing-based assessments have gained relevance in natural resources management as they are much less time-consuming than assessments entirely based on field inventories, while reducing uncertainties associated with inference and estimation from sample surveys in areas showing fine-scale heterogeneity (Franklin, 2001). In addition, these assessments have also proved their validity to improve the retrieval of representative field data through stratification methods (Adam et al., 2010; Czaplewski and Patterson, 2003; Grafström and Ringvall, 2013). Currently, a variety of remotely sensed images are available for providing information about forest ecosystems by a range of airborne and satellite sensors, from multi-spectral to hyperspectral sensors, with different spatial and temporal resolutions. The use of multispectral images has also supposed an important step ahead in the impact assessment of the impact of forest disturbances; specifically, to assess wildfire impact, as it has become an essential source of data to map burned area (Chuvieco, 2012; Fernández et al., 1997; Kushlat and Ripple, 1998; Lentile et al., 2006; Miller and Yool, 2002; Pereira, 1999; Roy et al., 2002). In this context, burn severity has become a key variable in order to assess the impact of fire on forest ecosystems and can be defined as the loss of soil and aboveground organic matter in a burned area, being used to analyse nonimmediate fire effects (Keeley, 2009; Morgan et al., 2014). Burn severity has been mapped through indices such as the Normalized Burn Ratio (NBR; Key and Benson, 1999) considering the absolute difference change detection protocol (dNBR), or as a relative difference (RdNBR) (De Santis, 2008; French et al., 2008; Keeley, 2009). These indices combined with other metrics derived from available satellite multi-spectral images have been used for mapping burn severity in Chapter 4 94 factor was applied to the intercept. In addition to the criteria used for variable selection, in this case the variables should have no obvious biases or dependencies in the predictions, had to be significant at the 0.05 level, and the resulting models had to be parsimonious and avoid excessive complexity, since the ultimate objective was to observe the relationship of the variables, and not a full predictive model. The resulting models under this approach were evaluated quantitatively by examining the magnitude and graphical distribution of the residuals for all the variables included in the model, aiming at finding obvious dependencies that indicate systematic discrepancies. The accuracy of the model predictions was determined by calculating the bias and precision, based on the absolute and relative estimates of bias and RMSE, as well as the coefficient of determination. With all candidate variables used in the mixed models, an alternative analysis was performed based on BRT. This approach combines statistical and machine learning techniques aiming at the improvement of the performance of a single model by fitting many models and combining them for prediction (Schapire, 2003). In this case, the modelling approach allows the use of incomplete data, does not take any assumption about the shape of the relationship of the variables and the predicted, permitting multiple interactions. In this case, besides the selection of variables to be included in the models, the models are defined by additional parameters: number of trees, learning rate (or shrinkage, related to the reduction of the impact of any additional tree), bag (random fraction of the residuals is selected to build the tree, per iteration) and number of interactions between variables. The number of trees was estimated through optimization (Ridgeway, 2006) and for the other parameters, several models were tested sequentially, being alternative learning rates fixed ranging from 0.8 to 0.001, and the number of interactions from 1 to 5. The models and statistical analysis were developed in R version 3.2.4 (R core development team, 2017). The BRT models were based on the dismo (Elith et al., 2008) extension of the gbm package (Ridgeway, 2006). The mixed models were estimated using the nlme package (Pinheiro et al., 2018). RdNBR calibration with field data 95 4.3 Results The preliminary analysis showed a significant correlation between the variables considered and the RdBNR. The strongest correlations were for the variable LIGpre and LIGloss (R=0.59 and 0.57, respectively) and CHARH, CHARHmean, TREECloss, SHRCloss (R=0.44, 0.40, 0.37, 0.12, respectively), with the rest presenting lower correlations. For the dummy variables, both Firetd and Forest showed a high explanatory power (Fig. 2). As expected, correlations between the independent variables were found (LIGpre and LIGloss; CHARH and CHARHmean) and, based on this, a subset of variables was selected by considering those variables presenting better results for the statistical models. One of the variables that showed the highest relationships in the preliminary analysis (LIGloss) had discontinued records (N=241), which limited the use of the rest of the data since mixed models require complete cases to be fitted. Therefore, two alternative models were fitted, with LIGloss being substituted with surrogate variables (N=650). The final models considered under this approach were: RdNBRi j = β0 + β1 LIGlossi j + β2 CHARCHi j + β3 Foresti j + β4 Firetdi j + µj + εij [1] RdNBRi j = β0 + β1 SHRClossi j + β2 TREEClossi j + β3 CHARCHi j + β4 Foresti j + β5 Firetdi j + µj + εij [2] where RdNBR is the response variable related to LIGloss, TREECloss, SHRCloss, CHARH, Forest, Firetd all as defined in Table 1. Subscripts i and j refer to measurement i in fire j, β1-βn are the parameters to be estimated, µj and εij are independent and identically distributed random between-fire and betweenmeasurement factors, with mean equal to 0 and constant variances σfire and σobs, respectively. Chapter 4 96 Fig. 2. Histogram of the distribution of the frequencies of the RdNBR (Relativized differenced of Normalized Burn Ratio) values, (between -500 and 1500, used as thresholds in the study). Relationships between RdNBR and CHARCH (%), TREECloss (%), LIGloss (%), SHRCloss (%), Forest and Firetd. RdNBR calibration with field data 97 All variables included in the linear mixed-effects models (Eq1 and Eq2) were statistically significant at p-value<0.05. The linear regression mixed models presented a coefficient of determination for the random and fixed parts of 0.67 and 0.61, for Eq1 and Eq 2, respectively. When only considering the fixed part, these values decreased to 0.60 and 0.41, respectively, as a large percentage of the variability was explained through the random factor (Table 2), reflecting the differences due to specific conditions related to the fire event. Concerning the BRT approach, the best model used a learning rate of 0.01, 5 interactions and 550 trees. In this case, the coefficient of determination was 0.56, but it was reduced to 0.45 when it was estimated using a cross-validation approach (Fig. 3). Table 2. Estimates, standard error (S.E.) and significance level of the parameters and variance components of the mixed models (Eq 1 and Eq 2) to predict the relativized difference of normalized burn ratio (RdNRB). BRT: predictions based on boosted regression trees. The weight for each variable is the average for models with 1, 3 and 5 interactions (tree complexity, N=3), for a learning rate 0.01. Variable Value Std. Error DF t-value p-value Eq 1 β0 321 62.53 229 5.1415 < 0.001 CHARCH 3.055 0.479 229 6.380 < 0.001 LIGloss 2.568 0.612 229 4.198 < 0.001 Forest -109.854 45.722 229 -2.403 0.017 Firetd -252.589 72.841 7 -3.468 0.010 σfire 100.68 σmeasurement 201.28 Eq 2 β0 429.1 56.25 618 7.627 < 0.001 CHARCH 1.042 0.276 618 3.768 < 0.001 TREECloss 3.060 0.490 618 6.246 < 0.001 SHRCloss 1.663 0.404 618 4.110 < 0.001 Forest -82.910 28.57 618 -2.901 < 0.004 Firetd 252.406 62.09 26 4.064 < 0.001 σfire 139.83 σmeasurement 203.08 BRT Weight (mean) Std. Error CHARCH 28.9 % 1.5 Firetd 28.4 % 2.2 TREECloss 19.1 % 1.0 LIGloss 11.4 % 0.6 Forest 7.6 % 1.4 SHRCloss 4.6 % 1.7 Chapter 4 98 Fig. 3 Observed versus predicted estimates for the two mixed models: Eq 1 (N = 241) and Ep 2 (N = 650) and estimates based on Boosted Regressions Trees (BRT). Concerning model assessment, there was not observed strong systematic biases in the predicted RdNBR and the values of the variables. (Fig. 4), although in Eq2, the use of TREECloss tend, to some extent, underestimate the RdNBR. The bias was -2.749 (-0.72%) and -1.471 (-0.38%) for Eq1 and Eq2, respectively, and -0.045 (- 0.012%) for the BRT model (RMSE=218, 245 and 211, for Eq1, Eq2 and BRT, respectively). In the case of the BRT approach, it failed systematically to produce RdNBR estimates below -67. The marginal effect on RdNBR of each single variable included in the statistical analysis was determined from partial dependency plots, which showed the effect of each variable on the response after accounting for the average effects of all other variables for the linear mixed models and with one, three and five interactions for BRT. All relationships were largely consistent along the modelling approaches (Fig. 5). The effects on RdNBR increased with increasing values of CHARH and LIGloss, following a linear trend. A similar effect was found considering TREECloss, although in this case, the values of Eq2 slightly resulted in higher RdNBR than in the BRT approach. For SHRCloss there were divergences between the mixed models, which accounted for a positive relationship, and the BRT estimates, which identified a negative threshold after 80%. In the BRT estimates, alternative values in the number of interactions, learning rates or number of trees used in the calibration, did not alter the nature of the relationships. RdNBR calibration with field data 99 Eq 1 Eq 2 Fig. 4 Mean residuals as a function of the variables included in the fixed part of Eq 1 and Eq 2. The residuals have been grouped in 6 tiles of equal number of observations. The figure represents the mean residual value as well as two time the standards errors (discontinuous lines). Predicted: relativized difference of normalized burn ratio (RdNBR). Regarding the dichotomous variables Forest and Firetd, the effects on the independent variable were higher in plots measured in forest areas between two or three months after the fire occurrence. The estimated contributions of these Chapter 4 100 variables (Table 2) were highest for CHARCH, TREECloss and LIGloss (average over 20%) and the lowest for SHRCloss (average less than 5%), suggesting the explanatory effects of the latter could be explained by combinations of the other variables. It must be taken into account that the weights slightly changed with the number of interactions, which was particularly affecting the weights of Forest (decreasing till 6% for 5 interactions) and SHRCloss (increasing till 7.3% for 5 interactions). Fig. 5 Partial dependence plots of the relativized difference of normalized burn ratio (RdNRBR) in the different models tested. The light grey lines result from boosted regression trees (BRT) constructed with 1, 3 and 5 interactions, whereas the straight lines result from two linear mixed models (Eq 1 and Eq2). LIGloss, included in Eq 1 and in the BRT models, presents lower observations (N=241) than the other variables (N=650). 4.4 Discussion The present study addresses the empirical relationships between a burn severity index derived from remote sensing and field data based on 28 fire events in several countries. Due to the large areas entailed by fires, the use of Landsat based indices fits the purposes of the analysis, as their spectral and spatial resolutions make them particularly suitable for assessing burn severity (Epting et RdNBR calibration with field data 101 al., 2005; Quintano et al., 2018; Soverel et al., 2010; Van Wagtendonk et al., 2004). Particularly, the RdNBR has been used to a large success (De Santis, 2008; French et al., 2008; Keeley, 2009). But for their use, it is crucial to calibrate it with field variables, in order to analyze the performance of remote-sensing outputs (French A et al., 2008; Miller et al., 2009) and to assess their accuracy. Most of studies combining field data with remotely sensed assessments of burn severity used the CBI, or adaptations like the GeoCBI, as these indices were designed to define burn severity ecologically, and to measure ground effects visually at each vegetative strata on large and heterogeneous plots. The results of these relationships were statically satisfactory in most of these studies. However, the integration of multiple variables in a combined index makes the specific assessment of these relationships difficult. In this study, we aimed to analyse the relations of the different specific field variables (based on the ForFireS assessment) in order to better understand the information that these indices are reflecting, therefore using RdNBR as the response variable subject to study. As expected, the variables reflecting fire intensity (as mean bark char height, CHARH), or burn severity, (as loss of tree, ligneous and shrub cover, TREECloss, LIGloss and SHRCloss, respectively), were positively related with RdNBR, indicating a loss of above ground biomass and reduced photosynthetic reflectance by living biomass, similar to other studies comparing field indicators and satellite indexes on burn severity (Miller et al., 2009; Miller and Quayle, 2015). Topographical factors, that were not significant in this study, have clear influence on fire behaviour although the association between topography and burn severity cannot be generalized for all ecosystems and regions (Dragozi et al., 2016; Gonzalez-Olabarria et al., 2006; González et al., 2007) due to the prevalence of determined local conditions on the studied regions (for instance, variation on humidity between different elevations and aspects), or how the slope influences fire behaviour (higher slopes increase the rate of spread and intensity, but usually reduce the residence time of fire). Burn severity is expected to be more severe in coniferous forests than in broadleaved forests (Fernandes et al., 2010) but in our study it did not appear as significant variable. The pre-fire tree species composition could play an important role in determining post-fire reflectance and recovery and, subsequently, burn severity (Dragozi et al., 2016). It is clear that, depending on the fire intensity and the specific response of the tree species to fire, different ecosystem responses such as resprouting and regeneration can be expected or be accentuated as time goes on (Keeley et al., 2008). Such variation on vegetation characteristics not only defines an immediate response to fire, but also ecological responses overtime (Keeley et al., Chapter 4 102 2008; Liu, 2016). Even more, the season of the year when each fire takes place is expected to influence initial fire severity, and delayed vegetation response (Knapp and Keeley, 2006; Valor et al., 2017). The results confirm that remote sensing indexes provide an overall acceptable assessment of fire included impact on forest vegetation, which can be translated into estimates of biomass lost even when predictive models are applied across large and heterogeneous regions. Yet it must be taken into account that the interactions between fire seasons, vegetation types and ecological responses, among others, would require of adjustment factors. This specific adjustment was addressed by the use of random factor, which included specific conditions of each fire in a mixed-effects approach. In fact, the statistical analysis showed the important differences between fire events on the RdNBR response, suggesting that even if general relations can be inferred for any type of forest, specific variations on such relations are to be expected for each fire event as it is characterized by a specific behaviour, due to topography, weather, and fuel arrangement (McRae et al., 2005; Estes et al., 2017). The vegetation composition and development stage prior to fires is also expected to vary between fires, especially when fires are allocated across a large and highly variable region such as the South Europe plus the Canary Islands (Valavanidis and Vlachogianni, 2011). In addition to the general differences of the RdNBR response depending on field measured burn severity, the use of BRT showed that the relationships are not necessary linear and that there are several interactions between the measurements of severity and their reflection in the RdNBR values. BRT is a useful statistical technique for analysis of ecological data (De’Ath, 2007) because nonlinear effects and interactions are mechanically considered due to trees are invariant to transformations of the predictors and interactions are automatically included. This is very suitable for complex analysis as it allows addressing the relationship among the explanatory variables and the predicted variables with nearly no assumptions. Although the results of the used of BRT were largely consistent with the mixed models approach, this analysis revealed that the relationship with loss of shrub cover during the fire could be over-estimated in the models. It must notice that the use of BRT does not account for between-fire differences in the same way that is handled by a mixed model. RdNBR calibration with field data 103 Finally, a factor influencing the relation between ground measurements and RdNBR were the time lapse between the fire event and the remotely sensed burn severity assessment. The impact of the elapsed time could be derived from the RdNBR sensitivity to ash cover which declines with time since fire (Miller and Quayle, 2015) and the regeneration and response of vegetation after the fire. In that sense, as the time since fire increases, the usage of fire severity measures taken on the field to generate burn severity indexes and relate them to satellite derived images may arise problems, due to the concept of burn severity itself and the increased impact of ecological responses (Keeley, 2009). In order to reduce this time, Quintano et al. (2018) combined the use of Landsat 8 OLI and Sentinel-2A MSI data with the dNBR index and its relative version because Sentinel-2A MSI provides higher spatial and temporal resolutions and can be used in case of unavailability of a pre-fire or a post-fire Landsat image. Although the authors showed that this approach can be adequate for postfire management intervention, the accuracy of burn severity maps was lower than the maps exclusively created with Landsat imagery (Quintano et al., 2018). The results of this study provide an assessment of relevant variables that should be prioritized when designing field inventories to assess burn severity, especially when a strong relation is found between satellite derived data and the real impacts of fire on forest Note that fire may also damage soils, a variable that was not measured and correlated in this study. The study has confirmed that the time lapse between the fire event and the severity assessment influences not only the correlation but the interpretation of the severity concept, and therefore, a careful adjustment of the time lapse must be performed according to the objectives of the analysis. On late assessments, the growth of herbs and bushes may play down estimation on tree mortality, whereas early assessments can miss relevant processes, as delayed mortality, plant recovery through resprouting, or recruitment of smaller plants protecting soil against erosion. A satellite assessment on fire severity should be therefore planned according to the expected use of the obtained information, and always try to reduce potential misinterpretations by evaluating the post-fire response traits of the affected species, either trees, bushes or herbs. Capítulo 5 Discusión general, Futuras líneas de trabajo y Conclusiones “El tiempo invertido en los arboles nunca es una pérdida de tiempo” Katrina Mayer Capítulo 5 112 Capítulo 5. Discusión general, Futuras líneas de trabajo y Conclusiones Este capítulo presenta una valoración general de las aportaciones realizadas sobre la aplicación de diversas fuentes de información remota en la modelización de los recursos forestales a diferentes escalas en la presente tesis doctoral. Los trabajos realizados han generado herramientas eficientes para la evaluación de estos recursos que permiten ser discutidas de forma general y exponer una serie de conclusiones como síntesis de las que ya han sido mostradas en cada uno de los capítulos de forma independiente, y de las cuales también se derivan posibles líneas de trabajo futuras. Hay que recalcar que esta tesis se enmarca en la empresa föra forest technologies a través del Programa de Investigación de Doctorados Industriales del Ministerio de Economía y Competitividad del Gobierno de España. Así, todas las aportaciones de los trabajos han sido integradas dentro del objetivo principal del programa, es decir, transferir el conocimiento generado en la propia empresa. 5.1 Discusión general La combinación de datos procedentes de distintas fuentes de información, con diferentes niveles de resolución espacial y espectral, ha demostrado un gran potencial para la evaluación de recursos forestales. Las herramientas generadas han sido efectivas para la resolución de diferentes problemas: la identificación de especies a nivel de árbol individual; la clasificación de diferentes tipos de masa en bosques de cobertura completa y; la evaluación de la capacidad de determinar la severidad tras el paso de un incendio. La integración de diferentes fuentes de información incrementa el poder de los modelos generados (Koch 2011). En el caso del capítulo 2, la información LiDAR de baja resolución ha permitido la delineación automática de las copas de los árboles individuales, pero la resolución espacial de la información combinada debe estar acorde con la escala y la unidad mínima de trabajo (Clark et al., 2005). A la escala de árbol individual, la resolución espacial de la información espectral es crítica. Una resolución muy baja generaría ruido por incluir gran cantidad de información externa al objeto, si bien un exceso de resolución espacial podría generar demasiada variabilidad dentro del objeto (Ghosh et al., 2014). Discusión general 113 La recurrencia temporal de la información procedente de sensores remotos permite abordar cambios en los sistemas forestales, tales como la evolución de la cobertura tras la ocurrencia de una perturbación (Chuvieco, 2012; Lentile et al., 2006) o el estado de los ecosistemas y sus recursos (Keeley, 2009). El hecho de estudiar efectos alejados en el tiempo implica recurrir a repositorios (como el de Landsat) que combinen buena resolución espectral y largas series temporales, si bien esto implica cierto sacrificio en cuanto a resolución espacial. En fenómenos ante los que el sistema reacciona con rapidez, el tiempo transcurrido entre el evento a estudiar y la toma de datos puede ser clave en la interpretación del efecto del fenómeno (Harris et al., 2011). La disponibilidad de información periódica es crítica si el objetivo principal es evaluar cambios en la cobertura (capítulo 4) o discriminar coberturas que cambian drásticamente según la estación del año en la que nos encontremos. Por ello, es importante recoger el efecto del paso del tiempo sobre la cobertura, por ejemplo, incluyendo variables que representen esa temporalidad (capítulo 4) o, incluyendo series temporales de imágenes y datos LiDAR en varios momentos en el tiempo con el objetivo de recoger la respuesta de la vegetación a la estacionalidad o el crecimiento (capítulo 3). Si bien es cierto que la recurrencia de la toma de datos, por ejemplo, de LiDAR aerotransportado, tiene un coste económico elevado, ha de ser valorada la posibilidad de realizar vuelos con mayor recurrencia temporal a costa de sacrificar superficie a volar. La resolución espectral es otro aspecto clave a tener en consideración. Si bien grandes clases, como el agua o la vegetación se pueden identificar con facilidad utilizando amplios rangos de longitud de onda, cuando se persigue un análisis más detallado, como pueden ser diferentes tipos de masa o de especies, su discriminación es más compleja, requiriendo de rangos de longitud de onda más estrechos. En estos casos, la disponibilidad de información con alta resolución espectral, como es el caso de las imágenes hiperespectrales, puede ser un limitante (Dalponte et al., 2014). Así, el hecho de ampliar los rangos del espectro puede ayudar a detectar con mayor precisión diferencias en la sensibilidad espectral de los componentes de la vegetación, permitiendo identificar aspectos adicionales como la variación en la densidad del dosel, procesos de senescencia o fenómenos de estrés (Li et al., 2018; Warner et al., 2017). En este sentido, es interesante el potencial del índice NDVI-705, que incluye la respuesta en la banda del red-edge, recogiendo la reflectancia de la vegetación en el rango en el que la clorofila absorbe gran parte de la luz, aportando Capítulo 5 114 diferencias en la separabilidad de las clases (capítulo 3). Del mismo modo, es interesante la estrecha relación demostrada, mediante modelos mixtos, entre el índice RdNBR y ciertas variables fácilmente medibles, entre las que destacan la altura de quemado en la corteza y el lapso de tiempo entre el incendio y su evaluación. Dichos modelos recogen con mayor eficiencia, la importancia de la variabilidad entre incendios y el impacto inducido por el fuego en la vegetación (capitulo 4). El análisis de grandes cantidades de datos requiere de algoritmos cada vez más sofisticados. En este sentido, se ha desarrollado una amplia gama de algoritmos de aprendizaje automático (Nitze et al., 2012; Vega Isuhuaylas et al., 2018; Xu et al., 2018), tanto para resolver problemas de clasificación, como de regresión. La precisión de este tipo de modelos suele ser alta. Sin embargo, la capacidad de explicación de la función modelada por estos algoritmos “caja negra” puede verse limitada por su complejidad. Frente a esto, cabe la alternativa de simplificar este tipo de modelos o buscar otros, más sencillos y fácilmente entendibles por los gestores. Así, se puede llevar a cabo una selección previa de variables con el objetivo de evitar sobreajuste, y reducir la dimensionalidad, lo que puede favorecer la comprensión de estos modelos. Esta simplificación puede realizarse de forma automática (Domingo et al., 2018), a través de herramientas como VSURF (Genuer et al., 2015), o mediante criterio experto, basado en el estudio de las matrices de correlación entre variables (Domingo et al., 2017). En este sentido, los resultados de esta tesis parecen indicar que las variables descriptivas de la estructura o del crecimiento suelen tener mayor importancia en los modelos y, por tanto, aportan mayor poder discriminatorio que las variables espectrales, aunque la información espectral multitemporal también aporta valor, siempre que exista cierta separabilidad espectral entre los grupos. Un aspecto adicional que considerar es la elección entre modelos paramétricos y no paramétricos. A pesar de que los modelos paramétricos son más rápidos y requieren de menos datos, su poder predictivo puede verse limitado, ya que diversos fenómenos ecológicos no responden linealmente (Park et al., 2015; Wu et al., 2018). Por contra, los modelos basados en aprendizaje automático son más complejos y lentos, sobre todo, debido al proceso de entrenamiento al que deben de ser sometidos, pero permiten obtener predicciones más robustas (De’Ath 2017; Zhu et al., 2017). En algunos casos el uso de métodos de regresión Discusión general 115 clásicos puede permitir la obtención de mejores ajustes frente a los modelos no paramétricos, como es el caso del capítulo 4, donde se enfrentan modelos mixtos y modelos de aprendizaje automático (BRT) y donde se observa claramente el efecto de la variabilidad entre incendios. En todo caso, también debe tenerse en cuenta que el efecto del tamaño de la muestra puede ser significativo en cuanto al poder predictivo de modelos paramétricos y no paramétricos. La elección de la técnica más adecuada depende del problema a abordar. De forma general, ninguna técnica se revela como claramente superior a las demás en el estudio de los sistemas ecológicos y la mejor aproximación suele depender de los objetivos (Brosofske et al. 2014). En todo caso, hay que considerar que pequeñas diferencias en los ajustes de los modelos pueden ser importantes en su aplicación en grandes áreas y, por tanto, debe haber un compromiso entre permitir errores en el modelo y que el modelo sea generalista. La disponibilidad de una gran cantidad de datos recogidos por sensores remotos permite comprender procesos que ocurren en grandes áreas. Su uso requiere de un análisis previo de las herramientas a utilizar, considerando la resolución espacial, espectral y temporal, así como el coste económico de la obtención y gestión de dichos datos. Existe información de libre acceso que cubre la superficie global con resoluciones medias y bajas. En cambio, cuando requerimos información a muy alta resolución espacial, la disponibilidad en grandes áreas es menor y normalmente a un coste muy elevado, tanto económico, como computacional. Así mismo, la tendencia parece ir hacia la implementación de modelos cada vez más complejos incluyendo aprendizaje reforzado y actualizaciones periódicas de la información de forma que los ajustes mejoren con el paso del tiempo, aunque esto suponga un aumento de la complejidad de los modelos y del coste computacional. También es importante remarcar que este último aspecto cada vez resulta más fácil de solventar, dado que el almacenamiento y la gestión de los datos es cada vez más eficiente a través de servicios de soportes en la nube para Big Data. La aplicación de estas herramientas en el sector forestal está suponiendo un fenómeno disruptivo, con cambios radicales en prácticas establecidas durante décadas, como son los inventarios forestales, obteniendo modelos con un nivel de detalle muy superior y aportando herramientas clave para la toma de decisiones por el gestor. Futuras líneas de trabajo 117 5.2 Futuras líneas de trabajo De los trabajos realizados durante esta tesis se pueden derivar diferentes propuestas de futuro: Aplicar la metodología empleada para discriminar entre las dos especies de pinos, sobre masas mixtas con mayor diversidad de especies. Realizar la clasificación de especies a nivel de árbol individual utilizando información espectral o SAR de libre acceso, como Sentinel, lo que permitiría valorar si la posible disminución de la capacidad predictiva del modelo puede ser asumida por la disminución del coste económico. Refinar la metodología de individualización de copas para la monitorización de árboles de forma automática, utilizando para ello la segunda cobertura de vuelos LiDAR–PNOA, la cual está siendo llevada a cabo y puesta a disposición de los usuarios actualmente. Desagregar con mayor detalle la discriminación de las tipologías de masa presentes en los bosques de cobertura completa, incluso llegando al nivel de especie. Evaluar el compromiso entre la recurrencia de los vuelos LiDAR-PNOA y la superficie volada mediante vuelos de RPAS con cámaras multiespectrales y LiDAR de cara a realizar monitorización de sistemas forestales. Conclusiones 119 5.3 Conclusiones A continuación, se presentan una serie de conclusiones específicas que se derivan de los objetivos que se han desarrollado en cada uno de los trabajos realizados. 1.- La combinación de datos LiDAR e imágenes de satélite Plèiades para clasificación de especies es más eficiente que el uso de cada una de las fuentes de datos individualmente. Esta combinación ha permitido discriminar a nivel de árbol individual entre dos especies similares en bosques mixtos de coníferas en la cuenca Mediterránea a partir de datos de entrenamiento en masas puras. 2.- La clasificación de las dos especies en masas mixtas puede considerarse satisfactoria pese a la semejanza de ambas especies, especialmente en cuanto a su respuesta espectral. La compensación de los errores a nivel de rodal muestra que la fortaleza del modelo recae en que ha sido entrenado sobre masas puras, donde es inequívoca la caracterización de la especie, y aplicado posteriormente sobre una masa mixta. 3.- La combinación de datos LiDAR e imágenes Sentinel multitemporales ha permitido clasificar las diferentes tipologías de masa en bosques de cobertura completa con alta densidad y continuidad vertical como son los bosques de monteverde. 4.- Las metodologías empleadas son válidas en su aplicación a grandes áreas, tanto a nivel regional como de paisaje, para la evaluación de los recursos. Los trabajos desarrollados aportan herramientas automáticas que mejoran la estimación del valor económico de la producción de los recursos forestales y dan apoyo al gestor en la toma de decisiones. 5.- Las herramientas utilizadas de machine learning: Random Forest (RF); Linear Support Vector Machine (SVML); Radial Support Vector Machine (SVMR) y Artificial Neural Networks (ANN), han sido capaces de clasificar con elevada precisión las diferentes tipologías de monteverde en Tenerife. Referencias “Uno no es lo que es por lo que escribe sino por lo que ha leído” Jose Luis Borges Referencias 128 Referencias Akbari H, Kalbi S (2017) Determining Pleiades satellite data capability for tree diversity modeling. IForest 10:348–352. doi: 10.1016/j.jsv.2018.08.051 Arozena ME, Panareda JM, Rivero B (2015) La herencia del carboneo en el paisaje de la laurisilva canaria. 1535–1542 Assman E (1970) The Principles of Forest Yield Study. Pergamon Press Oxford, New York, NY 506 pp Asta J, Erhardt W, Ferretti M, Fornasier F (2002) Mapping lichen diversity as an indicator of environmental quality. Environment 7:273–279. doi: Article Badreldin N, Sanchez-Azofeifa A (2015) Estimating forest biomass dynamics by integrating multi-temporal Landsat satellite images with ground and airborne LiDAR data in the Coal Valley Mine, Alberta, Canada. Remote Sens 7:2832–2849. doi: 10.3390/rs70302832 Behncke B, Fornaciai A, Neri M, et al (2016) Lidar surveys reveal eruptive volumes and rates at Etna, 2007-2010. Geophys Res Lett 43:4270–4278. doi: 10.1002/2016GL068495 Bergen KM, Dronova I (2007) Observing succession on aspen-dominated landscapes using a remote sensing-ecosystem approach. Landsc Ecol 22:1395–1410. doi: 10.1007/s10980007-9119-1 Blázquez-Casado Á (2017) Modelling methodologies focussed on different machine learning using Rstudio for stand classification and productivity Blázquez-Casado Á, González-Olabarria JR, Martín-Alcón S, et al (2015) Assessing PostStorm Forest Dynamics in the Pyrenees Using High-Resolution LIDAR Data and Aerial Photographs. 12:841–853 Brosofske KD, Froese RE, Falkowski MJ, Banskota A (2014) A review of methods for mapping and prediction of inventory attributes for operational forest management. For Sci 60:733–756. doi: 10.5849/forsci.12-134 Calama R, Tome M, Sanchez-Gonzalez M, et al (2010) Modelling non-wood forest products in Europe: a review. For Syst 19:69–85. doi: 10.5424/fs/201019S-9324 Campbell JB, Wynne RH (2011) Introduction to remote sensing. Guilford Press Carus M (2017) Biobased Economy and Climate Change—Important Links, Pitfalls, and Opportunities. Ind Biotechnol 13:41–51. doi: 10.1089/ind.2017.29073.mca Castillo JAA, Apan AA, Maraseni TN, Salmo SG (2017) Estimation and mapping of aboveground biomass of mangrove forests and their replacement land uses in the Philippines using Sentinel imagery. ISPRS J Photogramm Remote Sens 134:70–85. doi: 10.1016/j.isprsjprs.2017.10.016 Referencias 129 Chrysafis I, Mallinis G, Siachalou S, Patias P (2017) Assessing the relationships between growing stock volume and Sentinel-2 imagery in a Mediterranean forest ecosystem. Remote Sens Lett 8:508–517. doi: 10.1080/2150704X.2017.1295479 Chuvieco, E., (2012). Remote sensing of large wildfires: in the European Mediterranean Basin. Springer Science & Business Media. Clark ML, Roberts DA, Clark DB (2005) Hyperspectral discrimination of tropical rain forest tree species at leaf to crown scales. Remote Sens Environ 96:375–398. doi: 10.1016/j.rse.2005.03.009 Daehne A, Corsini A (2013) Kinematics of active earthflows revealed by digital image correlation and DEM subtraction techniques applied to multi-temporal LiDAR data. Earth Surf Process Landforms 38:640–654. doi: 10.1002/esp.3351 Dalponte M, Ørka HO, Ene LT, et al (2014) Tree crown delineation and tree species classification in boreal forests using hyperspectral and ALS data. Remote Sens Environ 140:306–317. doi: 10.1016/j.rse.2013.09.006 De’Ath, G., 2007. Boosted trees for ecological modeling and prediction. Ecology 88, 243–251 Diéguez-Aranda U, Rojo A. A, Castedo-Dorado F, et al (2009) Herramientas selvícolas para la gestión forestal sostenible en Galicia Domingo D, Lamelas-Garcia M. T, Montealegre-Garcia A., L, De la Riva-Fernández J (2017) Comparison of regression models to estimate biomass losses and CO2 emissions using low-density airborne laser scanning data in a burnt Aleppo pine forest. Eur J of Remote Sens, 50:1, 384-396. Domingo D, Lamelas-Garcia M. T, Montealegre-Garcia A., L, García-Martín A, De la RivaFernández J (2018) Estimation fo total biomass in Aleppo pine forest stands applying parametric and nonparametric methods to low-denity airborne laser scanning data. Forests, 9, 158; doi:10.3390/f9030158. Dupuya S, Lainé G, Tassin J, Sarrailh JM (2013) Characterization of the horizontal structure of the tropical forest canopy using object-based LiDAR and multispectral image analysis. Int J Appl Earth Obs Geoinf 25:76–86. doi: 10.1016/j.jag.2013.04.001 Ediriweera S, Pathirana S, Danaher T, Nichols D (2014) Estimating above-ground biomass by fusion of LiDAR and multispectral data in subtropical woody plant communities in topographically complex terrain in North-eastern Australia. J For Res 25:761–771. doi: 10.1007/s11676-014-0485-7 FAO (2018) the World ’ S Forests. Rome FAO (2015) FOREST EUROPE 2015: State of Europe’s Forest. FAO (2011) FOREST EUROPE, UNECE and FAO 2011: State of Europe’s Forests 2011. Status and Trends in Sustainable Forest Management in Europe Referencias 130 Fisher CT, Cohen AS, Fernández-Diaz JC, Leisz SJ (2017) The application of airborne mapping LiDAR for the documentation of ancient cities and regions in tropical regions. Quat Int 448:129–138. doi: 10.1016/j.quaint.2016.08.050 Garber S (2007) NASA-hystory. https://history.nasa.gov/sputnik/index.html. Accessed 8 Nov 2018 García M, Riaño D, Chuvieco E, Danson FM (2010) Estimating biomass carbon stocks for a Mediterranean forest in central Spain using LiDAR height and intensity data. Remote Sens Environ 114:816–830. doi: 10.1016/j.rse.2009.11.021 García M, Saatchi S, Ustin S, Balzter H (2018) Modelling forest canopy height by integrating airborne LiDAR samples with satellite Radar and multispectral imagery. Int J Appl Earth Obs Geoinf 66:159–173. doi: 10.1016/j.jag.2017.11.017 Genuer R, Poggi JM, Tuleau-Malot C, (2015). VSURF: An R Package for Variable Selection Using Random Forests. https://journal.r-project.org/archive/2015/RJ-2015-018/RJ-2015-018.pdf [8 June 2018]. Ghosh A, Fassnacht FE, Joshi PK, Kochb B (2014) A framework for mapping tree species combining hyperspectral and LiDAR data: Role of selected classifiers and sensor across three spatial scales. Int J Appl Earth Obs Geoinf 26:49–63. doi: 10.1016/j.jag.2013.05.017 Glennie CL, Carter WE, Shrestha RL, Dietrich WE (2013) Geodetic imaging with airborne LiDAR: The Earth’s surface revealed. Reports Prog Phys 76:. doi: 10.1088/0034-4885/76/8/086801 Guerra-Hernández J, Tomé M, González-Ferreiro E (2016) Using low density LiDAR data to map Mediterranean forest characteristics by means of an area-based approach and height threshold analysis. Rev Teledetección 0:103–117. doi: 10.4995/raet.2016.3980 Harris, S., Veraverbeke, S., Hook, S., (2011). Evaluating spectral indices for assessing fire severity in chaparral ecosystems (Southern California) using modis/aster (MASTER) airborne simulator data. Remote Sens. 3, 2403–2419. doi:10.3390/rs3112403 Hassan R, Scholes R, Ash N (2005) ECOSYSTEMS AND HUMAN WELL-BEING: Current State and Trends, Volume 1 Hatten JR (2014) Mapping and monitoring Mount Graham red squirrel habitat with Lidar and Landsat imagery. Ecol Modell 289:106–123. doi: 10.1016/j.ecolmodel.2014.07.004 Heinzel J, Koch B (2012) Investigating multiple data sources for tree species classification in temperate forest and use for single tree delineation. Int J Appl Earth Obs Geoinf 18:101– 110. doi: 10.1016/j.jag.2012.01.025 Hirschmugl M, Gallaun H, Dees M, et al (2017) Methods for Mapping Forest Disturbance and Degradation from Optical Earth Observation Data: a Review. Curr For Reports 3:32–45. doi: 10.1007/s40725-017-0047-2 Hudak AT, Crookston NL, Evans JS, et al (2006) Regression modeling and mapping of coniferous forest basal area and tree density from discretereturn LiDAR and multispectral satellite data. Can J Remote Sens 32:126–138. doi: 10.5589/m06-007 Referencias 131 Immitzer M, Vuolo F, Atzberger C (2016) First experience with Sentinel-2 data for crop and tree species classifications in central Europe. Remote Sens 8:. doi: 10.3390/rs8030166 JRC-EFFIS (2017) Forest Fires in Europe, Middle East and North Africa 2017 Karna YK, Hussin YA, Gilani H, et al (2015) Integration of WorldView-2 and airborne LiDAR data for tree species level carbon stock mapping in Kayar Khola watershed, Nepal. Int J Appl Earth Obs Geoinf 38:280–291. doi: 10.1016/j.jag.2015.01.011 Kimberley MO, Watt MS, Harrison D (2017) Characterising prediction error as a function of scale in spatial surfaces of tree productivity. New Zeal J For Sci 47:. doi: 10.1186/s40490017-0100-8 Keeley, J.E., (2009). Fire intensity, fire severity and burn severity: A brief review and suggested usage. Int. J. Wildl. Fire 18, 116–126. doi:10.1071/WF07049 Knapp N, Fischer R, Huth A (2018) Linking lidar and forest modeling to assess biomass estimation across scales and disturbance states. Remote Sens Environ 205:199–209. doi: 10.1016/j.rse.2017.11.018 Koch B (2011) La teledetección como apoyo a los inventarios forestales nacionales EFN Latifi H, Nothdurft A, Koch B (2010) Non-parametric prediction and mapping of standing timber volume and biomass in a temperate forest: Application of multiple optical/LiDAR-derived predictors. Forestry 83:395–407. doi: 10.1093/forestry/cpq022 Lentile, L.B., Holden, Z.A., Smith, A.M.S., Falkowski, M.J., Hudak, A.T., Morgan, P., Lewis, S.A., Gessler, P.E., Benson, N.C., (2006). Remote sensing techniques to assess active fire characteristics and post-fire effects. Int. J. Wildl. Fire 15, 319–345. doi:10.1071/WF05097 Li C, Zhu X, Wei Y, et al (2018) Estimating apple tree canopy chlorophyll content based on Sentinel-2A remote sensing imaging. Sci Rep 8:1–10. doi: 10.1038/s41598-018-21963-0 Lim K, Treitz P, Wulder M, et al (2003) LiDAR remote sensing of forest structure. Prog Phys Geogr 27:88–106. doi: 10.1191/0309133303pp360ra Lohani B, Ghosh S (2017) Airborne LiDAR Technology: A Review of Data Collection and Processing Systems. Proc Natl Acad Sci India Sect A - Phys Sci 87:567–579. doi: 10.1007/s40010-017-0435-9 Magnussen S, Nord-Larsen T, Riis-Nielsen T (2018) Lidar supported estimators of wood volume and aboveground biomass from the Danish national forest inventory (2012– 2016). Remote Sens Environ 211:146–153. doi: 10.1016/j.rse.2018.04.015 Maltamo M, Næsset E, Vauhkonen J (2014) Forestry Applications of Airborne Laser Scanning: Concepts and Case Studies Marinelli D, Paris C, Bruzzone L (2018) A Novel Approach to 3-D Change Detection in Multitemporal LiDAR Data Acquired in Forest Areas. IEEE Trans Geosci Remote Sens 56:3030–3046. doi: 10.1109/TGRS.2018.2789660 Referencias 132 Matasci G, Hermosilla T, Wulder MA, et al (2018) Three decades of forest structural dynamics over Canada’s forested ecosystems using Landsat time-series and lidar plots. Remote Sens Environ 216:697–714. doi: 10.1016/j.rse.2018.07.024 Meng J, Li S, Wang W, et al (2016) Estimation of forest structural diversity using the spectral and textural information derived from SPOT-5 satellite images. Remote Sens 8: doi: 10.3390/rs8020125 Muhati GL, Olago D, Olaka L (2018) Quantification of carbon stocks in Mount Marsabit Forest Reserve, a sub-humid montane forest in northern Kenya under anthropogenic disturbance. Glob Ecol Conserv 14:e00383. doi: 10.1016/j.gecco.2018.e00383 Mura M, Bottalico F, Giannetti F, et al (2018) Exploiting the capabilities of the Sentinel-2 multi spectral instrument for predicting growing stock volume in forest ecosystems. Int J Appl Earth Obs Geoinf 66:126–134. doi: 10.1016/j.jag.2017.11.013 Næsset E, Gobakken T, Holmgren J, et al (2004) Laser scanning of forest resources: The nordic experience. Scand J For Res 19:482–499. doi: 10.1080/02827580410019553 Naidoo L, Cho MA, Mathieu R, Asner G (2012) Classification of savanna tree species, in the Greater Kruger National Park region, by integrating hyperspectral and LiDAR data in a Random Forest data mining environment. ISPRS J Photogramm Remote Sens 69:167–179. doi: 10.1016/j.isprsjprs.2012.03.005 Nitze I, Schulthess U, Asche H (2012) Comparison of machine learning algorithms random forest, artificial neuronal network and support vector machine to maximum likelihood for supervised crop type classification. Proc tje 4th GEOBIA, May 7-9 - Rio Janeiro - Brazil 35–40. doi: 10.2355/isijinternational.49.425 Nordkvist K, Granholm A-H, Holmgren J, et al (2012) Combining optical satellite data and airborne laser scanner data for vegetation classification. Remote Sens Lett 3:393–401. doi: 10.1080/01431161.2011.606240 Onojeghuo AO, Onojeghuo AR (2017) Object-based habitat mapping using very high spatial resolution multispectral and hyperspectral imagery with LiDAR data. Int J Appl Earth Obs Geoinf 59:79–91. doi: 10.1016/j.jag.2017.03.007 Palahí M, Pukkala T, Bonet A, et al (2009) Management of Even-Aged Pine Stands of Catalonia Par H, Jeong S J, Ho C H, Kim J, Brown M E, Schaepman M E (2015). Nonlinear response of vegetation green-up to local temperature variations in temperate and boreal forests in the Northern Hemisphere. Remote Sensing of Environment, 165, 100-108. Pearse GD, Dash JP, Persson HJ, Watt MS (2018) Comparison of high-density LiDAR and satellite photogrammetry for forest inventory. ISPRS J Photogramm Remote Sens 142:257–267. doi: 10.1016/j.isprsjprs.2018.06.006 Peura M, Silveyra Gonzalez R, Müller J, et al (2016) Mapping a ‘cryptic kingdom’: Performance of lidar derived environmental variables in modelling the occurrence of forest fungi. Remote Sens Environ 186:428–438. doi: 10.1016/j.rse.2016.09.003 Referencias 133 Popescu SC (2007) Estimating biomass of individual pine trees using airborne lidar. Biomass and Bioenergy 31:646–655. doi: 10.1016/j.biombioe.2007.06.022 Popescu SC, Wynne RH, Nelson RF (2003) Measuring individual tree crown diameter with lidar and assessing its influence on estimating forest volume and biomass. Can J Remote Sens 29:564–577. doi: 10.5589/m03-027 Potapov P V., Turubanova SA, Tyukavina A, et al (2015) Eastern Europe’s forest cover dynamics from 1985 to 2012 quantified from the full Landsat archive. Remote Sens Environ 159:28–43. doi: 10.1016/j.rse.2014.11.027 Rego C, Boieiro M, Gonçalves Y, et al (2014) The drosophilids (Diptera: Drosophilidae) from a Laurisilva patch in Madeira with two new records for this island. Bocagiana 3141:1–8 Rodriguez-Veiga P, Saatchi S, Wheeler J, et al (2017) Methodology for Regional to Global Mapping of Aboveground Forest Biomass: Integrating Forest Allometry, Ground Plots, and Satellite Observations. Earth Obs L Emerg Monit 7 Rodríguez F, Lizarralde I (2015) Inventario forestal con tecnología LiDAR, ¡Más barato y más preciso! Forestalis 24:33–34 Sarrazin MJD, Aardt JANV, Asner GP, et al (2012) Fusing small-footprint waveform LiDAR and hyperspectral data for canopy-level species classification and herbaceous biomass modeling in savanna ecosystems. Can J Remote Sens 37:653–665. doi: 10.5589/m12-007 Sasaki T, Imanishi J, Ioki K, et al (2012) Object-based classification of land cover and tree species by integrating airborne LiDAR and high spatial resolution imagery data. Landsc Ecol Eng 8:157–171. doi: 10.1007/s11355-011-0158-z Seidl R, Schelhaas M, Rammer W, Verkerk PJ (2015) Increasing forest disturbances in Europe and their impact carbon storage. 4:806–810. doi: 10.1038/nclimate2318.Increasing Shi Y, Skidmore AK, Wang T, et al (2018) Int J Appl Earth Obs Geoinformation Tree species classi fi cation using plant functional traits from LiDAR and hyperspectral data. Int J Appl Earth Obs Geoinf 73:207–219. doi: 10.1016/j.jag.2018.06.018 Stereńczak K, Lisańczuk M, Erfanifard Y (2018) Delineation of homogeneous forest patches using combination of field measurements and LiDAR point clouds as a reliable reference for evaluation of low resolution global satellite data. For. Ecosyst. 5:1 Sullivan FB, Ducey MJ, Orwig DA, et al (2017) Comparison of lidarand allometry-derived canopy height models in an eastern deciduous forest. For Ecol Manage 406:83–94. doi: 10.1016/j.foreco.2017.10.005 Tesch SD (1980) The evolution of forest yield determination and site classification. For Ecol Manage 3:169–182. doi: 10.1016/0378-1127(80)90014-6 Thers H, Brunbjerg AK, Læssøe T, et al (2017) Lidar-derived variables as a proxy for fungal species richness and composition in temperate Northern Europe. Remote Sens Environ 200:102–113. doi: 10.1016/j.rse.2017.08.011 Referencias 134 Thom D, Seidl R (2016) Natural disturbance impacts on ecosystem services and biodiversity in temperate and boreal forests. Biol Rev Camb Philos Soc 91:760–781. doi: 10.1111/brv.12193 Tielidze LG (2015) Glaciers change over the last century, Caucasus Mountains, Georgia, observed by the old topographical maps, Landsat and ASTER satellite imagery. Cryosph Discuss 9:3777–3806. doi: 10.5194/tcd-9-3777-2015 Tonolli S, Dalponte M, Neteler M, et al (2011) Fusion of airborne LiDAR and satellite multispectral data for the estimation of timber volume in the Southern Alps. Remote Sens Environ 115:2486–2498. doi: 10.1016/j.rse.2011.05.009 Urban M, Berger C, Mudau TE, et al (2018) Surface moisture and vegetation cover analysis for drought monitoring in the southern Kruger National Park using Sentinel-1, Sentinel-2, and Landsat-8. Remote Sens 10:. doi: 10.3390/rs10091482 Vanclay J.K. (1994) Modelling forest growth and yield: applications to mixed tropical forestl. School of Environmental Sicence and Management Papers, 537. Vauhkonen J (2018) Predicting the provisioning potential of forest ecosystem services using airborne laser scanning data and forest resource maps. For Ecosyst 5:1–19. doi: 10.1186/s40663-018-0143-1 Vázquez-Piqué J, Pereira H (2008) What to take into account to develop cork weight models?: Review and statistical considerations. Inst Nac Investig y Tecnol Agrar y Aliment 17:199–215 Vega Isuhuaylas LA, Hirata Y, Ventura Santos LC, Serrudo Torobeo N, (2018). Natural forest mapping in the Andes (Peru): a comparison of the performance of machine-learning algorithms. Remote Sens 10: 782. Vepakomma U, St-Onge B, Kneeshaw D (2008) Spatially explicit characterization of boreal forest gap dynamics using multi-temporal lidar data. Remote Sens Environ 112:2326– 2340. doi: 10.1016/j.rse.2007.10.001 Verhegghen A, Eva H, Ceccherini G, et al (2016) The potential of sentinel satellites for burnt area mapping and monitoring in the Congo Basin forests. Remote Sens 8:1–22. doi: 10.3390/rs8120986 Wang H, Glennie C (2015) Fusion of waveform LiDAR data and hyperspectral imagery for land cover classification. ISPRS J Photogramm Remote Sens 108:1–11. doi: 10.1016/j.isprsjprs.2015.05.012 Warner, T.A., Skowronski, N.S., Gallagher, M.R., (2017). High spatial resolution burn severity mapping of the New Jersey Pine Barrens with WorldView-3 near-infrared and shortwave infrared imagery. Int. J. Remote Sens. 38, 598–616. doi:10.1080/01431161.2016.1268739 Wong J, Prokofieva I (2014) Report presnting synthesis of regional sectorial reviews to describe the “State of the European NWFP” StarTree deliverable D1.3. 96 pp Referencias 135 Wu B, Yu B, Wu Q, et al (2016) Individual tree crown delineation using localized contour tree method and airborne LiDAR data in coniferous forests. Int J Appl Earth Obs Geoinf 52:82– 94. doi: 10.1016/j.jag.2016.06.003 Wu C, Tao H, Zhai M, Lin Y, Wang K, Deng J, Yang H (2018). Using nonparametric modeling approaches and remote sensing imagery to estimate ecological welfare forest biomass. Journal of Forestry Research, 29(1), 151-161. Wulder MA, White JC, Alvarez F, et al (2009) Characterizing boreal forest wildfire with multitemporal Landsat and LIDAR data. Remote Sens Environ 113:1540–1555. doi: 10.1016/j.rse.2009.03.004 Wulder MA, White JC, Cranny M, et al (2008) Monitoring Canada’s forests. Part 1: Completion of the EOSD land cover project. Can J Remote Sens 34:549–562. doi: 10.5589/m08-066 Xu C, Manley B, Morgenroth J (2018) Evaluation of modelling approaches in predicting forest volume and stand age for small-scale plantation forests in New Zealand with RapidEye and LiDAR. Int J Appl Earth Obs Geoinf 73:386–396. doi: 10.1016/j.jag.2018.06.021 Xu Q, Man A, Frederickson MM, et al (2017) Quantification of the uncertainty in the aboveground biomass estimates derived from small-footprint airborne LiDAR. Prep 216:514–528. doi: 10.1016/j.rse.2018.07.022 Youssef AM, Pourghasemi HR, El-Haddad BA, Dhahry BK (2016) Landslide susceptibility maps using different probabilistic and bivariate statistical models and comparison of their performance at Wadi Itwad Basin, Asir Region, Saudi Arabia. Bull Eng Geol Environ 75:63– 87. doi: 10.1007/s10064-015-0734-9 Zhao K, Suarez JC, Garcia M, et al (2018) Utility of multitemporal lidar for forest and carbon monitoring: Tree growth, biomass dynamics, and carbon flux. Remote Sens Environ 204:883–897. doi: 10.1016/j.rse.2017.09.007 Zhu XX, Tuia D, Mou L, Xia GS, Zhang L, Xu F, Fraundorfer F, 2017. Deep Learning in Remote Sensing: A Comprehensive Review and List of Resources. IEEE Geosci Remote Sens Mag 5: 8-36