Influence of pansharpening in obtaining accurate vegetation maps

CANADIAN JOURNAL OF REMOTE SENSING https://doi.org/./.. Influence of Pansharpening in Obtaining Accurate Vegetation Maps Edurne Ibarrola-Ulzurruna,b, Consuelo Gonzalo-Martín b, and Javier Marcelloa aInstituto de Oceanografía y Cambio Global (IOCAG), Universidad de Las Palmas de Gran Canaria (ULPGC), Parque Científico Tecnológico Marino de Taliarte, s/n, , Telde, Las Palmas, Spain; bEscuela Técnica Superior de Ingenieros Informáticos, Universidad Politécnica de Madrid, UPM, Campus de Montegancedo, , Pozuelo de Alarcón, Madrid, Spain ARTICLE HISTORY Received  February  Accepted  August  ABSTRACT In recent decades, there has been a decline in ecosystem services. Thus, the development of reliable methodologies to monitor ecosystems is becoming important. In this context, the availability of very high resolution sensors offer practical and cost-effective means for good environmental management. However, improvements in the data received are becoming necessary to obtain higher quality information in order to get reliable thematic maps. One improvement is pansharpening, which enhances the spatial resolution of the multispectral bands by incorporating information from a panchromatic image. The main goal of this work was to assess the influence of pansharpening techniques in obtaining precise vegetation maps. Thus, pixeland object-based classification techniques were implemented and applied to fused imagery using different pansharpening algorithms. Worldview-2 high resolution imagery was used due to its excellent spatial and spectral characteristics. The Teide National Park, in The Canary Islands (Spain), was chosen as the study area since it is a vulnerable heterogeneous ecosystem. The vegetation classes of interest considered were established by the National Park conservation managers. Weighted Wavelet ‘à trous’ through Fractal Dimension Maps pansharpening algorithm demonstrated a superior performance in the image fusion preprocessing step, while the most appropriate classifier to generate accurate vegetation thematic maps in heterogenic and mixed ecosystems was the Bayes method after the segmentation stage, even though Support Vector Machine achieved the highest overall accuracy. RÉSUMÉ Au cours des dernières décennies, les ressources naturelles ont diminué. Pour ces raisons, le développement de méthodologies fiables de surveillance des écosystèmes devient de plus en plus important. Dans ce contexte, la disponibilité de satellites à très haute résolution offre des moyens pratiques et rentables pour une bonne gestion environnementale. Cependant, il est d’abord nécessaire d’introduire des améliorations dans l’acquisition de données afin d’obtenir des informations de meilleure qualité permettant la création des cartes thématiques fiables. Une de ces améliorations est le “pansharpening”, qui augmente la résolution spatiale des bandes multispectrales en incorporant de l’information à partir de l’image panchromatique. Le but principal de ce travail était d’évaluer l’influence des techniques de pansharpening dans l’obtention de cartes de végétation précises. Pour ce faire, on a mis en place des techniques de classification par pixel et orientée-objet, et on les a appliquées à l’imagerie fusionnée en utilisant différents algorithmes de “pansharpening”. L’imagerie à haute résolution Worldview-2 a été utilisée en raison de ses excellentes caractéristiques spatiales et spectrales. On a choisi le Parc National du Teide, aux Canaries (Espagne), comme zone d’étude, étant donné qu’il s’agit d’un écosystème vulnérable et hétérogène. Les classes de végétation considérées ont été établies par les responsables de la conservation du Parc National. L’algorithme de pansharpening “Weighted Wavelet ‘à trous’ through Fractal Dimension Maps”a démontré la précision la plus élevée à l’étape de prétraitement de fusion d’images, alors que le classificateur le plus approprié pour générer des cartes précises de végétation, dans les écosystèmes hétérogènes et mixtes, a été la méthode Bayes, après l’étape de segmentation, bien que Support Vector Machine a obtnenu le plus haut précision globale. Introduction Biodiversity supports a wide variety of ecological functions together with the services provided by ecosystems (Isbell et al. 2011).Itsconservationisrelatedtoglobal environmental changes, such as the land use changes, CONTACT Edurne Ibarrola-Ulzurrun edurne.ibarrola@alu.ulpgc.es climate change, and sustainable developments. However,duringthelastcentury,humanactivitygaveraise to alterations in ecosystems and thus, biodiversity is suffering a fast decline (Khare and Ghosh 2016). This creates a demand to preserve environmental resources Copyright © CASI Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018 2E. IBARROLA-ULZURRUN ET AL. (Pakzad et al. 2003). Remote sensing has become an essential tool for evaluating and monitoring ecosystems. It is able to provide consistent data on the Earth at various scales, reducing the collection of field data and ground-based observations (Khare and Ghosh 2016), making it a cost-effective tool for land-cover classification, monitoring, and environmental management (Aplin 2004). Procedures for both classifying land cover and monitoring land cover changes are extensively used in environmental management. However, classifying remotely sensed data in a thematic map remains a challengebecauseofthemanyfactorsinvolved,whichmay affect the success of the classification (Lu and Weng 2007), i.e., the complexity and heterogeneity of small species in the ecosystem, quality of the remotely sensed data available,aswellastheimagepreprocessingandclassification approaches. Remote sensing classification involves the selection of training samples, image pre-processing, a suitable classification model, post-classification processing, and accuracy assessment (Lu and Weng 2007). In this context, pixel-based and object-based classification approaches can be distinguished. In the traditional pixel-based classification approaches, each pixel is identified by comparing its spectral signature value with the training samples and that pixel is labeledwiththeproperclassbasedonacertainsupervised algorithm (Gao et al. 2009), but neither spatial concepts nor contextual information are incorporated. Thus, the misclassification rate could be high due to the spectral similarity characteristics of some classes, the spectral variability of the canopy reflectance and the presence of mixed pixels located on the boundary between classes (Peña-Barragán et al. 2011). On the other hand, object-based image analysis or OBIA, in the geographic context (GEOBIA), was proposed as a new classification approach due to the launch of very high resolution (VHR) sensors. It offers a methodological framework, from a machine-based interpretation of complex classes, defined by the spectral, spatial, structural and hierarchical properties (Blaschke et al. 2008; Goodin et al. 2015, Oruc et al. 2004). The idea of OBIA is the segmentation of the image followed by successive analyses, usually at different hierarchical levels, in order to create relationships within segments or objects (Garcia-Pedrero et al. 2015; Lantz and Wang, 2013; Peña-Barragán et al. 2011). Given that, segmentation could be defined as a bottomup region-merging process in which the image is subdivided into homogeneous regions according to several parameters (band weights, scale, color, shape, texture, etc.) defined by the operator, with the objective of creating object delimiting borders (Peña-Barragán et al. 2011). However, segmentation could be a challenge because the quality of the classification results depends largely on the segments obtained, thus, it is necessary to establish a balance between efficacy and efficiency (Garcia-Pedrero et al. 2015). As previously mentioned, remote sensing imagery requires a number of corrections and enhancements, before starting with the classification approach. Since, VHR sensors provide both a multispectral image and a panchromatic band, the quality enhancement would be carried out using an image fusion technique (pansharpening algorithm) in which the spatial resolution of the multispectral image can be improved by incorporating information from the panchromatic image (Carper et al. 1990; Chavez et al. 1991;Shettigara1992). Extensive research into image fusion has been carried out during the last decades. Image fusion can be categorized into 3 levels: pixel level, feature level, and knowledge or decision level. Specifically, pansharpening is performed at pixel level and ideal pansharpening algorithm should have 2 main attributes (Li et al. 2012): (i) enhancing high spatial resolution; and (ii) reducing spectral distortion. The simplest pansharpening methods, at the conceptual and computation level, are Intensity-Hue-Saturation (IHS; Carper et al. 1990;Shettigara1992), Principal Component Analysis (PCA; Chavez et al. 1991) and Brovey Transforms (Gillespie et al. 1987). However, these techniques have problems becausethey providespectral distortedfused images.New approaches such as Wavelet transformation and the High Pass Filtering (Alimuddin et al. 2012;Alparoneetal. 2016;Chavezetal.1991; González-Audícana et al. 2004; Kpalma et al. 2014;Marcelloetal.2013;Pohl2014;Vivone et al. 2015) have been proposed to address particular problems with the previous techniques (Aiazzi et al. 2002, 2006; Nuñez et al. 1999). In addition, images acquired from VHR sensors require atmospheric corrections to transform the top of atmosphere radiance into ground reflectance by removing the atmosphere absorption and scattering effects (Marcello et al. 2016), and image orthorectification, as the topographic relief decreases or increases the radiance illuminated because of the difference in land elevation in some areas. Once images have been corrected and enhanced, the generation of products for the management of natural resources can be carried out. In this context, the main goal in this study was to implement and assess pixeland object-based classifiers on VHR imagery that were fused using different pansharpening techniques. Thus, the influence of fusion on the classification results was analyzed. The ultimate purpose was the development of a reliable processing methodology which, when applied to remote sensing imagery, serves to obtain accurate vegetation maps and information for the conservation of natural resources. Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018 CANADIAN JOURNAL OF REMOTE SENSING 3 Figure . Study area of Teide National Park (Canary Islands, Spain). Specifically,a very complexshrublandecosystemhasbeen selected to perform the analysis. The article is structured as follows: The 2nd section includes the study area and a description of the data sets, the image fusion methods, and the different classification techniques applied at pixeland object-based levels. The analysis and evaluation of the different techniques, as well as the thematic maps obtained are presented in the 3rd section. Finally, the last section includes a critical analysis of the results and summarizes the main outcomes and contributions. Material and methods Study area and dataset A heterogeneous ecosystem of The Canary Islands (28° 06N; 16°33W) was selected for this study: the Teide National Park located on the island of Tenerife (Figure 1). The Teide National Park, created in 1954, is a protected area with 18,990 hectares (Martín-Osorio et al. 2005), located on the island of Tenerife. The Teide National Park area is a giant crater situated in the center of the island. Teide is made up by several overlapping volcanoes in the middle of the giant crater, making it the highest mountain in Spain (3,718 m; González-Lemus et al. 2009). As already mentioned, the vegetation is vulnerable to environmental changes. Thus, the plants respond to thermic and hydric stress, a characteristic of a mountain ecosystem with a shrub physiognomy (Arozena-Concepción and Beltrán-Yanes 2006). The following non-herbaceous vegetation species were selected for the study by the experts of the National Park, due to their abundance and importance at the ecological level: Spartocytisus supranubius (Teide broom), Pterocephalus lasiospermus (Rosalillo de cumbre), Descurainia bourgaeana (Hierba pajonera), and Pinus canariensis (Canarian pine; Figure 2). All of them are addressed in this article as “vegetation classes.” Moreover, urban, road, and bare soil classes were also included in the classification. TheTeidebroom(Spartocytisus supranubius;Bonet et al. 2009) is a broom-like shrub reaching a height of 2 m–3 m, moreover it is a strong competitor and its growth is a major factor affecting the composition of plant species (Kyncl et al. 2006). S. supranubius is 1 of the most important plant species as is Pterocephalus lasiospermus (Rosalillo de cumbre) and Descurania bourgaeana (Hierba pajonera; Arozena-Concepción and Beltrán-Yanes 2006). P. lasiospermus is a small shrub characterized by pink flowers appearing in the spring and D. bourgaeana is a shrub that can be easily distinguished by its semi-spherical shape and its flowers of yellow petals. Another significant plant species is the Canary pine (Pinus canariensis),theonlynativepinein the archipelago, which is found in the northern area of the park. This species has a great resistance to fire and cold temperatures (Garzón-Machado et al. 2011;Ottoetal. 2012;WildpretdelaTorre2001). Other similar shrubland ecosystems can be identified around the world, for example: Pico do Pico in the Azores, Mt. Halla in South Korea, and Hawaii or the Galapagos Islands, some areas Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018 4E. IBARROLA-ULZURRUN ET AL. Figure . Vegetation field campaign in the Teide National Park: (a) Spartocytisussupranubius;(b)Pterocephalus lasiospermus;(c)Descurainia bourgaeana;and(d)Pinus canariensis. of Okanagan Desert in Canada. Thus, the importance of studying this ecosystem is its similarity to other Mediterranean, temperate or tropical parts of the world (Ruocco et al. 2014). Worldview-2 (WV-2) orthoready imagery was used in the study. The WV-2 satellite, launched by DigitalGlobe on October 8, 2009, was the first commercial satellite to have a very high spatial resolution sensor with 1 panchromaticand8multispectralbands.TheTeideimage was acquired on May 16, 2011, in the spring season as the vegetation species has greater spectral separability (González-Lemus et al. 2009; Wildpret de la Torre 2001). Image corrections and enhancements Numerous methodological problems are associated with the use of data from VHR sensors (Franklin and Wulder 2002), thus, several image enhancements must be carried out in order to reduce these problems andtoobtainhighqualityproductswhichwillallow a comprehensive and useful analysis of the natural resources. Pansharpening techniques The first step was to generate a pixel-level fused image, with a high spatial and spectral quality, using a pansharpening process. The enhancement is important in this type of heterogeneous ecosystem due to the small size of some vegetation classes. After a comprehensive review of the state-of-art (Ibarrola-Ulzurrun et al. 2017), 7 pansharpening methods were applied (for detailed information see the references): rGram-Schmidt (GS): A low-resolution PAN image is simulated and the GS transformation is performed on the simulated PAN plus the MS bands. Then, theoriginalPANreplacesthefirstGSbandand the inverse transform is applied (Laben and Brower 2000). rFast Intensity Hue Saturation (FIHS): It uses the spectral bands and the IHS transformation to estimate the new component I. The spatial detail is extracted, computing the difference between the panchromatic band and I.Finally,itisinjectedinto themultispectralbands(Marcelloetal.2013;Tuetal. 2001;Vivoneetal.2015). rHyperspherical Color Sharpening (HCS): It is an algorithm specifically designed for the WV-2 sensor, based on the transformation between any native color space and the hyperspherical color space (Li et al. 2012,2013; Padwick et al. 2010;Tuetal.2012; Wu et al. 2015). rBased Modulation Transfer Function (MTF): The trendy implementations provided by Vivone et al. 2015 (http://openremotesensing.net/knowledgeba se/a-critical-comparison-among-pansharpeningalgorithms/) of 2 different MTF fusion algorithms were used: the one based on Generalized Laplacian Pyramid (MTF_GLP; Aiazzi et al. 2002,2006)and the Generalized Laplacian Pyramid plus a High Pass Modulation (MTF_GLP_HPM; Burt and Adelson 1983; Aiazzi et al. 2002,2006;Amroetal.2011). rWavelet ‘à trous’ (WAVE_ATROUS): The fused images are obtained by adding the panchromatic wavelet planes to the multispectral image approximation (Amolins et al. 2007;Amroetal.2011; Dutilleux et al. 1987; Gonzalo-Martín and LilloSaavedra 2008; Lillo-Saavedra and Gonzalo 2006; Marcello et al. 2013;Waldetal.1997). rWeighted Wavelet ‘à trous’ method through Fractal Dimension Maps (WATࣹFRAC; Lillo-Saavedra and Gonzalo 2006): It is based on the wavelet ‘à trous’.A mechanism that controls the trade-off between the spatial and spectral quality by introducing a weighting factor for each band αi(x,y)forthepanchromatic wavelet coefficients is established and defined as Fractal Dimension Map with the same size as the original image (Lillo-Saavedra et al. 2011). These maps are generated for each of the source images with the box-counting algorithm and by applying a windowingprocess.Eachelementinthesemapsprovides a different weighting value for each pixel and each band. Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018 CANADIAN JOURNAL OF REMOTE SENSING 5 Quality evaluation is a fundamental issue to benchmark and optimize different pansharpening algorithms (Marcello et al. 2013; Rodríguez-Esparragón 2015;Wald 2000;WangandBovik2002). A visual and quantitative assessment was undertaken in order to evaluate the pansharpening results in the different fused images. The quality assessment was carried out for the whole set of bands. In the study, a number of statistical evaluation indices were used to measure the quality of the fused images at thespectral,spatialandgloballevel(Alparoneetal.2008; Kpalma et al. 2014;ShridharandAlvarinho2013). The indicesappliedtomeasurethespectralqualityofthe fused images, which take the original multispectral bands as a reference, were: Spectral Angle Mapper (Kruse et al. 1993; best value closer to 0) and Spectral ERGAS (Wald, 2000; best values closer to 0, but usually between 0–3). The spatial quality assesses the spatial detail injected and the panchromatic band was chosen as a reference. The indices considered were: Spatial ERGAS (Lillo-Saavedra et al. 2005; best values closer to 0, but usually between 0– 3); Frequency Comparison (Rodríguez-Esparragón et al. 2014; values between 0–1, best closer to 1); and Zhou Index (Zhou et al. 1998; values between 0–1, best closer to 1). Finally, at global level the Q8 index, which is a generalization of the Q index, (Wang and Bovik 2002)andofthe Q4 index (Alparone et al. 2004), with values between 0–1 (best closer to 1), was used to measure correlation, mean shifts, and radiometric distortion simultaneously (Marcello et al. 2013). Asthe Teide ecosystemisveryheterogeneousandcomposed by small shrubs, pansharpening will be very important to achieve reliable vegetation maps. Also, both qualities will be critical, the spatial to discriminate the small plants and the spectral to differentiate similar covers. As indicated in Thomas et al. (2008), the ideal fusion method does not exist yet, but the fused results generally correspond to a tradeoff between a good geometrical representation of structures and a good representation of spectral information. Orthorectification and atmospheric correction After detailed assessment of different image and modelbased atmospheric corrections algorithms using WV-2 imagery and field spectroradiometer data collected simultaneously (Marcello et al. 2016), the FLAASH algorithm (Adler-Golden et al. 1999)achievedthebestperformance. The appropriate atmospheric parameters were adjusted (atmospheric model, aerosol model, aerosol optical thickness, adjacency, etc.) using climatologic information and fielddataandthe average reflectanceestimationsachieved a Root Mean Square Error (RMSE) around 3% when compared to the in-situ measurements. Orthorectification,usingtheRationalPolymodalCoefficientsmodel, was carried out using ENVI 5.0 software (ENVI 2004)and the average RMSE after the orthorectification was 3.11 m forgeodesicpointslocatedinthepark. Segmentation at object-based approach The OBIA process starts with a segmentation of the input images into local groups of pixels (segments) that become spatial units in the later classifications and accuracy assessment. The object shape, size and spectral properties depend on both the segmentation approach and the research goals. The ideal purpose of the segmentation is to approximate meaningful landscape entities recognizable at a given image resolution (Dronova 2015). Image segmentation in eCognition is not a straightforward process and finding useful segmentation levels is an iterative process (Walsh et al. 2008). The most popular segmentationcarriedoutbyeCognitionusers(Burnettetal.2003; Dronova 2015;LuandWeng2007; Walsh et al. 2008)is Multiresolution Segmentation. It is regarded as a regionbased algorithm, which starts by considering each pixel as a separate object. Subsequently, pairs of image segments are merged to form bigger segments. The decision to merge is based on the homogeneity criterion, which measures how homogeneous or heterogeneous an image object is within itself. The derived image objects in eCognition are determined by the following parameters (Baatz et al. 2001): rWeight of image channels to assign different importance to each image band. rScale to determine the maximum heterogeneity of the objects allowed. Smaller scales increase the dimensionality and the division of the objects into the sub-groups, while larger scales combine the multi-segments into 1 (Marangoz et al. 2006). rShape to adjust the homogeneity of the object generation. rCompactness to determine whether the objects will become more compact or smoother. Once the Multiresolution Segmentation was performed, Spectral Difference algorithmwasusedtomergeneighboring image objects when the difference between their layers mean intensities are below the value given by the maximum spectral difference (Baatz et al. 2001). Classification models Several classification techniques were applied in the study. The first step was to determine the classes appearing in the image and to obtain a set of training and testing samples. Thesizeandrepresentativenessofthesetofthetraining samples are critical for image classification (Lu and Weng 2007). The second step was to determine the classification Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018 6E. IBARROLA-ULZURRUN ET AL. Table . Pixel-based and object-based classification algorithms and their parameters. Classification Approach Classification Algorithm Parameters Pixel-based approach Maximum Likelihood Threshold: . Mahalanobis Distance Maximum distance error: None Support Vector Machine Kernel: Radial basis function Gamma in kernel function: . Penalty parameter:  Object-based approach Bayes eCognition does not allow to set parameters with this classifier Nearest Neighbor eCognition does not allow to set parameters with this classifier K-Nearest Neighbor K:  Support Vector Machine Kernel: Radial basis function Gamma in kernel function: . Penalty parameter:  model to be used and which classifier apply in each model. An evaluation of the parameters chosen in each classifier was performed. However, these parameters are image dependent, and the user should analyze which parametersaremoresuitablefortheirimages.The2classification approaches considered here were pixel-based, which was carried out using ENVI 5.0 image processing software (Exelis Visual Information Solutions, Inc., a subsidiary of Harris Corporation); and object-based using eCognition Developer (Trimble Geospatial). eCognition is optimized for the cost-effective OBIA classification of VHR imagery. The classification models selected for carrying out the pixel-based classification approach are Maximum Likelihood, Mahalanobis Distance, and Support Vector Machine. Regarding the OBIA approach, Bayes, Nearest Neighbor, K-Nearest Neighbor, and Support Vector Machinewereselected(Table 1). For choosing the classifiers’ parameters, different trials were performed and evaluated until obtaining the most suitable parameter value for each classifier. Thematic maps were obtained after applying the different classification techniques using the same training samples. Afterwards, the accuracy of the classification was measured by using the testing samples collected. In order to obtain a reliable thematic map, the training and testing samples were selected with the help of Teide Figure . True color images of the Teide National Park: (a) original multispectral image; and fused imagery using: (b) GS; (c) FIHS; (d) HCS; (e) MTF_GLP; (f) MTF_GLP_HPM; (g) WAVE_ATROUS; and (h) WATࣹFRAC. Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018 CANADIAN JOURNAL OF REMOTE SENSING 7 Figure . RGB color composite: (a) WATࣹFRAC fused image and (b) WAVE_ATROUS fused image; and thematic maps using: (c) Maximum Likelihood in WATࣹFRAC fused image; (d) Mahalanobis Distance in WAVE_ATROUS fused image; and (e) Support Vector Machine in WATࣹFRAC fused image. National Park experts from well-known sites around the park and using the ground truth samples obtained during thefieldworkcarriedoutinthestudyarea,whichwere used for the training and testing phases in the classification process. The method used for choosing the samples was a random method in which a total of 362,631 pixels (ca. 70% of the total samples) were chosen as training samples and 159,182 pixels (ca. 30% of the total samples) for the testing. The statistical accuracy assessment technique used in the study was the standardized confusion Error Matrix which reports 2 measurements of accuracy: Overall Accuracy and Kappa coefficient (Congalton 1991). It is more appropriate in difficult classification approaches, assuming that the map categories are mutually exclusive and exhaustive and that each location belongs to a single category. On the other hand, the Kappa coefficientisameasureofoverallstatisticalagreement,which takes non-diagonal elements into account (Lu and Weng 2007). Results and discussion The following results are presented as: (i)analysisand evaluation of the fused images quality obtained through the different pansharpening techniques; (ii) accuracy of Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018 8E. IBARROLA-ULZURRUN ET AL. Figure . Classification overall accuracy in percentage (%) of each class and classifier in the fused imagery. X axis: classes; Y axis: percentage value. the pixel-based thematic maps obtained for each fused image; and (iii) accuracy of the OBIA classification results obtained for each fused image. Pansharpening results In this section, a visual and a quantitative assessment of the fused images are presented. To facilitate the visual inspection and for a detailed spatial analysis, a zoom of thecompletesceneisshowninFigure 3. The visual interpretation at the spectral level in the image indicates that every algorithm seems to correctly preserve the spectral information in the fused image. Moreover, the differences are clearer spatially than spectrally. For HCS and MTF-based techniques, although they maintain the spectral information well, the spatial details are not satisfactorily injected, thus not achieving a good spatial enhancement. This smoothed aspect appearing in the image is because the algorithm makes uniform the areas with a common texture. Thus, Dimensional Fractal Maps are calculated into the MS and the PAN image, giving information regarding roughness in the image areas. Then, a small amount of the spatial information is incorporated in homogeneous areas, maintaining a more homogeneous spectral response, closer to the original multispectral image, avoiding in this way, the ‘salt&pepper’ effect. As it was mentioned, in order to carry out an objective evaluation of the pansharpening techniques, 6 quality indices were computed. The results are presented in Table 2. Spectral Quality indices and Q8 confirm that MTF methods provide better spectral performance while WAVE_ATROUS gets, in general, the lowest spectral quality, even though there is not great difference between the highest and the lowest result. As regards the spatial performance, WAVE_ATROUS is confirmed as the best spatial quality method. Furthermore, these results confirm the trade-off between the spectral and spatial quality of pansharpening techniques. As next presented, WATࣹFRAC will be the most suitable technique followed by WAVE_ATROUS, taking into account the balance between a good geometrical representation of structures and a good representation of spectral information (Thomas et al. 2008). Pixel-based classification results Before performing the classification process, the JeffriesMatusita Distance algorithm was carried out in order to check the ROIs separability. This distance provides for mostoftheclasspairsavalueofhigherthan1.8,indicating a good separability for most of the ROIs. Three supervised classification methods were applied to each fused image (Table 1). Table 3 shows the Overall Accuracy as well as the Kappa coefficient for each classification technique applied to each fused image. The best result for both Maximum Likelihood and SupportVectorMachineclassificationisobtainedinthe WATࣹFRACfusedimage,whereastheWAVE_ATROUS fused image obtains the best accuracy with Mahalanobis Distance classification. Every classifier performs a good classification, which is more significant both in Maximum Likelihood and Support Vector Machine, with Kappa coefficients of 0.76 and beyond. Furthermore, the Support Vector Machine achieves the highest accuracy with the WATࣹFRACfusedimage.Finally,itisobservedhowthe WAVE_ATROUS and WATࣹFRAC algorithms, which obtain higher values in the quantitative spatial indices (Table 2), improve the classification compared with the original multispectral imageandtheremainingpansharpening techniques. These results highlight the importance of the spatial information in the Teide National Park where vegetation species are mixed and are limited in size. In more detail, thematic maps for the best pansharpeningalgorithmareshowninFigure 4 for Maximum Likelihood, Mahalanobis Distance, and Support Vector Machine. According to Figure 4c, although the best Maximum Likelihood Classification obtains an Overall Accuracy of 89.12%, some misclassifications appear in this thematic map. For instance, some pixels which are classified as urban (red) appear in bare soil areas, as well as some road pixels in areas of bare soil. Moreover, due to the field observations and information from the Teide managers, an excess of D. bourgaeana appears in the classified image (too many yellow pixels). In addition, S. supranubius must be less abundant around P. canarie nsi s in the ecosystem (top left in the thematic map). For the best Mahalanobis Distance method, similar misclassified pixels to those of the Maximum Likelihood Classification Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018 CANADIAN JOURNAL OF REMOTE SENSING 9 Figure . Multiresolution segmentation and spectral difference segmentation results. Table . Quality results for the complete WV- bands (best results in bold). Spectral Quality Spatial Quality Global Quality SAM Spectral ERGAS Spatial ERGAS FC Zhou Q GS . . . . . . FIHS . . . . . . HCS . . . . . . MTF_GLP 3.212 . . . . 0.961 MTF_GLP_HPM_NM . 0.334 . . . . WAVE_ATROUS . . 0.749 0.935 0.994 . WATࣹFRAC . . . . . . Table . Accuracy assessment of the pixel-based classification algorithms for each fused image (the best results are in bold). Classification Techniques Maximum Likelihood Mahalanobis Distance Support Vector Machine Pansharpening Algorithms Overall Accuracy Kappa Overall Accuracy Kappa Overall Accuracy Kappa Multispectral .% . .% . .% . GS .% . .% . .% . FIHS .% . .% . .% . HCS .% . .% . .% . MTF_GLP .% . .% . .% . MTF_GLP_HPM .% . .% . .% . WAVE_ATROUS .% . 80.79% 0.73 .% . WATࣹFRAC 89.12% 0.85 .% . 92.75% 0.89 Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018 16 E. IBARROLA-ULZURRUN ET AL. González-Lemus, N., Carracedo-Gómez, J.C., and DurbánVillonga, M. 2009. “El Parque Nacional del Teide: Patrimonio mundial de la UNESCO. Anuario de Estudios Atlánticos, Vol. 1(No. 55): pp. 519–568. Gonzalo-Martín, C., and Lillo-Saavedra, M. 2008. “Influence of source images spatial characteristics on the global quality of fused images.” Revista de Teledetección, Vol. 30: pp. 33–46. Goodin, D.G., Anibas, K.L., and Bezymennyi, M. 2015. “Mapping land cover and land use from object-based classification: An example from a complex agricultural landscape.” International Journal of Remote Sensing, Vol. 36(No. 18): pp. 4702–4723. Ibarrola-Ulzurrun, E., Gonzalo-Martin, C., Marcello-Ruiz, J., Garcia-Pedrero, A., and Rodriguez-Esparragon, D. 2017. “Fusion of high resolution multispectral imagery in vulnerable coastal and land ecosystems.” Sensors, Vol. 17(No. 2): p. 228. Isbell, F., Calcagno, V., Hector, A., Connolly, J., Harpole, W.S., Reich, P.B., Scherer-Lorenzen, M., Schmid, B., Tilman, D., and Van Ruijven, J. 2011. “High plant diversity is needed to maintain ecosystem services.” Nature, Vol. 477(No. 7363): pp. 199–202. Khare, S., and Ghosh, S. 2016. “Satellite remote sensing technologies for biodiversity monitoring and its conservation.” International Journal of Advanced Earth Science and Engineering, Vol. 5(No. 1): pp. 375–389. Kpalma, K., El-Mezouar, M.C., and Taleb, N. 2014. “Recent trends in satellite image pan-sharpening techniques.” Paper presented at the 1st International Conference on Electrical, Electronic and Computing Engineering, Vrnja˘ cka Banja, Serbia, June 2014. Kruse, F., Lefkoff, A., Boardman, J., Heidebrecht, K., Shapiro, A., Barloon, P., and Goetz, A. 1993. “The spectral image processing system (SIPS)—Interactive visualization and analysis of imaging spectrometer data.” Remote Sensing of Environment, Vol. 44(No. 2): pp. 145–163. Kyncl,T.,Suda,J.,Wild,J.,Wildová,R.,andHerben,T. 2006. “Population dynamics and clonal growth of Spartocytisus supranubius (Fabaceae), a dominant shrub in the alpine zone of Tenerife, Canary Islands.” Plant Ecology,Vol. 186(No. 1): pp. 97–108. Laben, C.A., and Brower, B.V. 2000. “Process for enhancing the spatial resolution of multispectral imagery using pansharpening.” U.S. Patent No. 6,011,875. Washington, DC: U.S. Patent and Trademark Office. Lantz, N.J., and Wang, J. 2013. “Object-based classification of Worldview-2 imagery for mapping invasive common reed, Phragmites australis.” Canadian Journal of Remote Sensing, Vol. 39(No. 4): pp. 328–340. Li, X., He, M., and Zhang, L. 2013. “Hyperspherical color transform based pansharpening method for WorldView-2 satellite images. Paper presented at the IEEE 8th Conference on Industrial Electronics and Applications, Melbourne, Australia, June 2013. Li, X., Li, L., and He, M. 2012. “A novel pansharpening algorithm for WorldView-2 satellite images.” Paper presented at the International Conference on Industrial and Intelligent Information, Singapore, March 2012. Lillo-Saavedra, M., and Gonzalo, C. 2006. “Spectral or spatial quality for fused satellite imagery: A trade-off solution using the wavelet à trous algorithm.” International Journal of Remote Sensing, Vol. 27(No. 7): pp. 1453–1464. Lillo-Saavedra, M., Gonzalo, C., Arquero, A., and Martinez, E. 2005. “Fusion of multispectral and panchromatic satellite sensor imagery based on tailored filtering in the Fourier domain.” International Journal of Remote Sensing,Vol. 26(No. 6): pp. 1263–1268. Lillo-Saavedra, M., Gonzalo, C., and Lagos, O. 2011. “Toward reduction of artifacts in fused images.” International Journal of Applied Earth Observation and Geoinformation, Vol. 13: pp. 368–375. Lu, D., and Weng, Q. 2007. “A survey of image classification methods and techniques for improving classification performance.” International Journal of Remote Sensing, Vol. 28(No. 5): pp. 823–870. Marangoz, A.M., Oruç, M., Karakı¸s, S., and ¸Sahin, H. 2006. “Comparison of pixel-based and object-oriented classification using IKONOS imagery for automatic building extraction–Safranbolu testfield.” Paper presented at the Fifth International Symposium Turkish-German Joint Geodetic Days, Berlin, Germany, March 2006. Marcello, J., Eugenio, F., Perdomo, U. and Medina, A. 2016. “Assessment of atmospheric algorithms to retrieve vegetation in natural protected areas using multispectral high resolution imagery.” Sensors, Vol. 16(No. 10): p. 1624. Marcello, J., Medina, A., and Eugenio, F. 2013. “Evaluation of spatial and spectral effectiveness of pixel-level fusion techniques.” Geoscience and Remote Sensing Letters,Vol.10(No. 3): pp. 432–436. Martín-Osorio, V.E., Wildpret-De La Torre, W., and Hernández-Bolaños, B. 2005. “Avances significativos enlaelaboracióndelabasededatosgeorreferenciadade Flora y Vegetación del Parque Nacional del Teide, Tenerife,IslasCanariasmedianteunSistemadeInformación Geobotónica.” Vieraea: Folia Scientarum Biologicarum Canariensium, Vol. 33: pp. 345–358. Núnez, J., Otazu, X., Fors, O., Prades, A., Palà, V., and Arbiol, R. 1999. “Multiresolution-based image fusion with additive wavelet decomposition.” IEEE Transactions on Geoscience and Remote Sensing, Vol. 37: pp. 1204–1211. Oruc, M., Marangazoz, M., and Buyuksalih, G. 2004. “Comparison of pixel-based and object-oriented classification approaches using Landsat-7 ETM spectral bands.” Paper presented at the XX ISPRS Congress, Istanbul, Turkey, July 2004. Otto, R., García-Del-Rey, E., Méndez, J., and FernándezPalacios, J.M. 2012. “Effects of thinning on seed rain, regeneration and understory vegetation in a Pinus canariensis plantation (Tenerife, Canary Islands).” Forest Ecology and Management, Vol. 280: pp. 71–81. Padwick, C., Deskevich, M., Pacifici, F., and Smallwood, S. 2010. “WorldView-2 pan-sharpening.” Paper presented at the ASPRS 2010 Annual Conference, San Diego, CA, April 2010. Pakzad, K., Lúcio, G., Mota, A., Meirelles, M.S., Luiz, H., Coutinho, C., and Feitosa, R.Q. 2003. “Automatic interpretation of vegetation areas in Brazil.” Paper presented at the Joint Workshop of ISPRS Working Groups I/2, I/5, IC WG II/IV and the EARSeL Special Interest Group, Hannover, Germany, October 2003. Peña-Barragán, J.M., Ngugi, M.K., Plant, R.E., and Six, J. 2011. “Object-based crop identification using multiple vegetation indices, textural features and crop phenology.” Remote Sensing of Environment, Vol. 115(No. 6): pp. 1301–1316. Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018 CANADIAN JOURNAL OF REMOTE SENSING 17 Pohl, C. 2014. “Challenges of remote sensing image fusion to optimize earth observation data exploitation.” European Scientific Journal, Vol. 9: pp. 355–365. Rodríguez-Esparragón, D. 2015. Evaluación y desarrollo de métricas de calidad espacial y espectral para aplicaciones de fusión de imágenes multiespectrales de teledetección de alta resolución. Thesis, Universidad Las Palmas de Gran Canaria, Gran Canaria, Spain. Rodríguez-Esparragón, D., Marcello-Ruiz, J., Medina-Machín, A., Eugenio-González, F., Gonzalo-Martín, C., and GarcíaPedrero, A. 2014. “Evaluation of the performance of the spatial assessment of pansharpened images.” Paper presented at the IEEE International Geoscience and Remote Sensing Symposium (IGARSS), Quebec City, Canada, July 2014. Ruocco, M., Bertoni, D., Sarti, G., and Ciccarelli, D. 2014. “Mediterranean coastal dune systems: Which abiotic factors havethemostinfluenceonplantcommunities?”Estuarine, Coastal and Shelf Science, Vol. 149: pp. 213–222. Shettigara, V.K. 1992. “A generalized component substitution technique for spatial enhancement of multispectral images using a higher resolution data set.” Photogrammetric Engineering and Remote Sensing, Vol. 58: pp. 561–567. Shridhar, J.D., and Alvarinho, L.J. 2013. “A spectral index ratiobased Antarctic land-cover mapping using hyperspatial 8band WorldView-2 imagery.” Polar Science,Vol.7(No.1): pp. 18–38. Thomas, C., Ranchin, T., Wald, L., and Chanussot, J. 2008. “Synthesis of multispectral images to high spatial resolution: A critical review of fusion methods based on remote sensing physics.” IEEE Transactions on Geoscience and Remote Sensing, Vol. 46(No. 5): pp. 1301–1312. Tu, T.M., Hsu, C.L., Tu, P.Y., and Lee, C.H. 2012. “An adjustable pan-sharpening approach for IKONOS/QuickBird/ GeoEye-1/WorldView-2 imagery.” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, Vol. 5(No. 1): pp. 125–134. Tu, T.M., Su,S.C., Shyu,H.C., andHuang,P.S. 2001. “A new look at IHS-like image fusion methods.” Information Fusion,Vol. 2(No. 3): pp. 177–186. Vivone,G.,Alparone,L.,Chanussot,J.,DallaMura,M.,Garzelli, A., Licciardi, G., Restaino, R., and Wald, L. 2015. “A critical comparison among pansharpening algorithms.” IEEE Transactions on Geoscience and Remote Sensing, Vol. 53(No. 5): pp. 2565–2586. Wald, L. 2000. “Quality of high resolution synthesised images: Is there a simple criterion?” In Proceedings of the Third Conference Fusion of Earth Data: Merging Point Measurements, Raster Maps and Remotely Sensed Images,edited by T. Ranchin and L. Wald, pp. 99–103 Nice, France: SEE/URSICA. Wald, L., Ranchin, T., and Mangolini, M. 1997. “Fusion of satellite images of different spatial resolutions: assessing the quality of resulting images.” Photogrammetric Engineering and Remote Sensing, Vol. 63(No. 6): pp. 691–699. Walsh, S.J., Mccleary, A.L., Mena, C.F., Shao, Y., Tuttle, J.P., González, A., and Atkinson, R. 2008. “QuickBird and Hyperion data analysis of an invasive plant species in the Galapagos Islands of Ecuador: Implications for control and land use management.” Remote Sensing of Environment,Vol. 112(No. 5): pp. 1927–1941. Wang, Z., and Bovik, A.C. 2002. “A universal image quality index.” IEEE Signal Letter,Vol.9(No.3):p.4. Wildpret de la Torre, W. 2001. “Consideraciones sobre la fitobiodiversidad de las cañadas del Teide.” Paper presented at the Curso de acreditación de conocimientos para ejercer de guía en el Parque Nacional del Teide, Tenerife, Spain. Wu, B., Fu, Q., Sun, L., and Wang, X. 2015. “Enhanced hyperspherical color space fusion technique preserving spectral and spatial content.” Journal of Applied Remote Sensing,Vol. 9(No. 1): pp. 097291–097291. Zhou, J., Civco, D., and Silander, J. 1998. “A wavelet transform method to merge Landsat TM and SPOT panchromatic data.” International Journal of Remote Sensing, Vol. 19(No. 4): pp. 743–757. Downloaded by [ULPGC. Biblioteca] at 03:10 15 January 2018