Comparative analysis of UAV-based LiDAR and photogrammetric systems for the detection of terrain anomalies in a historical conflict landscape

Contents lists available at ScienceDirect Science of Remote Sensing journal homepage: www.sciencedirect.com/journal/science-of-remote-sensing Full length article Comparative analysis of UAV-based LiDAR and photogrammetric systems for the detection of terrain anomalies in a historical conflict landscape Marcel Storcha,b,∗, Benjamin Kisliukc, Thomas Jarmera, Björn Waskea, Norbert de Langeb aOsnabrück University, Inst. of Computer Science, Remote Sensing and Digital Image Analysis, Wachsbleiche 27, 49090 Osnabrück, Germany bOsnabrück University, Inst. of Computer Science, Environ. Informatics and Municipal Planning, Wachsbleiche 27, 49090 Osnabrück, Germany cGerman Research Center for Artificial Intelligence, Plan-based Robot Control Group, Hamburger Str. 24, 49084 Osnabrück, Germany ARTICLE INFO Keywords: Historical terrain anomalies UAV-based remote sensing LiDAR vs. photogrammetry Vegetation impact ABSTRACT The documentation of historical artefacts and cultural heritage using high-resolution data obtained from unmanned aerial vehicles (UAVs) is of paramount importance in the preservation of historical knowledge. This study compares three UAV-based systems for the detection of historically relevant terrain anomalies in a conflict landscape. Two laser scanners, a high-end (RIEGL miniVUX-1UAV) and a lower priced model (DJI Zenmuse L1), along with a cost-effective optical camera system (photogrammetry using Structure from Motion, SfM) were employed in two study sites with different densities of vegetation. In the study area with deciduous trees and little low vegetation, the DJI Zenmuse L1 system performs comparably to the RIEGL miniVUX-1UAV, with higher completeness but lower correctness. The SfM method demonstrated inferior performance with respect to correctness and the F1-score, yet achieved comparable or higher completeness values compared to the laser scanners (maximum 1.0, median 0.84). In the study area characterized by dense near-ground vegetation, the detection results are less optimal. However, the RIEGL miniVUX-1UAV system still demonstrates superior results in anomaly detection (F1-score maximum 0.61, median 0.53) compared to the other systems. The DJI Zenmuse L1 data showed lower performance (F1-score maximum 0.56, median 0.46). Both laser scanners exhibited enhanced results in comparison to the SfM approach, with a maximum F1-score of 0.12. Hence, the SfM method is viable under specific conditions, such as defoliated trees without dense low vegetation. Therefore, lower-cost systems can offer cost-effective alternatives to the high-end LiDAR system in suitable environments. However, limitations persist in densely vegetated areas. 1. Introduction The identification of historical traces and remains in landscapes where historical conflicts have occurred is of significant value. It is imperative to document historical remains in order to preserve knowledge before it is lost, which is particularly important given that many such landscapes are subject to significant transformation processes. These can be of a physical nature, for example, through processes caused or accelerated by climate change. Similarly, land development – encompassing construction, land modification and land use changes – poses a significant threat to historical remnants and can, in turn, exacerbate the impacts of climate crisis through e.g. vegetation loss, altered water balance and increased erosion. In addition, transformation processes can be of a discursive nature, which can result in the emergence of certain narratives about the conflicts over time. The documentation of historical traces enables the questioning and, if necessary, refutation of false narratives about the conflicts, engage in a discussion on the ∗Corresponding author at: Osnabrück University, Inst. of Computer Science, Remote Sensing and Digital Image Analysis, Wachsbleiche 27, 49090 Osnabrück, Germany. E-mail address: [email protected] (M. Storch). culture of remembrance, and eventually facilitate the development of cultural heritage. Historical conflicts in First and Second World Wars, in particular, have therefore been the subject of investigation within conflict landscape research (Adam et al.,2022;Stichelbaut and Cowley, 2016;Van Der Schriek,2022). In this context, the term historical terrain anomalies is used to describe structures in the terrain surface that typically exhibit an unnatural and regular form (e.g., rectangular or round shaped) and are most likely the result of a historical conflict. A substantial body of research demonstrates that remote sensing technology can be employed effectively in the context of detecting historical traces and remains and investigating conflicted landscapes. It is a more expedient and therefore more efficient method than ground inspection alone. The majority of analyses have been conducted using airborne (i.e. from an aircraft) and spaceborne systems (Barthelme et al.,2024;Duncan et al.,2023;Luo et al.,2019;Verhoeven,2017; https://doi.org/10.1016/j.srs.2024.100191 Received 24 October 2024; Received in revised form 18 December 2024; Accepted 27 December 2024 Science of Remote Sensing 11 (2025) 100191 Available online 4 January 2025 2666-0172/© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/bync-nd/4.0/ ). M. Storch et al. Weldegebriel et al.,2024). Recently, the utilization of drone-based (UAV, unmanned aerial vehicle) remote sensing data for detecting historical remains has become increasingly important over the last few years (Agudo et al.,2018;Calleja et al.,2018;Campana,2017;Opitz and Herrmann,2018). While the area coverage achievable with a drone is considerably less than that of an aircraft, the significantly higher resolution enables a more comprehensive recording of the ground situation, thereby facilitating a more detailed analysis. Furthermore, drone flights can be flexibly adapted to suit the specific area under investigation and the research question at hand with regard to flight planning and data acquisition parameters (Opitz and Herrmann,2018; Risbøl and Gustavsen,2018;Storch et al.,2021). A variety of sensor technologies are employed in the field of UAVbased remote sensing. However, in this context, a comprehensive and systematic comparison of these sensors still needs to be carried out, which is the focus of this paper. A significant distinction can be made between the utilization of passive optical-reflective camera systems on the one hand and laser scanning (LiDAR, light detection and ranging) systems on the other hand. Laser beams are emitted and their return signals are analyzed, thereby creating a three-dimensional point cloud of the area under investigation (Luo et al.,2019). LiDAR is capable of penetrating vegetation to a certain extent. This represents a significant advantage of LiDAR over other technologies in the detection of historical remains. The deterioration of such remains is particularly prevalent in open, unprotected terrain. The preservation conditions are also poor in agricultural areas, where tillage and other forms of intervention frequently result in significant disturbance. In contrast, historical traces and remnants are usually well preserved under forest cover, where they are better protected from external influences and erosion processes. Consequently, laser scanning represents an appropriate methodology for the scanning and identification of historical traces and remains due to its capability to effectively penetrate vegetation layers (Crow et al., 2007;Ronchi et al.,2020) and UAV-LiDAR has been frequently used for this purpose (Adam et al.,2022;Storch et al.,2021;Khan et al., 2017;Masini et al.,2022;Storch et al.,2023;Zhou et al.,2020). In many instances, including those previously outlined, drone-based laser scanning systems of the highest quality are utilized for these investigations, such as those belonging to the RIEGL (mini)VUX-SYS series. Their system accuracies and precision are typically in the range of millimeters to 1–1.5cm (Zhou et al.,2020;Dreier et al.,2021; Kükenbrink et al.,2022). Nevertheless, the high costs associated with system acquisition can often act as a barrier to the implementation of such systems, irrespective of the specific issue or purpose in question (Zhou et al.,2020;Hu et al.,2020). Consequently, research is being conducted to determine the feasibility of utilizing alternative, cost-effective systems. For instance, Ronchi et al. (2020) examined the use of LiDAR and multispectral data for the detection of historical landscape anomalies. Vilbig et al. (2020) compared airplane-based LiDAR and UAV photogrammetry to map a historic site. Zhou et al. (2020) compared the RIEGL VUX-1UAV and the Velodyne VLP-16 with respect to their capacity to penetrate vegetation and resulting ground point densities. Hu et al. (2020) developed a low-cost UAV-LiDAR system based on the DJI Livox MID40 for use in forest areas. Its capabilities were then compared with those of higher-quality and more expensive scanners, including the RIEGL VUX-1UAV, RIEGL miniVUX-1UAV, and HESAI Pandar40. Salach et al. (2018) and Štroner et al. (2023) each evaluated a UAV-based LiDAR solution (YellowScan Surveyor and DJI Zenmuse L1, respectively) for DTM generation in comparison to an alternative that is even more cost-effective: Photogrammetric point clouds were derived from RGB images acquired by a digital camera that was mounted on a drone. Kükenbrink et al. (2022) also compared two UAV-based LiDAR systems (RIEGL) with terrestrial photogrammetry in forest environments. The studies referenced demonstrate that low-cost variants can serve as a reasonable alternative in certain circumstances. In some instances, they were able to achieve comparable outcomes with regard to accuracy or the target value to be determined. Nevertheless, some are not explicitly concerned with scanning historically relevant remains (Kükenbrink et al.,2022;Hu et al.,2020;Salach et al.,2018;Štroner et al.,2023). Others who do, however, limit themselves to a primarily manual and visual interpretation and evaluation of the data in relation to the historically relevant objects (Ronchi et al.,2020;Zhou et al.,2020; Vilbig et al.,2020). The implementation of automated detection and classification procedures would yet be essential for carrying out an objective and comparable analysis. Hence, one aim of this study is to ascertain the extent to which the classification outcomes using a lower-priced laser scanning system are comparable to those of the higher-priced system, and thus potentially offer an alternative solution. Furthermore, an additional drone is employed to gather RGB image data from the designated survey areas, representing an even more costeffective alternative to laser scanning. A photogrammetric method is used to obtain 3D point clouds from the image data, and the classification results of the detection of terrain anomalies are compared with those of the two laser scanners. The second objective of this paper is therefore to ascertain whether photogrammetrically generated 3D data could represent a viable alternative to the more expensive laser scanner systems for this research domain. Nevertheless, a number of studies have demonstrated that the quality of photogrammetrically generated 3D data can vary significantly in comparison to laser scanning data, particularly in areas with high vegetation density (Vilbig et al.,2020;Salach et al.,2018;Štroner et al.,2023). For this reason, two study sites were selected for analysis, characterized by different densities of vegetation. In light of the aforementioned considerations, the third objective is to assess terrain anomaly detection across varying vegetation densities, using the three UAV-based systems, while evaluating their suitability under realistic, sub-optimal conditions. Therefore, to summarize, the overall objective of this paper is to conduct a comparative analysis of different drone-based systems in the context of detecting historical terrain anomalies. Drone-based laser scanning data and photogrammetrically generated 3D data were collected in a historical conflict landscape with varying vegetation densities, using both a high-cost and a low-cost laser scanning system, as well as an RGB drone. In order to ensure the objectivity of the aforementioned comparisons, it is essential that the evaluation is conducted automatically through the implementation of a classification procedure, as opposed to a manual approach. The use of template-matching approaches or object-based image analysis often presents the disadvantage of requiring the incorporation of prior knowledge or several parameters (Lambers et al.,2019). In contrast, the effectiveness and possible applications of machine learning (ML) in detecting historical traces and remnants have been demonstrated in a few studies (Storch et al.,2023;Guyot et al.,2018). Recently, there has furthermore been a notable increase in the utilization of deep learning in the field of remote sensing (Osco et al.,2021), with applications extending to the detection of historical remains (Duncan et al.,2023;Trier et al., 2021;Trotter et al.,2022). However, its application to drone data can present challenges in this regard, as a substantial and sufficient amount of training data is typically required. This can be problematic in instances where the number of historical remains in small survey areas is limited. Consequently, we have chosen to utilize ML methodologies that are not associated with the domain of deep learning in the present investigation. Finally, the metrics of completeness and correctness, as well as their harmonic mean (F1-score), are utilized to validate the accuracies of the different results. Completeness is a measure of the number of reference instances that have been correctly identified in the classification process. It provides insight into the extent to which the ground truth has been captured. Correctness measures how many of the objects identified in the classification are actually correct. It therefore reflects the accuracy of the output. Science of Remote Sensing 11 (2025) 100191 2 M. Storch et al. Fig. 1. Flow chart depicting the complete workflow for detecting terrain anomalies. In order to avoid the potential limitations of relying on a single classification method, this study employs two distinct ML techniques to assess if they perform in a similar way: A one-class support vector machine (Schölkopf et al.,2001), which can be applied in an unsupervised manner, and the MaxEnt algorithm (Phillips et al.,2006;Phillips and Dudík,2008), which is a supervised method. Both methods are based on a binary approach, in that there is only one class to identify. Consequently, they are referred to as one-class classifiers (OCC) and represent two widely used and well-established OCC in remote sensing, as evidenced by their application in numerous studies (Mack et al., 2014;Mack and Waske,2017;Rapinel and Hubert-Moy,2021;Shi et al.,2021;Yang et al.,2021;Zu et al.,2024). 2. Material and methods 2.1. Overview of methodological framework This section provides a structured overview of the methodological workflow employed in this study. Fig. 1depicts a schematic representation of the complete workflow in the form of a flow chart. (1) Drone data is acquired in a historical conflict landscape (Section 2.2) using three different sensor systems (Section 2.3). Point cloud data is generated from each of the three different data sources (Section 2.4). (2) Ground points are separated from non-ground points (Section 2.5) and a Differential Morphological Profile (DMP) is used on the ground data to extract the terrain features (Section 2.6.1). (3) Subsequently, two distinct classification techniques for anomaly detection are employed in the analysis of each data set (One-Class Support Vector Machine, OCSVM, Section 2.6.2, and the MaxEnt classifier, Section 2.6.3). (4) In order to compare and evaluate the performance of the different acquisition systems for the detection of terrain anomalies, a final validation of the classification results is carried out (Section 2.6.4). 2.2. Study site The area under investigation is situated within the Eifel region in the west of Germany close to the Belgium-German border, specifically within the valley of Kall (Hürtgenwald municipality). In November 1944, the US Army and the Wehrmacht engaged in intense combat in this region. Due to the challenging topography and the frequent retreats, the American soldiers constructed a multitude of defensive structures, commonly referred to as ‘‘foxholes’’, to safeguard themselves from enemy assaults and rifle fire (Miller,2003). The remnants of these circular shelters can still be observed today as topographical anomalies within the landscape of the valley of Kall. More recent processes, including deforestation, the consequences of climate change, and the impact of reenactment, are contributing to the deterioration of these remains, as evidenced by Adam et al. (2022). The study area is situated to the southeast of Aachen in close proximity to the village Kommerscheidt (see Fig. 2). The valley of Kall is predominantly oriented in a north-south direction. It is partially forested and exhibits alternating zones of high and low vegetation. In a manner similar to that described by Storch et al. (2023), two sub-areas were identified within the valley, each with an area of approximately 0.5 hectares. Since these two sub-areas have not been excessively altered by forestry activities or the influence of reenactors in recent years, it can be assumed that fox holes can still be detected. In the following sections, the two study areas will be referred to as study area Aand study area B. Study area A is situated on the eastern side of the valley of Kall. The tree population in this region of the valley remains almost intact, it is thus characterized by a dense assemblage of deciduous trees, while the immediate ground cover is largely absent, with the exception of a sparse accumulation of fallen leaves. The second study area (study area B) is located on the west side of the valley. In contrast to study area A, this region is distinguished by dense low vegetation, including bushes and shrubs, which persist year-round, along with forest cover. For nadir images of the two survey areas, captured via drone and supplemented by ground photographs of potential foxholes, the reader is directed to Fig. 2. 2.3. Data acquisition Data acquisition took place on April 4, 2023. The deciduous trees in this region have not yet reached their full foliage development at this time of year. The sky was predominantly clear, with minimal cloud cover. Three different drone-based systems were used to collect data: Two laser scanners and one optical-reflective camera system. The latter is the cheapest system used and is to be utilized to generate 3D point clouds from the images using digital aerial photogrammetry methodology (Structure-from-Motion, SfM). The high-priced laser scanning system is the RIEGL miniVUX-1UAV scanner, mounted under a DJI Matrice 600 carrier drone. The scanner operates via a rotating mirror, thereby generating scan lines that run transverse to the direction of flight. It is able to capture up to five echoes per emitted pulse at a pulse repetition frequency (PRF) of 100 kHz. The lower-priced laser scanning system is the DJI Zenmuse L1 scanner, mounted under a DJI Matrice 300 carrier drone. It consists of a Livox LiDAR module based on a risley prism setup. The L1 can receive up to three echoes at a scan rate of 240 kHz. Both devices exhibit an elliptical laser footprint; however, this is more than twice as large in the case of the DJI Zenmuse L1 as it is for the RIEGL miniVUX1UAV. The inertial measurement units (IMU) in both systems sample at a rate of 200 Hz. However, the accuracy of the Applanix APX-20 IMU implemented in the RIEGL miniVUX-1UAV is 1.6 times better than that of the DJI Zenmuse L1 in terms of roll and pitch and 4.3 times better in terms of heading (see Table 1,DJI,2024b;RIEGL,2024;Mandlburger et al.,2023). The optical-reflective system is a camera mounted under a DJI Phantom 4. We use the RGB-bands of the P4 multispectral system for our study, as it includes RTK functionality (see below), which is crucial for our work to ensure precise georeferencing. The camera takes images with a resolution of 2.1 megapixels. The wavelengths utilized for visible light imaging are 450 nanometers (nm) ±16 nm for the blue band, 560 nm ±16 nm for the green band and 650 nm ±16 nm for the red band DJI (2024a). All data must be georeferenced as accurately as possible so that the collected data can be compared. Hence, determining the trajectory of the drone as accurately as possible is a pivotal step in the generation and georeferencing of UAV point cloud data. It is therefore common to integrate a correction signal obtained by a fixed base station which is set up at a known and precisely measured location (Differential Global Navigation Satellite System, D-GNSS). In this way, centimeter-level accuracy can be achieved (Dreier et al.,2021). When the correction signal is obtained and integrated in real time, the process is called Real Time Kinematics (RTK). Since both the DJI Science of Remote Sensing 11 (2025) 100191 3 M. Storch et al. Fig. 2. Location and photos of the research area. Photos taken by M. Adam and M. Storch. Table 1 Sensor specifications of the two UAV-based laser scanners (DJI,2024b;RIEGL,2024;Mandlburger et al.,2023). RIEGL miniVUX-1UAV with DJI Zenmuse L1 IMU Applanix APX-20 UAV Laser Wavelength Near-Infrared Near-Infrared Scan Rate 100 kHz 240 kHz Max. Number of Echoes 5 3 Laser Footprint @ 50 m 8 cm ×3 cm 25 cm ×3.5 cm IMU Sampling Rate 200 Hz 200 Hz IMU Accuracy: Roll, Pitch 0.015◦0.025◦ IMU Accuracy: Heading 0.035◦0.15◦ Matrice 300 and the DJI Phantom 4 are RTK-capable, an RTK correction was realized by using the DJI base station (DJI D-RTK 2). The location of the reference point was previously measured precisely using a GNSS receiver (Stonex S9III+), which receives its D-GNSS correction signal from known reference ground stations (satellite positioning service of the German national survey, SAPOS). However, the DJI Matrice 600 is not compatible with the D-RTK 2 station. Therefore, an RTK correction could not be applied during the flight. In this case, the Stonex S9III+ was used to record the raw satellite signals while the DJI Matrice 600 performed the flight. The correction of the raw trajectory was then subsequently applied after the flight during data processing (Section 2.4.1). The flights with the different drone systems were carried out in both areas in succession on the same day. We opted for a flying altitude of 50 meters above ground level (AGL) in order to maintain a safe distance from the trees within the study area and to guarantee continuous visual contact between the drone pilot and the UAV throughout the entire flight duration. The RGB data recording achieves an 80% side and front overlap. 2.4. Point cloud generation 2.4.1. RIEGL miniVUX-1UAV scanner The processing of the LiDAR data recorded by the RIEGL miniVUX1UAV system commences with the correction of the flight trajectory obtained from the drone. As it was not possible to utilize RTK correction with the Matrice 600 in our setup (see Section 2.3), its flight path needs to be rectified prior to the generation of point clouds derived from the scanner data. It is therefore necessary to obtain the GNSS correction data from the GNSS base station. The GNSS observables and the raw IMU (inertial measurement unit) data are imported into the Applanix POSPac software. Subsequently, the corrected trajectory (smoothed best estimate of trajectory, SBET), is combined with the raw scanner data in RIEGL’s RiPROCESS program, and global registration is performed to generate georeferenced point clouds. Furthermore, RIEGL’s RiPRECISION plugin in RiPROCESS is utilized to automatically identify tie-planes, thereby enhancing cross-flight line registration. A more detailed explanation of this workflow can be found in a variety Science of Remote Sensing 11 (2025) 100191 4 M. Storch et al. of other studies (Storch et al.,2021;Dreier et al.,2021;Brede et al., 2017;ten Harkel et al.,2019). 2.4.2. DJI Zenmuse L1 scanner The DJI Zenmuse L1 UAV sensor has been introduced in 2021 as a relatively affordable UAV-based 3D LiDAR sensor. It means to fill the gap between photogrammetric 3D reconstruction and typically more expensive but very precise LiDAR systems. It can be exclusively used as a payload for the DJI Matrice 300 carrier drone and combines an IMU, RGB photo sensor and a LiDAR module based on the principle of the Livox Mid-40. It offers two scan modes which are called line scan mode and non-repetitive (DJI,2024b) enabled by its setup comprising a limited number of LiDAR transceivers and a beam deflection utilizing a Risley prism. This results in the generation of multiple parallel figureof-eight-like scanning trajectories, which are subsequently employed in a narrow, stretched, nearly linear scanning band or rotated into a round petal shape that only repeats after a full iteration of scans. This hybrid approach permits the reduction of costs while maintaining an adequate number of measurement points (Mandlburger et al.,2023; Brazeal et al.,2021). In this study, the non-repetitive scan mode was selected in order to enable the acquisition of a greater number of measurement points at oblique angles along the flight axis than would be feasible in the linear mode. This can be particularly advantageous in vegetated areas. The resulting data can only be processed with the proprietary DJI software Terra into 3D surface models of the environment. In combination with the (RTK corrected) GNSS data of the drone, it is possible to create full color 3D environment models. 2.4.3. Optical-reflective RGB images and structure from motion 3D information is generated from the RGB images using Structure from Motion (SfM), which is a photogrammetric technique that reconstructs 3D structures from a series of 2D images, in this case RGB images taken from the Phantom 4 drone. The images must capture the objects from different angles, in order to ensure sufficient overlap and perspective variations. Therefore, a very high overlap both in and across the flight direction was ensured (80% overlap in the direction of flight, front overlap, as well as 80% overlap transverse to the direction of flight, side overlap, see Section 2.3). Distinctive points (features) such as corners, edges or textures are recognized in the images. Algorithms such as SIFT (Scale-Invariant Feature Transform) extract these features. Subsequently, they are identified in the various images and correlated with each other. The relative position and orientation of the cameras (camera poses) is determined on the basis of the corresponding features. Finally, a dense 3D point cloud is generated using the estimated camera poses and corresponding features. This is done by triangulation, in which the 3D coordinates of the points are calculated by determining the intersection points of the projections of these points in the images (see Jiang et al.,2020 for a comprehensive overview of SfM technology). The procedure is implemented using the Agisoft Metashape software. 2.5. Ground filter When generating digital terrain models from 3D point cloud data, a crucial step is to determine which data points represent the terrain. This is known as (ground) filtering and involves the point-by-point classification of the data points into ground points and non-ground points. For this purpose, there is a wide range of filter algorithms that work in different ways, e.g. surface-based, TIN-based, and morphologybased, or a combination of different approaches. The various algorithms have advantages and disadvantages depending on the conditions and characteristics of the respective study area. Studies have shown that morphologically based methods can lead to good results in complex terrain Storch et al. (2021), Moudr` y et al. (2020), Klápště et al. (2020). We therefore opt for the Simple Morphological Filter (SMRF) for ground filtering. This filtering process is based on morphological operations (erosion and dilation). First, the point cloud is converted into a raster, then a structure mask with increasing size is used to erode local minima, followed by dilation to restore the original shape. By combining these two steps, the points close to the ground are separated from the higher objects. After the morphological operations, a threshold is applied to perform the final classification of the ground and non-ground points (Pingel et al.,2013). The procedure is based on the following parameterization: It is recommended that the maximum radius should be set to 10 mor greater, and that the slope tolerance parameter should not exceed a value of 10%. The elevation threshold is then calculated by multiplying the slope tolerance by the window radius and the cell size (Pingel et al.,2013). In our case, we set the cell size to 20 cm. This leads to a maximum elevation threshold for the initial opened surface to one meter (0.1 × 0.2 × 10 m∕0.2 m). This corresponds to the approximate vertical extent of the terrain anomalies to be expected in the study areas, which consequently remain preserved as part of the minimal surface of the SMRF algorithm. Subsequently, the final threshold to classify each LiDAR point into ground or non-ground is set to 10 cm. This falls within the usual value range for this parameter (typically 0.05 m–0.2 m) in other studies (Moudr` y et al.,2020;Klápště et al.,2020;Štroner et al., 2022). The parameterization is applied to the point clouds of all three drone systems in both study areas in an identical manner, so that all data sets are treated equally for comparability. 2.6. Methodology 2.6.1. Differential morphological profile Since a terrain model is simply a grid of elevation values, spatial information seems particularly relevant for data analysis. Filtering techniques, such as morphological filters, can simultaneously suppress the unimportant details and preserve the spatial characteristics of the other regions, such as certain shapes in the terrain model. In this study, Differential Morphological Profiles (DMP) are used. The DMP was originally developed by Pesaresi and Benediktsson (2001) for segmenting and classifying high-resolution satellite imagery. A morphological profile (MP) is created by repeatedly applying erosions and dilations to an image, more specifically opening (erosion followed by dilation) and closing (dilation followed by erosion) operations. A structuring mask (moving window) defines the local neighborhood around each pixel for these operations. Its shape depends on the type of features being detected. It can be e.g. square, round or rectangular shaped, and also the size (in pixels) of the structuring element must be specified. Then its size is increased in each iteration in order to create the MP. Finally, the DMP extends the MP by calculating the differences between consecutive MP images. In this way, the profiles capture the changes in the image structure as the size of the structuring element varies. When applied to an originally single-channel grayscale image, in this case an elevation grid, the DMP produces a multidimensional result. DMPs are widely used in remote sensing for classifying satellite images (Huang et al.,2016;Kemmouche et al.,2021), but also for object and anomaly detection in LiDAR data (Storch et al.,2023; Mongus et al.,2014). In this paper, a circular shaped structuring mask is employed because the terrain anomalies to be searched for (foxholes) are also round shaped. The initial diameter of the element is set at one meter and increased in five steps up to a size of two meters. This results in a DMP with ten bands and allows terrain anomalies of different sizes to be highlighted. As a result, each pixel has ten values (five times closing, five times opening). Subsequently, the output of the DMP is used as input features for the one-class support vector machine (Section 2.6.2) and the MaxEnt classifier (Section 2.6.3). Science of Remote Sensing 11 (2025) 100191 5 M. Storch et al. 2.6.2. One-Class Support Vector Machine AOne-Class Support Vector Machine (OCSVM) is a special type of support vector machine that is used for detecting anomalies, i.e. identifying anomalous data points that deviate from a single class. It was originally proposed by Schölkopf et al. (2001). As with other SVMs, a kernel function is used with the OCSVM to transform the data into a higher-dimensional space. Commonly used kernel functions are the linear kernel, the polynomial kernel and the radial basis function (RBF) kernel. The OCSVM is then only trained with the data points of one class. These pixels represent the normal, typical state – in this case, terrain without terrain anomalies. The mathematical basis of the OCSVM is to find a decision equation that separates the training data points in high-dimensional space. This equation is based on a parameter 𝜈, which controls what proportion of the data is considered an outlier and how strictly the boundary between normal and anomalous data is defined. After training, OCSVM is used on test data to classify the data points. A new data point (i.e., a pixel not seen during training) is classified as an anomaly if it lies on the negative side of the separating hyperplane in feature space. A number of studies have employed OCSVM for the purpose of anomaly detection in remote sensing data (Ananias and Negri,2021; Chen et al.,2024;Coca and Datcu,2021). A recent study has also demonstrated the efficacy of OCSVM in the detection of terrain anomalies in LiDAR data (Storch et al.,2023). Similarly, we also use the OCSVM in an unsupervised manner. The data sets are divided into patches, 2∕3 of them are used for training, 1∕3 is used for independent classification. Then this process is repeated until all potential combinations of patches have been utilized once to train the OCSVM. The mean signed distance to the separating hyperplane, calculated over all iterations, is employed as the basis for the final classification decision. In the event that the aforementioned distance is negative, the pixel in question is classified as anomalous. In order to ascertain the optimal parameter values for the OCSVM parameter 𝜈, a series of iterations is conducted using values between 0.01 and 0.05 (in 0.5hincrement), which is a parameter range also used in related research (Mack et al.,2014;Li and Xu,2010). The RBF kernel 𝐾(𝑥, 𝑥𝑖) = exp(−𝛾||𝑥−𝑥𝑖||2)is employed as the underlying kernel in this paper, as it is the most commonly applied kernel in numerous other studies (Mack et al.,2014;Mack and Waske,2017;Ananias and Negri,2021). However, as the final classification decision is based on a large number of individual OCSVM runs, the kernel parameter 𝛾is not subjected to additional variation. Instead, it is maintained at its default value, which is equivalent to the reciprocal of the number of features. 2.6.3. MaxEnt classifier The MaxEnt algorithm (Maximum Entropy) was initially developed and employed in the field of ecology for the purpose of modeling the potential geographical distribution of species (Phillips et al.,2006; Phillips and Dudík,2008). It is based on the estimation of probability densities. The objective is to ascertain the probability of occurrence of the target class, given the environmental conditions, i.e. 𝑃(𝑦|𝑥𝑖) which denotes the conditional probability of the presence of the target class 𝑦given the independent input variables 𝑥𝑖. Using Bayes’ rule, this can be transformed into [𝑃(𝑥𝑖|𝑦)𝑃(𝑦)]∕𝑃(𝑥𝑖). Since the a priori probability of the target class, i.e. 𝑃(𝑦)is not known, MaxEnt estimates the density ratio 𝑃(𝑥𝑖|𝑦)∕𝑃(𝑥𝑖). In order to estimate the conditional probability density 𝑃(𝑥𝑖|𝑦), information on the input variables 𝑥𝑖at sample locations where the target class 𝑦was observed are required. In contrast, the unconditional probability 𝑃(𝑥𝑖)is estimated from background samples, without information on whether the target class is present or not. The incorporation of unlabeled data into the modeling process, in conjunction with the positive samples, categorizes MaxEnt as a PU-classifier (Positive and Unlabeled) (Mack et al.,2014;Elith et al.,2011). Similar to the OCSVM, the ten bands of the DMP are employed as input data. Here, they represent the independent input variables 𝑥𝑖. Furthermore, selected locations (pixels) are incorporated into the model at which the target class, namely terrain anomalies indicating foxholes, where observed (see Section 2.6.4 on the acquisition of reference data). To avoid reliance on the selection or distribution of the training data, this classifier is also iterated multiple times. One-third of the reference data is utilized as training samples, while the remaining two-thirds are employed for validation purposes. This process is repeated 100 times, with a randomized new distribution between the training and validation samples each time. Finally, a threshold is required in order to convert the continuous output of MaxEnt into a binary classification. A value of 0.5 is typically selected for this purpose to decide for each location (i.e., for each pixel) if it is classified as an anomaly or not (Mack et al.,2014;Ortiz et al.,2013). 2.6.4. Validation The reference data is generated through a process of manual interpretation and digitization of the data. The positions in the data are independently annotated by three experts in order to identify any terrain anomalies that may indicate the presence of foxholes. Subsequently, the degree of overlap between the reference and classification datasets is ascertained. A match between the reference and the classification result is indicative of a correctly recognized instance (i.e., a terrain anomaly) which is referred to as a true positive (TP). In contrast, an object from the classification dataset that does not have a spatial match in the reference dataset is classified as a false positive (FP). An unrecognized object from the reference data set is classified as a false negative (FN). From this, completeness (or recall, Eq. (1)), correctness (or precision, Eq. (2)) and the F1-score (Eq. (3)) are determined, each of which has a value range from 0 to 1 (see e.g. Zou et al.,2016). Completeness (Recall) =TP TP +FN (1) Correctness (Precision) =TP TP +FP (2) F1-Score = 2⋅ Completeness ⋅Correctness Completeness +Correctness (3) In order to ascertain whether the results, e.g. with regard to the use of the different sensors, differ significantly from each other or whether these differences are too small and therefore random, a statistical test for the significance of the differences on independent samples is carried out. In order to apply a t-test, for instance, it is necessary to ensure that the data in question follows a normal distribution. In the event that this is not the case, a Mann–Whitney U test may be conducted. As previously stated, the selection of training and test data for each run is randomized. Consequently, the classification outcomes can be considered to be the results of independent random samples, and the aforementioned test procedures are therefore appropriate for use. 3. Results 3.1. Point cloud comparison Table 2summarizes the point density resulting from the different data acquisition methods in the two study areas. The lowest point density is achieved by the RIEGL miniVUX-1UAV scanner. It is three to four times higher for the data from the DJI Zenmuse L1 system. The point density obtained through the SfM method is situated between that of the data from the two laser scanners. This also applies to the filtered point clouds, i.e. points that represent the terrain (see the two bottom rows in Table 2). Fig. 3depicts approximately 60 m × 5 mexemplary sections of the filtered point clouds from study areas A and B, derived from data obtained by the different systems. Study area A (left) is typified by a deciduous forest without near-ground vegetation. Even though the SfM method (a3) is unable to derive any meaningful information about the tree vegetation from the RGB images, all three systems apparently Science of Remote Sensing 11 (2025) 100191 6 M. Storch et al. Table 2 Point densities of the point clouds using different acquisition techniques. RIEGL miniVUX-1UAV DJI Zenmuse L1 RGB images and SfM Area A Area B Area A Area B Area A Area B Point Density 245 pts/m2192 pts/m2791 pts/m2779 pts/m2309 pts/m2453 pts/m2 Avg. Horiz. Point Spacing 6cm 7cm 4cm 4cm 6cm 5cm Ground Point Density 105 pts/m278 pts/m2375 pts/m2313 pts/m2201 pts/m2150 pts/m2 Avg. Ground Point Spacing 10 cm 11 cm 5cm 6cm 7cm 8cm Fig. 3. 60 m × 5 mexemplary sections of the filtered point clouds from study areas A and B, derived from the RIEGL miniVUX-1UAV scanner (a1, b1), DJI Zenmuse L1 scanner (a2, b2), RGB images and Structure from Motion (a3, b3). demonstrated a comparable capability to record the earth’s surface topography (ground points are indicated by a brown color). In study area B (right), which is distinguished by markedly more pronounced ground-level vegetation in comparison to study area A, it is evident that there are differences in the performance of the systems. On average, the point density of the ground points of the DJI system is more than three times that of the RIEGL system (see Table 2). Nevertheless, in areas with dense vegetation, particularly bushes and shrubbery (left half of b1–b3 in Fig. 3), an examination of the ground point data from the DJI Zenmuse L1 (b2) reveals the presence of gaps. The coverage of the terrain from the SfM data set (b3) appears to be even more incomplete in these areas. The RIEGL miniVUX-1UAV (b1), however, captures the terrain surface in these regions without producing data voids. 3.2. Terrain anomaly detection Fig. 4depicts an example of the results of the anomaly detection implementation within the two study areas, using the data from the RIEGL miniVUX-1UAV and the OCSVM. The reference data (see Section 2.6.4) is delineated in red, the classification output shows in black color where an anomaly has been identified. Figs. 5and 6depict the quantitative validation results graphically in the form of box plots. Completeness, correctness and the F1-score are compared in terms of the data source (RIEGL miniVUX-1UAV shown in orange, DJI Zenmuse L1 shown in blue, optical images and SfM shown in green) and the anomaly detection algorithm used (OCSVM vs. MaxEnt). Tables 3and 4demonstrate whether the observed differences are statistically significant. As the Shapiro–Wilk test for normal distribution indicated that this was not consistently present across all data sets, the Mann–Whitney U test was carried out uniformly for all data. Finally, Table 5summarizes the median F1-scores. In study area A (Fig. 5), it is noticeable with regard to the completeness that the medians of the completeness of the classifications all lie between 0.68 and 0.84. In general, the results of the MaxEnt method are slightly better than those of the OCSVM. With regard to the data sources, however, there are differences in the results in relation to the two classification methods used. While the difference between the two methods is rather small for the data from the RIEGL system (difference in the median of 0.04), it is 0.1 for the DJI system and rises to 0.16 for the SfM data. For the latter two, the differences are statistically significant. This also applies when comparing MaxEnt from the SfM and DJI data to the MaxEnt classification using the RIEGL data (see Table 3). In contrast, when using OCSVM, the completeness of the classification results does not differ significantly. Similar to the completeness, it can be observed that the MaxEnt method achieves slightly better correctness values than the OCSVM, at least for the data from the two laser scanners used, even if the difference in the data of the DJI system is not significant. The reverse is true for data generated with SfM. Here, OCSVM achieves higher correctness values than the MaxEnt classifier. Overall, the correctness achieved by the MaxEnt method on the RIEGL data is the highest Science of Remote Sensing 11 (2025) 100191 7 M. Storch et al. Table 3 Significance of the differences in the validation metrics (study area A). RIEGL = RIEGL miniVUX-1UAV, DJI = DJI Zenmuse L1, SfM = Structure from Motion based on RGB images. For the entries marked with an X, the two results are statistically significantly different (𝛼= 0.05). Study area A Completeness Correctness F1-Score RIEGL - OCSVM RIEGL - MaxEnt DJI - OCSVM DJI - MaxEnt SfM - OCSVM SfM - MaxEnt RIEGL - OCSVM RIEGL - MaxEnt DJI - OCSVM DJI - MaxEnt SfM - OCSVM SfM - MaxEnt RIEGL - OCSVM RIEGL - MaxEnt DJI - OCSVM DJI - MaxEnt SfM - OCSVM SfM - MaxEnt RIEGL - OCSVM X X X X X X X X X RIEGL - MaxEnt X X X X X X X X X DJI - OCSVM X X X X X DJI - MaxEnt X X X X X SfM - OCSVM X X X Fig. 4. An exemplary presentation of the classification results for terrain anomaly detection in the two study areas. Here: Application of the one-class support vector machine (OCSVM) based on the RIEGL miniVUX-1UAV data sets. Top: Hillshade visualization of the terrain models. Bottom: Classification results. (median is 0.72). The differences to all other variants are statistically significant. Overall, this pattern is reflected in the F1-score. The highest F1scores are achieved by the MaxEnt method on the RIEGL and DJI laser scanner data (median 0.71 and 0.69, difference not significant). When using the OCSVM classifier, this is slightly lower: 0.67 (RIEGL) and 0.62 (DJI). However, this difference is statistically significant. The F1 values of the SfM data are significantly lower for both classification methods in comparison. Their maximum does not exceed 0.61, the medians are 0.56 and 0.51. In study area B (Fig. 6,Table 4), where dense understorey vegetation like bushes and shrubs are present, it can generally be stated that the validation metrics of the methods with the laser scanner data (RIEGL and DJI) are higher than the results with the data generated from SfM. In terms of completeness, it can be seen that the results are higher when using the RIEGL data than when using the DJI data. Taking into account the two different classification methods, the median of the completeness is 0.70 (RIEGL) and 0.60 (DJI) when applying the OCSVM (difference is significant), it is 0.86 (RIEGL) and 0.71 (DJI) when using MaxEnt (not significant). In a direct comparison of the classification methods, MaxEnt therefore achieves higher completeness values than OCSVM (the difference is only significant for the DJI data). At the same time, the different interquartile ranges (IQR) of the two classification methods are apparent. With a value of 0.43 and 0.29, the IQR is 4 and 3 times higher when using MaxEnt than the results obtained with OCSVM. When using the data generated using SfM, the median of completeness is around 0.30 in each case, regardless of the classification method. In general, it should be noted that, due to the object-based validation and the limited number of anomalies in this study area (the reference data set for study area B comprises 10 anomalies), the completeness can only attain values in 0.1 increments here. Consequently, it is possible that the median is equal to either the first or third quartile. This phenomenon is observed in the results derived from the RIEGL miniVUX1-UAV data set (both classifiers) and the DJI Zenmuse L1, utilizing the OCSVM. As with completeness, the data from the two laser scanners also show significantly higher results than the SfM in terms of correctness. A direct comparison of the data from the two laser scanners shows that the correctness values achieved with the RIEGL data are slightly higher than values achieved with the DJI data. However, the difference is not significant when using OCSVM as classifier (medians 0.44 vs. 0.40), it is significant when using MaxEnt (medians 0.18 vs. 0.13). Therefore, in contrast to completeness, the OCSVM correctness values are generally higher than those of MaxEnt. The F1-score therefore behaves in a similar way. The highest F1score is achieved by the OCSVM classifier on the RIEGL laser scanner data (maximum value 0.61, median 0.53), the second best result is obtained when using the OCSVM algorithm on the DJI data (maximum value 0.56, median 0.46). The difference is significant, as with all other variants. Once again, both classification methods achieve by far the lowest results on the data generated with SfM (do not exceed a value of 0.12). Table 5presents a summary of the median F1-scores achieved for the two study areas, three acquisition systems, and two classification methods. 4. Discussion The overall objective of this study was to compare different UAV sensor systems for detecting terrain anomalies. Data sets from two study sites were analyzed, using two different methods (OCSVM, MaxEnt). One aim was to evaluate the performance of a lower-cost laser scanner (in this case DJI Zenmuse L1) in comparison to a high-priced system (in this case RIEGL miniVUX-1UAV). Study area A was characterized by deciduous forest stands without near-ground vegetation. Here, the DJI system achieved a ground point density that is approx. 34 times higher than that of the RIEGL system. Despite the considerable disparity in terrain point density, the classification maps remain largely comparable. The completeness of the DJI data set is slightly higher than that of the RIEGL data set, whereas the opposite is true with regard to correctness. This general trend appears to be independent of the specific classifier, as this behavior can be observed with both the OCSVM and the MaxEnt classifier. In study area B, the vegetation consisted of additional dense low vegetation. Here, the data from the Science of Remote Sensing 11 (2025) 100191 8 M. Storch et al. Fig. 5. Validation quantities in study area A. Fig. 6. Validation quantities in study area B. Science of Remote Sensing 11 (2025) 100191 9