Habitat sampler—A sampling algorithm for habitat type delineation in remote sensing imagery

The management of habitats for the conservation and restoration of biodiversity in protected area networks requires an appropriate monitoring to increase our understanding of processes and dynamics in managed ecosystems. Remote sensing offers a unique potential for the derivation of coherent spatiotemporal information to report on natural or management‐induced habitat change. However, the methods used for the delineation of habitat types in remote sensing imagery depend on the extensive process of ground truth sampling as reference to construct image classifiers. In fact, the number of required reference samples is intrinsically unknown in complex scenes due to the heterogeneity of varying habitat conditions. Thus, most classifiers are not transferable in retrospective image analysis or between different ecosystems that is, however, required for an operational use of remote sensing‐based monitoring systems.


| INTRODUC TI ON
The loss of biodiversity is currently recognized as one of the major global challenges that affects ecosystems worldwide. As a consequence, environmental policies have implemented the conservation and management of species and habitats in protected area (PA) networks (Aichi Target 11 CBD, 2010). A critical step for maintaining and promoting biodiversity through PAs is set by the regular monitoring of the habitat status in terms of habitat extent, species composition and evolving pressures. The goal is to effectively control conservation measures for a target-oriented habitat management Chape et al., 2005;Geldmann et al., 2018;Watson et al., 2014). Management effectiveness evaluations are thus implemented to increase the conservation performance of PAs, particularly with regard to impacts on biodiversity outcomes (Coad et al., 2015;Gray et al., 2016). However, recently management effectiveness evaluations are only implemented for 9% of PAs, reporting on 20% of global PA coverage (UNEP-WCM, IUCN, & NGS, 2018).
A repeatable, standardized and replicable monitoring is still being strongly demanded to rapidly reveal spatiotemporal trends, improve the environmental impact assessment of active habitat management and enforce legal control mechanisms on PA status and configuration (Kati et al., 2015;Lengyel et al., 2008;Watson et al., 2016).
Biodiversity appears in form of spatially and temporally structured vegetation patterns that integrate processes and functions of ecosystems. Systemically relevant patterns can be described by habitat types as fundamental ecological unit. Variations in habitat type extent and composition allow for an explicit evaluation of species exchange between and within habitat boundaries at the scale of landscapes (Escudero et al., 2003;Loreau et al., 2003). Remote sensing technologies offer a unique perspective to measure states and dynamics of such habitat types (Kennedy et al., 2014). In particular, satellite imagery is the most suitable source to acquire coherent spatiotemporal information about the extent and configuration of habitat types within PA networks Regos et al., 2017;Rose et al., 2015). In this regard, image time-series analyses provide insights into mechanisms driving ecosystem change and hence enable the derivation of pressures, exchange pathways and adaptation processes of habitats on the landscape scale (Kennedy et al., 2014;Pasquarella et al., 2016). According to that, habitat characteristics tracked from space have the potential to increase our understanding of the ecology behind biodiversity and related ecosystem functioning (Pettorelli et al., 2018;Requena-Mullor et al., 2018;Vihervaara et al., 2017).
The extraction of spatially explicit information from remote sensing imagery still requires appropriate calibration between ecological field references and image data. Collection of reference samples is the most time-consuming step associated with the highest costs.
For retrospective image analysis, it may in many cases be almost impossible to retrieve detailed reference data. At the same time, the amount and distribution of reference samples directly affects model training and predictive accuracy, particularly in ecological applications of remote sensing where the target classes (e.g. habitat types) often exhibit high intraclass variability in structure and species composition (Cingolani et al., 2004;Pouteau & Collin, 2013;Rocchini et al., 2013;Tuia et al., 2009Tuia et al., , 2011a. This is due to the fact that in remote sensing imagery, vegetation units such as habitat types, are neither spectrally (plant vigour, phenology and life cycle) nor spatially (plant species gradients) unique. In addition, the typology of habitats often mixes criteria from phytosociological classification of plant communities with functional aspects of management targets or landscape units (Evans, 2006;Rodwell et al., 2018). In fact, terrestrial sampling of habitat type properties is prone to subjective bias (Vittoz & Guisan, 2007;Wang et al., 2013) and substantially affects the final spatial configuration of resulting maps (Stohlgren et al., 1997). In consequence, reference data sampling acts as a critical determinant for the accuracy of habitat type delineation in statistical classification approaches, which still impedes the implementation of operational remote sensing monitoring systems Haest et al., 2017).
It was already shown that in cases of appropriate reference data availability, space-borne remote sensing can be used to map the spatial extent and quality of habitats over large areas (Álvarez-Martínez et al., 2018;Cohen & Godard, 2004;Corbane et al., 2015;McDermid et al., 2005). In this regard, recent, freely available multispectral earth observation systems such as Landsat or the European Sentinel-2 satellites provide a suitable data base for implementing regular monitoring Macintyre et al., 2020;Turner et al., 2015). Comprehensive and representative reference sample collection further enables the modelling of the fuzziness of vegetation, particularly the mapping of ecological gradients in complex environments Foody, 1999;Neumann et al., 2016;Rocchini et al., 2013;Schmidtlein et al., 2007), though in most studies habitat types such as in the definition of the European Natura 2000 sites (Förster et al., 2008;Haest et al., 2010Haest et al., , 2017Stenzel et al., 2014) are discriminated using manually composed training sets (supervised classification) (Xie et al., 2008).
The process of image classification can be decoupled from reference data sampling by applying unsupervised clustering approaches.
Nevertheless, the resulting spectral clusters do not necessarily display ecological meaningful units. They need to be manually matched against habitat types on the basis of prior knowledge on the expected distribution of habitats (Hasmadi et al., 2017;Townshend & Justice, 1980) using, for example, post hoc spatial aggregation (Belward et al., 1990;Lark, 1995). In that respect, contextual features from the spatial-spectral domain improve classification accuracy; however, they need to be defined structurally on specific scales that is recently performed using deep learning networks (Cao et al., 2018;Zhang et al., 2017). The latter again introduces the demand of many training samples, whereas pixel-based clustering is limited in case of spatially complex vegetation structures (Jain et al., 1999;Palylyk & Crown, 1984;Townshend & Justice, 1980). The strong dependence on ground reference data and related poor model transferability using standard classification methods in remote sensing image analyses still creates a gap between monitoring demands and conservation measures  although the concept of habitat types has well established as a way to systematically evaluate ecological representation and management effects in PA networks (Müller et al., 2018;Schmidt et al., 2017b). The challenge is to find an adequate spectral sampling and related models for habitat type delineation in complex scenes. In order to effectively support nature conservation planning, semi-automated self-learning and sampling procedures that are capable of representing the complexity of spatiotemporal habitat dynamics with a minimum user interference can potentially advance relevant monitoring tasks.
In this paper, a novel procedure, the Habitat Sampler, is intro-

| ME THODS
The proposed procedure is divided into (a) the simultaneous sam- Each model m max is thereby related to a unique spatial distribution of references, Mn max , that is sampled as independent point locations in an image (sampling by training). In the second step (b), the model ensemble M is tested against a set H of predefined habitat types that are represented by one habitat spectrum Hs per type. A habitat spectrum is hereinafter referred to as the specific composition of spectral wavebands (mono-or multitemporal) that are provided to characterize a habitat type in remote sensing imagery. Hs is used as input for all classification models i × m max to consecutively predict the labels [1;2] that are subsequently compared to the reference labels in H. A final set of models M fin is selected that maximize the predictive power of one habitat type H fin compared to all others. Finally, only the selected models in M fin are applied to the input image and pixel-based predictions of class labels are summed up for a probability mapping of habitat type H fin (selective prediction). By defining a threshold for the spatial probability distribution, the image is reduced by the pixels that represent H fin and the procedure starts from the beginning to find M models for a habitat type from the shortened set Hs = H − H fin (reductive learning). The method is provided as plain language description for the algorithm (see Appendix S1) and for the workflow (see Appendix S2). Convergence is achieved due to two basic assumptions: (a) habitat types are spatially clustered which implies that correctly classified pixels are more likely to be spatially adjacent to similar class pixels, and (b) habitat types can be spectrally resolved in the scale of image pixels size. There are variables that can be set to control the processes of sampling and model building (Table S1.1).

| Model selection by habitat type prediction
The requirement for step two (b.1) is the availability of a list of predefined habitat types H 1…n that is commonly made available by experts for nature conservation purposes. There are two options to provide the related habitat spectra Hs for model test: (1) the expert marks each habitat type by one spatial point location within a scene of the input image or (2) the expert accesses a spectral library. The preferred option is (1) since no further pre-processing is required as spectral predictors are extracted directly from the input image. In option (1), one representative image pixel per habitat type is generated. They are mostly composed of mono-or multitemporal spectral wavebands stacks, for example in satellite imagery. Each model m max from the first step (a) in M is subsequently used to predict the defined habitat types one after another ( Figure S1.3). The habitat type that is predicted is hereinafter defined as target habitat H n . Its labels will be assigned to the value [2].
If the target habitat is predicted as [2] and all the others H ex as [1], the model produces a perfect split and is accepted as classification model in M fin . Thereby, the predictive distance P d evaluates the number of perfect splits over all models in a cumulated ratio starting with a failed prediction of target habitat [2] and the others [2] (H ex /H n = 2/2), which results in P d = 0. A perfect split always increases P d by adding up the labels [1/2] in the next steps (e.g. +[1/2] ->H ex /H n = 3/4; P d = 0.25 and +[1/2] ->H ex /H n = 4/6; P d = 0.44). A failed split instead decreases the level of P d by adding up [2/2] (e.g. +[1/2] ->H ex /H n = 3/4; P d = 0.25 and +[2/2] ->H ex /H n = 5/6; P d = 0.11). By adding up all model predictions, P d approaches asymptotically the value of 1 depending on the number of perfect and failed splits. After testing all models, the target habitat is changed and the models from M are tested again until all habitat types H 1…n are predicted once as target habitat. P d is then used as a criterion to assess the cumulated validity among the final model sets M fin . In consequence, the one target habitat that maximizes P d determines an optimal model set among all habitat types that is used as final model set in M fin . According to this, the models in the final set M fin represent only perfect classifiers for the related target habitat. They are based on spatial point locations used as reference samples for model training in the input imagery.

| Probability mapping and pixel reduction
The models in the final set M fin are finally applied to the input imagery to derive spatially explicit predictions of the selected target habitat. Each pixel is predicted as target habitat [2] or background [1], while all model predictions are summed up to generate a probability map ( Figure S1.4). The more often a pixel is predicted as [2] the higher the probability of the related target habitat. Finally, the user has to decide which threshold to use for the extraction of the respective habitat type pixels. The procedure is repeated until all habitat types are extracted from imagery. As a result, habitat type samples are selected independently from each other, as new point locations are sampled as references from the reduced image pixels.
In each step, the user needs to re-decide which probability distribution of pixel values is appropriate for the representation of the current target habitat. The remaining pixels in the final step cannot be assigned to any habitat category and hence represent undefined surface properties. It is used as an example of abandoned dry heath (Calluna vulgaris) establishment and degradation processes in the continental biogeographical region, particularly for the mapping of resulting differences in Calluna life cycle phases (Gimingham, 1972;Watt, 1955 (Table S1.2). The samples for habitat type discrimination are sampled autonomously according to the spatial sampling procedure in step (a) (Figure 1, Habitat Sampler).

| Satellite imagery
Satellite imagery was taken from the Landsat series that provide multispectral archive imagery from the Thematic Mapper (TM: 1992(TM: , 1997(TM: , 2002(TM: and 2009  that utilize ground control points for geometric alignment (Young et al., 2017) and atmospheric models for surface reflectance derivation (Masek et al., 2006;Vermote et al., 2016). Analysis at finer grain was performed on Copernicus Sentinel-2 A/B satellite imagery provided by the European Space Agency (ESA) in 2018. I used the same spectral regions from Landsat imagery supplemented with the three red-edge channels between 0.703 and 0.779 μm. All bands were resampled to 10-m spatial resolution and geometrically aligned applying the automated co-registration procedure AROSICS (Automated and Robust Open-Source Image Co-Registration Software) (Scheffler et al., 2017). Sentinel-2 at-sensor radiance was finally transferred into top-of-canopy spectral reflectance over radiative transfer modelling in SICOR (Sensor Independent Atmospheric Correction) (Doxani et al., 2018;Hollstein et al., 2016).
All bands were provided as multitemporal image stacks including only completely cloud-free scenes for each year (Table S1.2). Due to the high revisit interval of five days and a selection of only a small area extent, Sentinel-2 A/B provide n = 10 cloud-free dates for the construction of a time stack to extract reference habitat spectra.
The small area within the 2018 Sentinel scene provides a known reference extent that is used once in order to mark valid point locations at which the spectral predictors can be extracted. These habitat spectra (one per type) were used in step (b) of the habitat sampler for model selection in each year . For this purpose, the spectra are resampled to the respective satellite sensor and reduced to the number of available dates in the respective year.
The Sentinel-2 time series is thus used to construct a dense temporal representation of habitat spectra and to perform fine grain analyses for the delineation of management effects in 2018. The performance of autonomous habitat type sampling and prediction within the Habitat Sampler was validated against supervised classification, unsupervised image clustering and spectral unmixing that are commonly used for pattern recognition in remote sensing imagery. As benchmark to represent the case of many training samples I used n = 100 random 75%/25% splits of image spectra extracted from validation samples to train and test random forest (RF) (Breiman, 2001) and support vector machine (SV) (Boser et al., 1992;Vapnik & Lerner, 1963) classification. The accuracy of habitat type discrimination was estimated using the percentage amount of correctly classified habitat types: Overall Accuracy (OA) (Story & Congalton, 1986), and OA corrected for random predictions: Kappa (K) (Cohen, 1960;Kruskal & Goodman, 1954). The averaged benchmark accuracy was tested against Habitat Sampler results that are based on only one reference spectrum Hs per habitat type that were The Habitat Sampler was trained using i x m max random forest models for all associated habitat types H 1…9 . In (a), the hypothetical case of available training data is simulated using a split of digitized validation data, whereas in (b) model training is performed on only the 9 habitat spectra (reduced). In (c), single habitat type spectra Hs were defined as spectral endmembers (EM). Angles between EM and pixel spectra were calculated to assign habitat types on the basis of spectral similarity (spectral angle mapper SAM) (Kruse et al., 1993).

| Validation
In order to prove that habitat sampling does not simply reproduces images statistics as criterion for cluster partitioning (d), a k-means clustering was applied (Jain et al., 1999) with H 1…9 classes that were manually assigned to habitat types.

| Performance evaluation
Averaged classification accuracies for habitat type discrimination are generally highest (OAA > 87%; K > 0.87) for the three validation datasets that use random sample splits from image spectra (Table 1) In retrospective Landsat analysis, supervised classification and linear unmixing using habitat spectra as input crucially decreases classification precision (K « 0.5), while unsupervised image clustering can still deliver moderate accuracies (k-means: 0.5 < K < 0.6). Sentinel-2 supervised classification generally leads to higher class agreement during validation, except in unsupervised k-mean clustering. The validation performance of the habitat sampler is between 10.7% in 2009 and 6.1% in 2018 Sentinel-2 lower then benchmark validation data splits, while the benchmark case uses 75% training samples and the Habitat Sampler uses in average 3.8% of the validation dataset for autonomous sampling (one spectrum per habitat type).

| Spatiotemporal habitat type dynamics
The spatiotemporal evolution of heathland ecosystem patterns shows distinct gradients of heath life cycle development, succession TA B L E 1 Performance metrics overall accuracy (OA) and Kappa (K)_ averaged for n = 100 validation sample splits and spectral profile validation comparing different machine learning classification; HaSa: Habitat Sampler, RF: random forest, SVM: support vector machine, SAM: spectral angle mapper, bench: benchmark training on validation spectra, reduced: training on habitat spectra and degeneration from 1992 to 2018 (Figure 3). All processes that were delineated over associated habitat types (Figure 2)

| Spatial patterns of heathland management
Small-scale management patterns can be made visible by mapping probabilities of extracted habitat types (Figure 4). Pixels are only plotted for probabilities of corresponding habitat types that are above an individually set probability thresholds. The maximum probability for each habitat type is defined by 2 x n models stored in M fin that represent pixels where all models deliver a unique prediction of H fin . Each pixel is finally assigned to one habitat type above an allo- for a better understanding of landscape effects on biodiversity dynamics (Brose & Hillebrand, 2016;Loreau et al., 2003). Accordingly, the study states that ecologists and conservationists should jointly initiate new advances in remote sensing approaches for a more ecosystem-based design of habitat types that can be mapped and associated over specific processes of succession, life cycle traits and management induced disturbance regimes.

| Application and usage
The Habitat Sampler is implemented in R (R Core Team, 2020) and can be executed over a single script or as R-package. It is tested for windows and Unix environments and makes use of Leaflet (Cheng et al., 2019) to generate interactive maps in a web browser. User inputs are required over the R command line interpreter. The proposed procedure makes no assumptions about the input image.
There are no constraints for the spectral-temporal-spatial domain in which an image is sampled. The user is required to have information about expected habitat types and patterns that can be delineated in imagery as the habitat spectrum is marked per point location or extracted from spectral library a priori. Classifiers in M fin can belong to any machine learning method. It takes into F I G U R E 4 Copernicus Sentinel-2 true-colour composite of managed heathland areas and assigned habitat types in 2018; maps of habitat type probabilities derived from Habitat Sampler of Calluna heath series based on minimum threshold (grey); Calluna heath series are spatially represented as different life cycle phases (b.1-b.4) and natural succession (b.4-b.5) that are reset after the implementation of management measures (b5. -b.1) account that the relative performance of any classifier is specific to the unique features of an application (Khatami et al., 2016).
Classifiers convergence is usually fast (<10 steps), whereas divergent behaviour is suppressed by initiating new start configurations. Processing speed crucially depends on the input image size (pixel size, extent, number of layers). Computational efficiency is further determined by the maximum number of samples per step, the number of models saved in M, the sample buffer, the number of iterations and the selected classification algorithm itself that have to be defined individually on the basis of the expected ecosystem complexity (Table S1.1).
The proposed procedure autonomously generates a huge amount of labelled reference samples (see Figure 1 and Figure S2.3 for sample distribution) to maximize the predictive power of cumulated classifier outputs. Unknown or incomplete training samples due to a lack of ground truth data, particularly, in complex scenes of natural habitats are one of the major constraints for accurate image classification (Foody, 2010;Maxwell et al., 2018). Even sufficient amounts of reference data can lead to biased predictions due to random variations in partitioning into training and test samples (Bickel et al., 2009;Lyons et al., 2018). The Habitat Sampler results are comparable to cases where model training is based on comprehensive and representative reference samples, while the final habitat type output is predicted as gradual probability maps.
It is intended that the user individually defines a threshold for a discrete habitat type. Conventional habitat maps often fail to represent the full complexity of organism-environment relationships Zlinszky & Kania, 2016). On the other hand, clear spatial information is required for communicating and controlling effects of habitat management. In that respect, probability maps enable the preservation of information on alternatives to selected classes, particularly, with regard to continuous species gradients and encroachment parameter that are involved in the design of expert-driven habitat quality assessment schemes (Mairota et al., 2015;Nagendra et al., 2013).

| Features
The Habitat Sampler can be classified into the statistical category of active learning. In this context, active learning has evolved as a promising tool for the extraction of independent training samples in remote sensing images (Bruzzone & Persello, 2009;Tuia et al., 2011b;Zhang et al., 2016). Therein, the process of sampling is optimized to fully cover the statistical distribution of a target class, while automatization is realized by iteratively improving the performance of the In remote sensing, habitat spectra are defined as spectral-temporal wavebands that are saved as references datasets from imagery or in spectral libraries (Hueni et al., 2009;Milton et al., 2009). By this means, an external feature space can be generated for model training as substitute for image training. Although spectral end member analysis for hyperspectral airborne (Artigas & Yang, 2005;Dudley et al., 2015;Zomer et al., 2009) and Landsat time series (Hostert et al., 2003;Sonnenschein et al., 2011) data demonstrated the applicability of spectral libraries for vegetation mapping, effects of spatial non-stationary, acquisition scale and phenological shifts still hamper the transferability of externally calibrated models to spatially/ temporally independent images (Feilhauer & Schmidtlein, 2011). In the proposed procedure, the steps of model training and matching a habitat spectrum are decoupled via the criterion of predictive distance P d . Here, the classifiers M are only used to predict Hs of all habitat types H 1…N for a comparative filtering. The predictive resemblance to a given habitat spectrum is used to select an optimal model ensemble M fin that will be applied to imagery (selective prediction).
This way, external habitat spectra do not determine the constraints for image transfer as they are not involved in model training itself.
A target habitat type for which model predictions M fin are mapped will be extracted from the image; thus, the sample population is successively reduced by pixels that are already determined as habitat type (reductive learning). As in each step, the habitat distribution is sampled again uniformly from the remaining pixels, and the procedure creates a balanced training set that overcomes inaccurate model representation due to random clustering of pixel classes in training data splits (Ali et al., 2015;López et al., 2014) or from manually selected field references (Wang et al., 2013).

| Spatiotemporal dynamics of continental dry heathlands
Remote sensing-based vegetation mapping needs carefully designed classification systems in order to adequately represent the complexity of vegetation (Xie et al., 2008). In this study, I propose the introduction of associated habitat type sequences for a process-based mapping of heathland vegetation. In this context, habitat types are connected over successional gradients and life cycle phases that reveal the temporal evolution of life history traits and degradation of abandoned dry heaths in the continental biogeographical region.
The study shows that under rainfall limited conditions Calluna senescence as described in the mature phase, emerges much faster, after 10 years, then described for maritime Heather, where the building phase typically lasts between 7 and 15 years (Gimingham, 1972;Watt, 1955). The fast emergence of Calluna pioneer and building phases towards senescent stands and shrub encroachment indicates that cyclic processes are highly variable according growth forms and rates (Gimingham, 1988;Schellenberg, 2017). Further research is needed for an optimal implementation of habitat management according varying environmental backgrounds that essentially effects the actual restoration success (Fagúndez, 2012;Henning et al., 2017).
Although hierarchical classification of heathland age as well as grass and bush encroachment (Delalieux et al., 2012;Fenske et al., 2020;Siegmann et al., 2014;Thoonen et al., 2013), functional heathland signatures (Schmidt et al., 2017a) and species turnover in heath communities Neumann et al., 2015) have proven to provide valuable information for habitat quality assessment, the concept of associated habitat types contribute towards a more process-based monitoring that can be utilized for an effective restoration management. There are a few studies using multispectral imagery to classify structural patterns of heath stands (Fenske et al., 2020;Förster et al., 2008;Raab et al., 2018;Stenzel et al., 2014;Wood & Foody, 1989). Classification accuracies are overall precise (OA > 77%), which is comparable to the delineation of habitat types using the proposed procedure (OA 78%-84%). In fact, this study presents the first long-term retrospective analysis of heathland dynamics that is based on autonomous sampling of required references. It demonstrates that habitat management has a substantial effect not only on the quantity of habitat types but also on the spatial configuration of arising landscape patterns (see Figure 4). In that respect, spatial patterns of management activities have the potential to reflect differing life history traits and evolving patterns of functional diversity.
For an operational use of the proposed Habitat Sampler, an optimization of computation time and an increased degree of automatization will be continuously developed and provided on GitHub. In this regard, there will be a spectral time series library made available that can be used for heathland vegetation mapping and will be extendable to habitat types in various other ecosystems.

| CON CLUS IONS
In this study, a novel procedure is introduced, the Habitat Sampler, that autonomously generates independent sets of reference samples for the training of habitat type classifiers in remote sensing imagery. It combines the principles of selective sampling and active learning with predictive modelling of habitat type probabilities. For that purpose, I show how the steps of model training and habitat type discrimination can be decoupled by selective prediction of individual habitat spectra. The accuracy outperforms supervised classification, spectral unmixing and image clustering when using only one reference spectrum per habitat type as training sample input.
Thus, reference data dependencies can be substantially minimized.
The procedure is provided as a tool for use in conservation monitoring and management planning for a better representation of habitat dynamics, particularly in retrospective image analyses where in most cases no reference data are available. In a dry heathland area, I present that spatially explicit habitat probabilities can be delineated as associated habitat types that are connected over processes of ecological succession and cyclic life history dynamics. This way, the Habitat Sampler revealed the spatiotemporal evolution of pioneer grasslands, the degradation of heath vegetation as well as recovery dynamics, particularly under the influence of habitat management in open landscapes. There are no restrictions concerning the spatial, temporal nor spectral image domain on which an image is sampled.
The final distribution of generated reference samples is representative for the respective image itself and can be applied to any machine learning classifier. According to that, the Habitat Sampler has the potential to be used for an operational mapping of habitat dynamics and related landscape processes, particularly to implement an effective restoration management in protected areas of various ecosystems.

ACK N OWLED G EM ENTS
The research is part of the F&U-NBS-Verbund project NaTec-KRH.
It is based on the "Federal Program on Biological Diversity," a funding programme for the implementation of the "National Strategy

PE E R R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/ddi.13165.

DATA AVA I L A B I L I T Y S TAT E M E N T
The R scripts, a user manual and a test dataset with coregistrated and atmospherically corrected Sentinel-2 satellite imagery, is made available per GitHub repository on: https://github.com/carst ennh/ Habit atSam pler. The Habitat Sampler is additionally provided as R-source package, including test datasets. It can be installed from https://github.com/carst ennh/Habit atSam pler/R-package.