A High-Resolution Landscape Dataset for Concept-Based XAI With Application to Species Distribution Models

Augustin de la Brosse; Damien Garreau; Thomas Houet; Thomas Corpetti

cs.CVApr 14, 2026

A High-Resolution Landscape Dataset for Concept-Based XAI With Application to Species Distribution Models

Authors: Augustin de la Brosse, Damien Garreau, Thomas Houet, Thomas Corpetti

Organizations: Université Rennes 2, CNRS, Nantes Université, Univ Brest, LETG, UMR 6554, 35000 Rennes, France · LTSER Zone Atelier Armorique, Rennes, France · University of Würzburg, Center for Artificial Intelligence and Data Science, Germany

Abstract

Mapping the spatial distribution of species is essential for conservation policy and invasive species management. Species distribution models (SDMs) are the primary tools for this task, serving two purposes: achieving robust predictive performance while providing ecological insights into the driving factors of distribution. However, the increasing complexity of deep learning SDMs has made extracting these insights more challenging. To reconcile these objectives, we propose the first implementation of concept-based Explainable AI (XAI) for SDMs. We leverage the Robust TCAV (Testing with Concept Activation Vectors) methodology to quantify the influence of landscape concepts on model predictions. To enable this, we provide a new open-access landscape concept dataset derived from high-resolution multispectral and LiDAR drone imagery. It includes 653 patches across 15 distinct landscape concepts and 1,450 random reference patches, designed to suit a wide range of species. We demonstrate this approach through a case study of two aquatic insects, Plecoptera and Trichoptera, using two Convolutional Neural Networks and one Vision Transformer. Results show that concept-based XAI helps validate SDMs against expert knowledge while uncovering novel associations that generate new ecological hypotheses. Robust TCAV also provides landscape-level information, useful for policy-making and land management. Code and datasets are publicly available.

Figures & tables

Figure 1: Schematic overview of our implementation of Robust TCAV : After training a model to classify the classes Presence and Absence of insect ⓐ, we select concepts ( e.g. , hedgerow) ⓑ, and quantify its importance for the chosen class by computing its Robust TCAV score ⓒ.

Concept	Name	# of patches (% of total number of patches)
Hedgerow	Hedge	50 (7.7%)
Isolated tree	IsoTree	45 (6.9%)
Woodland	Wood	50 (7.7%)
Linear body of water	LinW	50 (7.7%)
Surface body of water	SurfW	20 (3.1%)
Wetland	Wet	50 (7.7%)

Table 1: List of the 15 concepts as well as the number of patches.

Figure 2: Left panel : global overview of one of the acquisition sites. Yellow dots correspond to the locations where Road concept patches were extracted. Right panel : The patch is extracted around the point vector from the multispectral image and the digital elevation models to obtain a 7-band patch.

Taxon	CerberusCNN	ARN-50	Pretrained ARN-50	PicoViT
Plecoptera	0.88 (0.01)	0.82 (0.04)	0.90 (0.07)	0.72 (0.02)
Trichoptera	0.79 (0.03)	0.74 (0.04)	0.75 (0.08)	0.66 (0.06)

Table 2: Model predictive performance per taxon: mean AUC and standard deviation over three training runs. Best scores are in bold.

Figure 3: Robust TCAV: Robust TCAV requires a set of examples for a concept ( e.g. , Road ), and random examples ⓐ, labeled test examples for the studied class (. e.g. , Presence of Plecoptera) ⓑ and a trained network ⓒ. We obtain the CAV ( $v^{\ell}_{C}$ , orange arrow) by computing the difference between the mean activations for the concept examples and for the random examples ⓓ. We use the directional derivative to quantify the sensitivity score $S_{C,k,\ell}$ of an individual training example to concept $C$ for the class $k$ . Robust TCAV score for class $k$ is the proportion of positive sensitivity scores among input examples of this class ⓔ.

Concept	CerberusCNN		ARN-50		PicoViT
	Pleco.	Trich.	Pleco.	Trich.	Pleco.	Trich.
Hedge	0.30 (0.25)	0.44 (0.14)	0.90 (0.00)	0.65 (0.00)	0.85 (0.00)	0.00 (0.00)
IsoTree	0.85 (0.02)	0.74 (0.00)	0.90 (0.00)	0.65 (0.00)	0.85 (0.00)	0.00 (0.00)
Wood	0.85 (0.00)	0.74 (0.00)	0.90 (0.00)	0.65 (0.00)	0.13 (0.06)	0.01 (0.06)
LinW	0.84 (0.08)	0.40 (0.14)	0.79 (0.28)	0.65 (0.00)	0.84 (0.04)	0.00 (0.00)
SurfW	0.15 (0.00)	0.26 (0.00)	0.10 (0.00)	0.97 (0.09)	0.16 (0.03)	1.00 (0.00)

Table 3: Mean TCAV scores (with standard deviation) for the Presence class across all models and taxa.

Rank	Plecoptera Presence	Trichoptera Presence
1	Wood	Wet
2	Wet	Wood
3	PermG	IsoTree
4	IsoTree	PermG
5	Hedge	Hedge
6	LinW	LinW

Table 4: Concept importance ranking for Plecoptera and Trichoptera Presence classes.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Concept	AUC	TCAV Presence	TCAV Absence
Hedge	0.95	1.00 (0.00)	0.13 (0.00)
Road	1.00	1.00 (0.00)	0.00 (0.00)
Wet	0.89	0.88 (0.00)	0.25 (0.00)
TempG	0.65	0.62 (0.00)	0.14 (0.00)
Wheat	1.00	1.00 (0.00)	0.09 (0.00)
Organic	0.92	0.67 (0.00)	0.00 (0.00)

Appendix

Table 5: Sanity Check: Discriminability (AUC) and Self-Influence (Robust TCAV) for selected landscape concepts.

Taxon	Train ( $n=198$ )	Validation ( $n=38$ )	Test ( $n=48$ )
Plecoptera	53%	45%	42%
Trichoptera	43%	45%	48%

Appendix

Table 6: Presence rate for each taxon across the training, validation, and test splits.

Figure 4: CerberusCNN is a convolutional architecture with three parallel branches to capture spatial features at multiple scales. The outputs of the branches are concatenated to feed a classification head.

Figure 5: Distribution maps obtained with the CerberusCNN for the five study sites. The blue vectors represent the water bodies. The color scheme is as follows: dark red is used to denote high probability, and white is used to indicate low probability. Study site names are on the left.

Concept	Plecoptera		Trichoptera
	Presence	Absence	Presence	Absence
Hedge	0.30 (0.25)	0.29 (0.13)	0.44 (0.14)	0.22 (0.04)
IsoTree	0.85 (0.02)	0.21 (0.00)	0.74 (0.00)	0.28 (0.00)
Wood	0.85 (0.00)	0.21 (0.00)	0.74 (0.00)	0.28 (0.00)
LinW	0.84 (0.08)	0.21 (0.01)	0.40 (0.14)	0.20 (0.05)
SurfW	0.15 (0.00)	0.79 (0.00)	0.26 (0.00)	0.72 (0.00)

Appendix

Table 7: TCAV scores (mean with standard deviation in parentheses) for each concept and taxon for CerberusCNN.

Concept	Plecoptera		Trichoptera
	Presence	Absence	Presence	Absence
Hedge	0.90 (0.00)	0.25 (0.00)	0.65 (0.00)	0.28 (0.00)
IsoTree	0.90 (0.00)	0.25 (0.00)	0.65 (0.00)	0.28 (0.00)
Wood	0.90 (0.00)	0.25 (0.00)	0.65 (0.00)	0.28 (0.00)
LinW	0.79 (0.28)	0.32 (0.18)	0.65 (0.00)	0.28 (0.00)
SurfW	0.10 (0.00)	0.75 (0.00)	0.97 (0.09)	0.95 (0.19)

Appendix

Table 8: TCAV scores (mean with standard deviation in parentheses) for each concept and taxon for ARN-50.

Concept	Plecoptera		Trichoptera
	Presence	Absence	Presence	Absence
Hedge	0.85 (0.00)	0.33 (0.03)	0.00 (0.00)	0.00 (0.00)
IsoTree	0.85 (0.00)	0.29 (0.00)	0.00 (0.00)	0.00 (0.00)
Wood	0.13 (0.06)	0.28 (0.05)	0.01 (0.06)	0.02 (0.07)
LinW	0.84 (0.04)	0.28 (0.01)	0.00 (0.00)	0.00 (0.00)
SurfW	0.16 (0.03)	0.71 (0.00)	1.00 (0.00)	1.00 (0.00)

Appendix

Table 9: TCAV scores (mean with standard deviation in parentheses) for each concept and taxon for PicoViT.

Rank	Plecoptera		Trichoptera
	Presence	Absence	Presence	Absence
1	Wood	Cereal	Wet	Wheat
2	Wet	Maize	Wood	Cereal
3	PermG	Convent	IsoTree	Convent
4	IsoTree	SurfW	PermG	Maize
5	Hedge	Organic	Hedge	TempG

Appendix

Table 10: Concept ranking for Plecoptera and Trichoptera ( Presence and Absence classes).

Concept	Plecoptera		Trichoptera
	Presence	Absence	Presence	Absence
Hedge	0.12 (0.18)	0.05 (0.05)	0.31 (0.08)	0.16 (0.04)
IsoTree	0.85 (0.02)	0.21 (0.00)	0.74 (0.00)	0.28 (0.00)
Wood	0.85 (0.00)	0.21 (0.00)	0.74 (0.02)	0.28 (0.00)
LinW	0.81 (0.14)	0.21 (0.03)	0.28 (0.11)	0.13 (0.06)
SurfW	0.15 (0.00)	0.78 (0.02)	0.26 (0.00)	0.72 (0.00)

Appendix

Table 11: TCAV scores (mean with standard deviation in parentheses) using a 0.05 threshold for each concept and taxon for CerberusCNN.

Concept	Plecoptera		Trichoptera
	Presence	Absence	Presence	Absence
Hedge	0.87 (0.17)	0.24 (0.05)	0.62 (0.14)	0.27 (0.06)
IsoTree	0.90 (0.00)	0.25 (0.00)	0.65 (0.00)	0.28 (0.00)
Wood	0.90 (0.00)	0.25 (0.00)	0.65 (0.00)	0.28 (0.00)
LinW	0.03 (0.17)	0.01 (0.05)	0.36 (0.32)	0.15 (0.14)
SurfW	0.02 (0.00)	0.17 (0.00)	0.00 (0.00)	0.00 (0.00)

Appendix

Table 12: TCAV scores (mean with standard deviation in parentheses) using a 0.05 threshold for each concept and taxon for ARN-50.

Concept	Plecoptera		Trichoptera
	Presence	Absence	Presence	Absence
Hedge	0.23 (0.08)	0.02 (0.03)	0.00 (0.00)	0.00 (0.00)
IsoTree	0.20 (0.04)	0.04 (0.01)	0.00 (0.00)	0.00 (0.00)
Wood	0.00 (0.00)	0.00 (0.00)	0.00 (0.00)	0.00 (0.00)
LinW	0.26 (0.10)	0.09 (0.02)	0.00 (0.00)	0.00 (0.00)
SurfW	0.10 (0.01)	0.42 (0.06)	0.00 (0.00)	0.00 (0.00)

Appendix

Table 13: TCAV scores (mean with standard deviation in parentheses) using a 0.05 threshold for each concept and taxon for PicoViT.

Concept-Based Explanations

Concept-based explanations make AI decisions legible through human-understandable features and test how those features affect predictions.

See all 73 papers

SAGE: A sampling-aware global evaluation benchmark for species distribution modeling
Sep 25, 2026Emilia Arens, Nina van Tiel, Robin Zbinden +8
EcoVision Lab, Department of Mathematical Modeling and Machine Learning, University of Zurich, Zurich, Switzerland · ECEO, École Polytechnique Fédérale de Lausanne, Sion, Switzerland · Swiss Federal Institute for Forest, Snow and Landscape Research, WSL Birmensdorf, Switzerland +2
Explainable AI for Biodiversity Monitoring and Ecological Image Analysis
Jun 26, 2026Brinnae Bent, Holly R. Houliston, Jiayi Zhou +2Explainable Artificial Intelligence Biodiversity Monitoring
Pratt School of Engineering, Duke University · British Antarctic Survey · Department of Applied Mathematics and Theoretical Physics, University of Cambridge +1
STELLAR: Spatio-Temporal Environmental Learning with Latent Alignment and Refinement for Long-Tailed Species Distribution Modeling
Jun 7, 2026Shufeng Kong, Tao Yu, Yuanyuan Wei +6Spatiotemporal GNNs
School of Software Engineering, Sun Yat-sen University, Zhuhai, China · School of Computer Science and Artificial Intelligence, Foshan University, Foshan, China · Department of Ecology and Evolutionary Biology, Cornell University, Ithaca, NY, USA +2

Sep 25, 2026J/K move · Enter open · S save

SAGE: A sampling-aware global evaluation benchmark for species distribution modeling

Emilia Arens, Nina van Tiel, Robin Zbinden, Damien Robert+7

EcoVision Lab, Department of Mathematical Modeling and Machine Learning, University of Zurich, Zurich, Switzerland · ECEO, École Polytechnique Fédérale de Lausanne, Sion, Switzerland · Swiss Federal Institute for Forest, Snow and Landscape Research, WSL Birmensdorf, Switzerland+2

Open paper AskPDF arXiv

Knowing where species occur is fundamental for biodiversity research and conservation. Species distribution models (SDMs) link species observations to environmental conditions to estimate their spatial distribution. However, accuracy varies with the underlying data and models, making it essential to know for which species models can be trusted. Deep-learning-based SDMs ("DeepSDMs") now jointly model thousands of species, drawing on hundreds of millions of community-science records. At this scale, averaging performance hides substantial species-level variability, particularly for rare species, often of greatest conservation concern. Records are also strongly biased, making occurrence counts misleading. Accounting for these factors is essential for a reliable and informative evaluation of multi-species SDMs. Here, we introduce a Sampling-Aware Global Evaluation (SAGE) benchmark, combining GBIF records for training with sPlotOpen vegetation plots for presence-absence evaluation across 5771 plant species. We propose an evaluation framework that groups species based on two properties, sampling effort and relative prevalence, which describe how densely a species' range is sampled and how frequently the species is recorded. Evaluating single-species SDMs and multi-species DeepSDMs, we find that Random Forests and DeepSDMs perform best overall, but neither dominates: DeepSDMs outperform single-species SDMs for infrequently recorded species while offering no consistent advantage for well-sampled ones. Crucially, this advantage emerges only when established bias-correction practices, such as spatial thinning and reweighting, are carried over to the deep-learning setting. SAGE helps identify the species and data conditions for which a given approach is beneficial, thereby supporting the development of more transparent and ecologically credible SDMs. Data and code: https://earens.github.io/sage/

Abstract

Concept-Based Explanations

Explore similar work

SAGE: A sampling-aware global evaluation benchmark for species distribution modeling

Explainable AI for Biodiversity Monitoring and Ecological Image Analysis

STELLAR: Spatio-Temporal Environmental Learning with Latent Alignment and Refinement for Long-Tailed Species Distribution Modeling