AI-Driven Habitat Change Detection in Alpine Protected Areas: Extended Abstract (Preliminary Results) - AI4Science Conference

AI-Driven Habitat Change Detection in Alpine Protected Areas: Foundation Models vs. Direct Change Detection Harald Kristen1[000−0003−0395−8428], Daniel Kulmer2[0009−0007−1993−1728], and Manuela Hirschmugl1,2[0000−0002−5224−0992] 1University of Graz, 8010 Graz, Austria 2Joanneum Research, 8010 Graz, Austria Abstract. Rapid climate change in alpine ecosystems demands frequent habitat monitoring, yet traditional manual mapping approaches are prohibitively expensive for the temporal resolution needed. This study presents a comprehensive comparison of two fundamental change detection paradigms using long-term alpine habitat data: (1) post classification - independent habitat maps compared temporally versus (2) direct change detection - bi-temporal imagery processed end-to-end. We systematically evaluate modern AI architectures across both paradigms on a unique 20year alpine habitat dataset from Gesäuse National Park, Austria. Postclassification approaches compare geospatial foundation models PrithviEO-2.0 and Clay v1.0 against traditional convolutional neural networks (CNNs) based on U-Net architecture. Direct change detection evaluates specialized transformers ChangeViT against U-Net baselines. Using very high resolution multimodal data including RGB, near-infrared (NIR), LiDAR elevation, and terrain attributes across 4,480 documented habitat changes over 15.3 km², our analysis reveals key performance differences. Within post-classification, foundation models significantly outperform CNNs with Clay achieving 0.60 Overall Accuracy (OA) versus U-Net’s 0.49 OA. Between paradigms, direct methods achieve superior change detection with Intersection Over Union (IoU) scores of 0.48-0.49 compared to post-classification approaches at IoU 0.28. The results provide actionable guidance for protected area managers choosing between AI paradigms for automated habitat monitoring. Keywords: Change Detection ·Remote Sensing ·Protected Areas · Foundation Models ·Alpine Habitats 1 Introduction Alpine protected areas face unprecedented landscape transformation due to climate change, with ecosystems warming at twice the global average rate [3]. Traditional habitat mapping relies on manual interpretation of aerial imagery, exemplified by the HabitAlp project [5], but this approach is too labor-intensive for the frequent monitoring that rapidly changing alpine ecosystems require. 2 H. Kristen et al. Recent advances in geospatial AI, particularly Geospatial Foundation Models (FMs) demonstrate remarkable capabilities when fine-tuned for Earth Observation tasks [7], [4]. Moreover, specialized Vision Transformer architectures have emerged for direct change detection [8]. However, a fundamental question remains: Which automated change detection paradigm, post-classification versus direct change detection, is optimal for ecological monitoring? This gap is significant given recent analyses showing that complex models do not always outperform established baselines, emphasizing the need for rigorous benchmarking [1]. The unique characteristics of ecological data sets may favor different approaches than urban or agricultural applications. Research questions 1. How well can AI procedures detect changes in protected area habitats? 2. Which paradigm is more suitable: direct change detection or post-classification methods? 3. Do foundation models improve performance compared to CNN-based approaches? 4. What is the added value of including additional input data sets such as LiDAR-derived metrics? 2 Methodology 2.1 Study area and Data Gesäuse National Park provides the HabitAlp dataset (2003, 2013) with 23 habitat classes focusing on forest structure. The dataset encompasses 15.3 km²with 4,480 documented changes. Data includes RGB imagery (2003, 2013, 2020, 2024), NIR, and LiDAR-derived products (nDSM, DTM, roughness). 2.2 Approach 1 - Post Classification We finetune U-Nets [6] using pretrained ImageNet weights [2] for semantic segmentation. We also fine-tune two geospatial foundation models: Prithvi-EO-2.0300M (Vision Transformer pretrained on 4.2M NASA HLS V2 samples) [7] and Clay v1.0 (ViT pretrained on 47M satellite images and 24M optical samples) [4]. Semantic segmentation is performed using data from 2013 with RGB + NIR + LiDAR input layers for training/validation, then applied to 2020/2024 data to identify habitat changes through semantic classification comparison. 2.3 Approach 2 - Direct Change We use pretrained U-Nets similar to approach 1, with stacked temporal bands for binary and multiclass change prediction. We also evaluate ChangeViT, which employs vision transformers for binary change detection, excelling at large-scale pattern recognition and detecting gradual and sudden changes [8]. End-to-end training is performed on bi-temporal RGB imagery (2003-2013) with labeled changes, then applied to 2013-2020/2024 data for prospective change detection. AI-Driven Habitat Change Detection in Alpine Protected Areas 3 2.4 Evaluation and Enhancement Evaluation metrics include Jaccard Index / Intersection Over Union (IoU), Overall Accuracy (OA), and F1-scores suitable for imbalanced datasets. Postprocessing filters eliminate change transitions that are ecologically impossible in the given timeframe (e.g., rock →forest, clearcut →mature stage), to improve semantic consistency. Multi-modal analysis evaluates RGB-only versus RGB + NIR versus RGB + NIR + elevation combinations across both paradigms. 3 Preliminary Results 3.1 Performance Comparison Table 1 summarizes performance across both paradigms. Metrics reflect different evaluation strategies due to ongoing data preparation: post-classification shows semantic segmentation accuracy on 2013 habitat data, while direct change detection shows change detection performance using 2003-2013 transitions. Table 1. Performance comparison across paradigms (different evaluation tasks) Paradigm Method IoU OA Semantic Segmentation Task (2013 habitats) Post-Classification U-Net RGB+NIR+nDSM 0.24 0.43 Prithvi-EO-2.0 0.26 0.54 Clay v1.0 0.28 0.60 Change Detection Task (2003-2013 transitions) Direct Change Detection U-Net Binary 0.48 0.63 ChangeViT Binary 0.48 0.64 U-Net Multiclass 0.35 0.51 With Clay achieving the best Post-Classification results, foundation models perform better than traditional CNNs. Multi-modal data integration proves crucial: adding elevation data improves U-Net performance from 0.20 to 0.43 OA. Binary change detection achieves high performance across architectures (IoU: 0.48-0.49), multiclass approaches remain more challenging. ChangeViT matches U-Net performance for binary tasks but offers better scalability. 3.2 Key insights Foundation models significantly outperform traditional CNNs for habitat classification, with Clay achieving 40% better OA than U-Net, while direct change detection methods excel at binary change identification. Elevation data consistently improves performance across all approaches, highlighting the importance of three-dimensional structural information for alpine habitat discrimination. 4 H. Kristen et al. 4 Expected Impact and Future Work This research provides a comprehensive benchmarking framework for AI-based habitat monitoring in protected areas, directly addressing conservation managers’ operational needs. Results demonstrate that paradigm choice depends on specific monitoring objectives: direct methods excel for change detection, while foundation models offer superior semantic understanding. Future work includes generation of 2020 training data and use of 2024 LiDAR data with forest structure attributes. Findings enable protected area managers to make informed decisions about AI adoption, potentially reducing monitoring costs while increasing temporal resolution for climate change adaptation. References 1. Corley, I., Robinson, C., Ortiz, A.: A Change Detection Reality Check (Apr 2024). https://doi.org/10.48550/arXiv.2402.06994, http://arxiv.org/abs/2402.06994, arXiv:2402.06994 [cs] 2. Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: ImageNet: A large-scale hierarchical image database. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. pp. 248–255 (Jun 2009). https://doi.org/10.1109/CVPR.2009.5206848, https://ieeexplore.ieee.org/document/5206848, iSSN: 1063-6919 3. European Environment Agency: Regional climate change and adaptation : the Alps facing the challenge of changing water resources. Publications Office, LU (2009), https://data.europa.eu/doi/10.2800/12552 4. Foundation, C.: Clay v1.0 (2023), https://huggingface.co/made-with-clay/Clay, doi = 10.57967/hf/1537 5. Hauenstein, P., Indra-Camathias, L.: CC-HABITALP Kartierung 1954 - 2003 - 2015. Vorläufiger technischer Schlussbericht, Hauenstein Geoinformatik, Tamins (Oct 2018) 6. Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional Networks for Biomedical Image Segmentation (May 2015). https://doi.org/10.48550/arXiv.1505.04597, http://arxiv.org/abs/1505.04597, arXiv:1505.04597 [cs] 7. Szwarcman, D., Roy, S., Fraccaro, P., Gíslason, P.E., Blumenstiel, B., Ghosal, R., Oliveira, P.H.d., Almeida, J.L.d.S., Sedona, R., Kang, Y., Chakraborty, S., Wang, S., Gomes, C., Kumar, A., Truong, M., Godwin, D., Lee, H., Hsu, C.Y., Asanjan, A.A., Mujeci, B., Shidham, D., Keenan, T., Arevalo, P., Li, W., Alemohammad, H., Olofsson, P., Hain, C., Kennedy, R., Zadrozny, B., Bell, D., Cavallaro, G., Watson, C., Maskey, M., Ramachandran, R., Moreno, J.B.: Prithvi-EO-2.0: A Versatile Multi-Temporal Foundation Model for Earth Observation Applications (Feb 2025). https://doi.org/10.48550/arXiv.2412.02732, http://arxiv.org/abs/2412.02732, arXiv:2412.02732 [cs] 8. Zhu, D., Huang, X., Huang, H., Shao, Z., Cheng, Q.: ChangeViT: Unleashing Plain Vision Transformers for Change Detection (Jun 2024). https://doi.org/10.48550/arXiv.2406.12847, http://arxiv.org/abs/2406.12847, arXiv:2406.12847 [cs]