Disentangling sampling bias to generate early-warning surfaces of forest ecological sensitivity: An integrated MaxEnt–CatBoost framework
Forest Ecosystems, cilt.17, sa.100516, ss.1-25, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 17 Sayı: 100516
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.fecs.2026.100516
- Dergi Adı: Forest Ecosystems
- Derginin Tarandığı İndeksler: Natural Science Collection (ProQuest), Earth, Atmospheric, & Aquatic Science Collection (ProQuest), Scopus, Science Citation Index Expanded (SCI-EXPANDED), BIOSIS, CAB Abstracts, Directory of Open Access Journals, Zoological Record
- Sayfa Sayıları: ss.1-25
- İnönü Üniversitesi Adresli: Evet
Özet
Biodiversity models often rely on opportunistic species occurrence data. However, biases due to different sampling efforts or accessibility can confound biodiversity patterns and obscure ecologically pertinent patterns. Such problems can be specifically challenging when aiming at identifying emerging ecological sensitivity in forested and peri-urban landscapes for spatial planning purposes. Here, we present an eco-informatics framework that decouples sampling bias from environmental suitability to derive planning-ready ecological sensitivity surfaces. First, accessibility-related observation patterns are modeled as a null model (NULL) with the maximum entropy (MaxEnt) algorithm. The NULL model is then removed from the full model (FULL) to produce a bias-corrected suitability signal (FULL-NULL). Second, the residual signal is aggregated to account for geometrical effects and ensure consistency in spatial adjacency through a hexagonal grid. Third, categorical boosting (CatBoost) is used to capture nonlinear associations between rarity-based ecological sensitivity patterns and environmental predictors. Fourth, robustness is evaluated via spatial block-based hold-out validation and local spatial autocorrelation analysis, and model results are interpreted using SHapley Additive exPlanations (SHAP). The resulting surfaces highlight transitional landscapes where ecological vulnerability is elevated but not yet fully expressed, supporting a spatial early-warning perspective. Rather than forecasting future ecological collapse, the proposed framework identifies areas exhibiting latent ecological sensitivity and potential vulnerability under current environmental conditions. Overall, the proposed framework provides a practical and transferable approach to convert biased species occurrence data into meaningful spatially-consistent information that can help to inform regional forest management, conservation planning, and decision-making.