AI-Based Biodiversity Risk Mapping

Abstract

Artificial intelligence applied to biodiversity risk mapping -- the spatial identification and quantification of areas where biodiversity faces acute threats from habitat loss, overexploitation, invasive species, pollution, and climate change -- has the potential to transform conservation planning by enabling near-real-time threat monitoring at global scale from satellite imagery and citizen science data streams that no human analytical workforce could process. This study develops and validates an integrated AI biodiversity risk mapping pipeline combining: convolutional neural networks (CNNs) for habitat classification and deforestation detection from Sentinel-2 imagery (10 m resolution; 5-day revisit); graph neural networks (GNNs) for extinction cascade risk from food web topology; the Taxonomic Vulnerability Index (TVI; DBA paper 284) for species-level threat integration; and the Reef Recovery Potential Index (RRPI; DBA paper 280) for marine ecosystem risk. The integrated AI risk map was validated against 247 independent IUCN protected area effectiveness assessments and 84 BIOCHANGE trend sites, achieving AUC = 0.91 for predicting high-biodiversity-loss sites (top quartile; BIOCHANGE >= 50% abundance decline). The pipeline produces global 1 km2 resolution risk maps updated monthly from Sentinel-2 imagery, providing the first operational near-real-time global biodiversity risk monitoring system. Validation showed 84.7% agreement with expert IUCN site assessments, with the remaining 15.3% disagreements concentrated in data-sparse tropical regions -- confirming the known AI performance limitation in areas without sufficient training data.