Anomaly detection on the Earth’s surface plays a pivotal role in understanding dynamic environmental changes, detecting natural disasters, and supporting timely decision-making. While remote sensing imagery has been widely adopted for such tasks, existing methods often suffer from limited adaptability to heterogeneous landscapes and poor robustness to temporal noise or data gaps. To address these challenges, we propose the Spatio-temporal Adaptive Grid for Surface Anomaly Detection (STAGA), a novel two-stage framework that combines adaptive spatial partitioning, unsupervised temporal modeling, and spatially aware inference to improve both accuracy and operational utility of surface anomaly detection. In the offline stage, STAGA first performs geo-adaptive partitioning, segmenting the Earth’s surface into grid units, and represents each with compact multi-temporal features from spectral bands, remote sensing indices, spatial gradients, and land cover compositions. To model the normal surface evolution patterns, we train lightweight autoencoders per grid unit to learn smooth and denoised temporal trends, forming a Geoscientific Knowledge Grid (GKG). In the online stage, new observations are compared against GKG to compute standardized residuals, which are then refined via a Conditional Random Field (CRF) that fuses spatial context and residual similarity for anomaly inference. We validate STAGA across four representative regions: Los Angeles, Mandalay Poyang Lake, and Gaza, using over five years of Sentinel-2 imagery. The results show that STAGA achieves high detection accuracy (recall up to 93%). Its adaptive partitioning and unsupervised temporal modeling significantly outperform traditional fixed-grid or threshold-based methods. STAGA provides a robust and scalable solution for real-world surface anomaly monitoring.
Shi et al. (Mon,) studied this question.