T-SAR-JEPA: Self-Supervised Temporal Anomaly Detection in SAR Amplitude Stacks via Latent Prediction
Researchers have developed T-SAR-JEPA, a novel self-supervised framework designed for detecting temporal anomalies in Synthetic Aperture Radar (SAR) amplitude data. The model utilizes a domain-adapted Vision Transformer (ViT) encoder and a temporal transformer to predict future latent states from sequential SAR acquisitions. Tested on the DFC 2026 dataset, T-SAR-JEPA demonstrated strong performance, achieving a 77.0% ROC-AUC for detecting a volcanic eruption, significantly outperforming several baseline methods. AI
IMPACT Introduces a new self-supervised method for temporal anomaly detection in SAR data, potentially improving monitoring capabilities.