A Self‐Supervised Seasonal Anomaly Embedding ViT for Label‐Free Drought Mapping in the Horn of Africa

Abstract Drought poses a major threat to agriculture and food security in the Horn of Africa (HOA), where monitoring efforts are hindered by sparse in situ observations and a lack of ground truth data. In this paper, a new self‐supervised drought classification model, Seasonal Anomaly Embedding with Vision Transformers (SAED‐ViT) is proposed using satellite‐derived seasonal anomalies of NDVI, Land Surface Temperature (LST), and precipitation. The method employs the masked autoencoders with Vision Transformers (MAE‐ViT) to learn robust spatiotemporal representations from 25 years of satellite Earth observation data (2000–2024). The learned latent features are clustered using unsupervised K‐Means to identify semantically meaningful drought regimes, which are then mapped to standardized severity classes without requiring predefined thresholds or labeled data. The results exhibit high spatial accuracy and temporal coherence across a broad range of agro‐climatic regions with capturing the large‐scale droughts in 2011, 2017, and 2022. Quantitatively compared and verified against Standardized Precipitation Evapotranspiration Index (SPEI), the agreement is strong ( r = −0.91, P‐value < 0.01) and better when compared to the conventional indexes like NDVI, TCI, and VHI. SAED‐ViT achieved robust and label‐free drought‐severity classification across multiple satellite data sources.