Eco‐Phase: A Phase‐Space Trajectory Framework for Unsupervised Drought Detection and LSTM‐Based Forecasting in the Horn of Africa

Abstract Early and precise detection of drought is essential for agricultural resilience in the Horn of Africa (HOA), which is increasingly exposed to climate variability. This paper proposes Eco‐Phase, an unsupervised, high‐resolution agricultural drought monitoring and early warning system based on NDVI–LST phase‐space dynamics. In contrast to conventional threshold‐based or rainfall‐driven indices, Eco‐Phase defines drought as changes in the shape, direction, and velocity of vegetation–temperature trajectories, measured by orbit area, loop direction, and velocity. These measures are summarized into a new Phase Anomaly Score (PAS) to quantify ecosystem stress. With a 23‐year Landsat archive (2000–2022), PAS was able to detect well‐documented droughts and hotspots in regions with high negative correlations with SPEI ( r  = −0.78), SPI ( r  = −0.72), VHI ( r  = −0.81), and TVDI ( r  = −0.75). Moreover, a cluster‐specific LSTM‐based neural network enabled the 1‐year‐ahead prediction of PAS with high accuracy ( r  = 0.689; RMSE = 0.0935). The model revealed increasing frequency and severity of drought in recent decades, specifically in southern Somalia, central Ethiopia, and Eritrea. Eco‐Phase enhances drought monitoring by capturing ecological dynamics, improving spatial resolution (30 m), and enabling early warning using unlabeled data. These capabilities improve proactive drought risk management and climate‐resilient policy planning in data‐constrained regions.