Authors - Barsa Priyadarshani Behera, Monalisa Jena, Ranjan Kumar Behera, Sung-Bae Cho Abstract - Drought prediction remains challenging due to complex physical interactions and limited observability of land-atmosphere processes. This study proposes a Causal-Chain Transformer that explicitly employs drought evolution through three sequential latent representations corresponding to heat stress, evaporation stress, and soil moisture stress. Using only past temperature and evaporation data over a xed historical window, the model predicts future drought occurrence at a predened lead time, while excluding current soil moisture to avoid target leak- age. Experiments on region-averaged NASA POWER and ERA5-Land datasets over Odisha, a state of India, show that the proposed model achieves the highest F1-scores (0.709 on NASA POWER and 0.467 on ERA5-Land), outperforming logistic regression, Long Short-Term Memory (LSTM), and standard Transformer baselines. The learned latent stress signals provide intrinsic interpretability, with early increases in heat and evaporation stress frequently preceding observed drought events, supporting its applicability for early-warning systems in agriculture- dependent regions.
