Leveraging Causality to Enhance Drought Forecasting With Deep Learning Models

ABSTRACT This study aims to improve the interpretability and efficacy of drought forecasting by integrating causality into the Long Short‐Term Memory (LSTM) structure. The proposed causal‐LSTM structure dynamically adjusts input weights based on identified causal relationships, tailored to the diverse climatic regions of Pakistan. As a basis for this analysis, the trend assessment first revealed a substantial degree of heterogeneity in Vegetation Health Index (VHI) trends across the country, motivating a region‐specific modelling approach. To identify the controls behind this variation, the transfer entropy approach together with the Peter and Clark Momentary Conditional Independence (PCMCI) framework was applied, which showed that precipitation (Prc), temperature metrics (average and minimum temperatures, AvgT and MinT), soil moisture (SM) and potential evapotranspiration (PET) were the predominant causal factors influencing VHI. Importantly, these causal drivers varied spatially, with water‐supply, atmospheric‐demand and mixed‐control mechanisms distributed across districts, indicating that drought in Pakistan is governed by locally varying pathways rather than a single dominant control. Based on these district‐specific causal structures, the performance evaluation of the baseline LSTM (using all possible drivers) and the Causal‐LSTM (using dynamically identified causal drivers) models for each district, using Nash‐Sutcliffe Efficiency (NSE) and Root Mean Square Error (RMSE), resulted in improved model performance. The Causal‐LSTM achieved a higher median NSE (approximately 0.82 vs. 0.76 in testing) and a lower median RMSE, indicating better accuracy and efficiency, and this improvement further translated into better drought‐severity classification, with higher accuracy and detection skill across mild, moderate and severe drought categories. Beyond accuracy, the Causal‐LSTM provides physically interpretable predictions by identifying the dominant drought drivers in each district and shows more stable performance across Pakistan's heterogeneous climate. Further efforts are needed to enhance the explainability of the model at local and global scales, ensuring transparency and clear insights into the causal factors driving drought.