Incorporating causality into deep learning architectures to improve flash drought forecasts

Abstract

Soil moisture flash droughts present challenges to agriculture and ecosystems, leading to widespread socioeconomic impacts. Predicting and providing early warnings for these events remains difficult. We propose a novel deep learning framework, the ResAttCauRec model, which integrates an attention mechanism and additional causal information into a CNN-LSTM (convolutional neural network with long short-term memory) backbone to capture the dependence of soil moisture on spatial-temporal meteorological variables. Our results demonstrate that the causality module acts as a regularization technique, enhancing model generalization and performance. This enables effective forecasts of flash droughts, achieving an F1 score of 0.41 compared to 0.06 for the baseline model. Model interpretation analysis reveals that the causality degree significantly improves predictive performance for key drivers including daily maximum temperature, evaporation, and surface pressure, alongside soil temperature and moisture. While normal droughts are influenced by long-term temperature trends, flash droughts are more sensitive to rapid atmospheric changes. Our analysis also highlights a concerning trend of increasing drought complexity and intensification, complicating reliable predictions. This study offers valuable insights into flash drought onset mechanisms and advocates for enhanced predictive models that better support agricultural and ecological practices. Additionally, we introduce an effective approach to enhance data-driven models by incorporating additional causal information, which not only facilitates forecast and interpretation of flash droughts but may also be extended to broader extreme weather events.