Integrating causal inference with ConvLSTM networks for spatiotemporal forecasting of root zone soil moisture

Accurate spatiotemporal prediction of the root zone soil moisture (RZSM) is essential for agricultural water management and drought early warning systems. However, current prediction methods often lack prior knowledge, which restricts their predictive accuracy. Causality is an essential form of prior knowledge, describing the relationship where one factor directly influences another, and helping to understand how specific changes lead to particular outcomes within a system. We propose a novel model named Causal-Convolutional Long Short-Term Memory (Causal-ConvLSTM) to predict RZSM in the Yangtze River Basin (YRB) with a lead time of 1-14 days. The Causal-ConvLSTM integrates causal inference into the ConvLSTM framework by employing a causal weight unit to directly incorporate causal relationships from spatiotemporal sequences. The experimental results demonstrate that the spatiotemporal prediction accuracy of the Causal-ConvLSTM model outperforms that of the three baseline models (ConvLSTM, LSTM, and PredRNN-V2) with a lead time of 1-14 days in the YRB. Specifically, the improvements of R2 for the proposed Causal-ConvLSTM over the ConvLSTM are 10.17 %, 18.40 %, and 25.65 % at 1, 7, and 14 lead times, respectively. The ablation experiments also show an improvement in the performance of the Causal-ConvLSTM model relative to the simple feature selection model. These findings establish that incorporating causality into spatiotemporal prediction frameworks significantly enhances predictive capabilities, and future research should prioritize the improvement of model performance and interpretability through the causality.