CA-moDWT: A Channel-Attentive Multiscale Framework for Global Soil Moisture Prediction

Soil moisture links precipitation and evapotranspiration with land-surface water and energy exchange, making it central to the terrestrial water cycle, drought monitoring, agricultural water management, and ecosystem assessment. Reliable soil moisture predictions are therefore essential for understanding land-surface moisture dynamics and supporting water-resource management. However, existing deep learning models face two key limitations. First, raw time series contain short-term fluctuations, recurring variations, and long-term trends that are difficult to distinguish. Second, multiscale decomposition produces numerous features that contribute unequally to prediction, and simply combining them may introduce redundant information. To overcome these limitations, we developed a multiscale prediction framework that combines the maximum overlap discrete wavelet transform (moDWT) with channel attention (CA). moDWT separates each dynamic input variable into components representing different temporal scales while preserving the length and alignment of the original sequence. The CA module then adaptively adjusts the relative contributions of the decomposed channels associated with different input variables and time scales before the temporal prediction stage. We evaluated this framework using three deep learning architectures: long short-term memory (LSTM), bidirectional LSTM (BiLSTM), and Encoder–Decoder LSTM. The CA-moDWT strategy consistently improved predictive performance across all three architectures. Relative to the baseline LSTM, CA-moDWT-LSTM achieved the largest overall improvement. It increased R2 from 0.7911 to 0.8563 and KGE from 0.8082 to 0.8850, while reducing ubRMSE from 0.0186 to 0.0163 m3 m−3 and Bias magnitude from 0.0165 to 0.0132 m3 m−3. These changes correspond to increases of 8.24% and 9.50% in R2 and KGE, respectively, and reductions of 12.37% and 20.00% in ubRMSE and Bias magnitude. Improvements were particularly evident in regions characterized by more dynamic soil moisture variability, including parts of North America, South America, Eastern Europe, and East Asia. These findings demonstrate that combining multiscale decomposition with adaptive feature weighting enables deep learning models to represent complex soil moisture dynamics more effectively and provides a promising approach for global next-day soil moisture prediction.