Abstract
Soil moisture is a critical parameter in precision irrigation management, yet its spatiotemporal heterogeneity and nonlinear dynamics pose significant challenges to accurate prediction. Existing deep learning methods, including recurrent networks and convolutional architectures, struggle to simultaneously capture long-range dependencies and location-specific temporal patterns across spatially distributed sensors. To address this gap, we propose Swin-MBUNet, a novel framework that integrates the Swin Transformer encoder for hierarchical feature extraction from multivariate time series with a multi-branch UNet decoder for multi-scale temporal feature fusion. The model processes multi-source sensor data including 27 soil moisture channels from a 3×3 spatial grid at three depths, enabling location-specific predictions that account for field-level spatial variability. Trained on multi-source meteorological and sensor data from winter wheat fields, the model achieves superior accuracy, convergence stability, and uncertainty-aware prediction compared with baseline architectures. Our contributions include: (1) formulating soil moisture forecasting as a multi-output temporal prediction problem with location-specific channels from a spatially distributed sensor network (2) designing a hybrid Swin Transformer–UNet framework tailored for precision agriculture (3) demonstrating improved prediction reliability and irrigation decision support through empirical evaluation.