MTC-TCN-Attn: A TCN-Attention-Based Multi-Task Collaborative Prediction Method for Wheat Soil Moisture

Wheat is one of the most important staple crops worldwide, and soil moisture plays a critical role in regulating crop growth, irrigation management, and agricultural water-use efficiency. Accurate soil moisture prediction is therefore essential for precision agriculture and intelligent irrigation scheduling. However, soil moisture variations are driven by complex interactions among meteorological conditions, soil layers, and temporal dynamics, resulting in strong nonlinear characteristics and long-term dependencies that remain challenging for conventional prediction approaches. To address these challenges, this study proposes a Multi-Task Collaborative Temporal Convolutional Network with Attention (MTC-TCN-Attn) model for multi-variable wheat soil moisture prediction. The proposed framework integrates a Temporal Convolutional Network (TCN) with a multi-head self-attention mechanism, where causal and dilated convolutions are employed to capture multi-scale temporal dependencies, while the attention module adaptively identifies informative time steps and enhances global temporal feature representation. Furthermore, a multi-task collaborative learning strategy based on shared feature representations is developed to exploit the intrinsic correlations among multiple soil-moisture-related variables and improve model generalization. Experiments conducted on wheat field data demonstrate that the proposed MTC-TCN-Attn model outperforms representative benchmark models, including SVR, LSTM, Transformer, and Multi-Task Long Short-Term Memory (MT-LSTM), across all five prediction targets in terms of R2, RMSE, and MAE. For the three principal soil-moisture-state variables, the proposed model achieved R2 values of 0.9147, 0.9651, and 0.9622 for surface soil moisture (SSM), subsurface soil moisture (SUSM), and soil moisture profile (SMP), respectively. Compared with single-task and multi-task TCN baselines, the proposed MTC-TCN-Attn model consistently reduces prediction errors and improves robustness by effectively integrating temporal dependency modeling and cross-task information sharing. The results indicate that the proposed framework provides an effective approach for multi-variable soil moisture forecasting and has promising potential for intelligent irrigation management and precision agricultural applications.