Multi-Relational Decker Duck Swarm Graph Convolutional Attention Network for Short-Term Soil Moisture Variations and Crop Yield Prediction

Precise soil moisture forecasting is essential in precision agriculture, allowing for effective management of water resources and estimation of crop yields. Soil moisture changes are driven by complex environmental conditions, necessitating strong predictive models to handle subtle dependencies. In this research, a Multi-Relational Decker Duck Swarm Graph Convolutional Attention Network (MDR-2DSG-CAN) is introduced for short-term soil moisture forecasting and crop yield estimation. The Soil Moisture Monitoring Dataset, which has 5,000 hourly samples from January 1, 2025, is used to evaluate the model. The methods include Cumulative Curve Fitting Approximation (CCFA) for pre-processing, Adaptive Causal Decision Transformers for feature extraction, and MDR-2DSG-CAN for prediction. MDR-2DSG-CAN is a new method that couples a Double Decker Convolutional Neural Network (DDCNN) with a Multi-Relational Graph Attention Network (MR-GAT), and its parameters are optimized via the Duck Swarm Algorithm (DSA). This composite framework improves the ability to simulate spatiotemporal correlations and intricate variable interactions influencing soil moisture processes. Large-scale experiments illustrate that MDR-2DSG-CAN attains 99.9% accuracy, surpassing traditional machine learning and deep learning approaches. The prediction of soil moisture involves understanding the properties of the soil, and an efficient optimization technique based on the Duck Swarm Algorithm improves generalizability and stability.