An Improved 3D U-Net Hybrid Network for High-Precision Soybean Mapping From Multitemporal Sentinel-2 Imagery

High-precision mapping of soybeans, a globally significant oil and protein crop, has long been constrained by the limitations of traditional remote sensing methods, particularly regarding their inefficiency in fusing spatio-temporal information and suboptimal classification accuracy. Consequently, the rigorous demands of modern precision agriculture for crop monitoring have often not been adequately addressed. To address these challenges, a novel hybrid model was developed in this study for soybean identification using multi-temporal remote sensing imagery. This model is built upon an improved three-dimensional (3D) U-Net architecture, which synergistically integrates Bidirectional Long Short-Term Memory (Bi-LSTM) networks and Random Forest (RF) algorithms. This integration enables the effective extraction of multi-level features and subsequent integrated classification. A notable innovation of this research involves the incorporation of the Greenness-Water Content Composite Index (GWCCI) into the feature set. GWCCI quantifies canopy greenness and water status, improving discrimination of soybean phenology. Experimental validation was conducted within the primary soybean production region of Hailun City, Heilongjiang Province. The results demonstrated that the proposed model significantly outperformed 10 baseline methods (including standard 3D U-Net, CNN, and RF) across four key metrics: Overall Accuracy (OA), User Accuracy (UA), F1-score, and Kappa coefficient. Specifically, when benchmarked against a standard 3D U-Net, the developed model exhibited a 3.2% increase in OA, a 9.3% increase in UA, a 2.6% improvement in F1-score, and a 6.4% increase in the Kappa coefficient. These quantitative improvements highlight its enhanced capability in distinguishing soybeans from background crops. Ultimately, this research not only provides a novel methodological framework for precise crop classification but also contributes a viable multi-dimensional feature fusion strategy directly applicable to crop monitoring and intelligent agricultural management.