Método de clasificación de cultivos para imágenes de teledetección multitemporales basado en SAFPN (3+2)D

Accurate crop classification plays a critical role in agricultural monitoring and food security assurance. Effectively exploiting spatiotemporal information from multi-temporal remote sensing data remains a key challenge in crop mapping.This study proposes an improved neural network model, termed the (3+2)D Split-Attention Feature Pyramid Network ((3+2)D SAFPN), which is built upon a hybrid 3D-2D Feature Pyramid Network ((3+2)D FPN). The model integrates a 3D FPN to capture spatiotemporal crop dynamics, a 2D FPN to extract multi-scale spatial features, a split-attention (SA) mechanism to enhance inter-channel information interaction, and a focal loss function to improve learning performance on minority crop classes. Multi-temporal Sentinel-2 imagery acquired in 2024 was used to construct a plot-level NDVI time-series dataset for Talhu Town, Wuyuan County, Bayannur City, Inner Mongolia. The dataset was divided into training, validation, and test sets with a ratio of 6:2:2.Experimental results demonstrate that the proposed (3+2)D SAFPN model achieved overall accuracies of 89.01% and 89.06% on the test and validation sets, respectively, with Kappa coefficients of 0.82 for both sets, outperforming the original (3+2)D FPN model. Furthermore, comparative experiments conducted on the public Munich dataset indicate strong generalization ability, with accuracy improvements of 2.88% on the test set and 2.44% on the validation set compared to the baseline model.The results indicate that the (3+2)D SAFPN model effectively integrates spatial, spectral, and temporal information from multi-temporal remote sensing imagery, providing a robust and high-accuracy solution for crop classification tasks. This approach shows strong potential for large-scale agricultural monitoring applications. The source code of the proposed model is publicly available at: https://gitee.com/btgw/YicongSun/ree/(3+2)D-SAFPN_torch.