A Multi-Branch Deep Learning Network for Crop Classification Based on GF-2 Remote Sensing

The accurate classification of staple crops is of great significance for scientifically promoting food production. Crop classification methods based on deep learning models or medium/low-resolution images have been applied in plain areas. However, existing methods perform poorly in complex mountainous scenes with rugged terrain, diverse planting structures, and fragmented farmland. This study introduces the Complex Scene Crop Classification U-Net+ (CSCCU+), designed to improve staple crop classification accuracy in intricate landscapes by integrating supplementary spectral information through an additional branch input. CSCCU+ employs a multi-branch architecture comprising three distinct pathways: the primary branch, auxiliary branch, and supplementary branch. The model utilizes a multi-level feature fusion architecture, including layered integration via the Shallow Feature Fusion (SFF) and Deep Feature Fusion (DFF) modules, alongside a balance parameter for adaptive feature importance calibration. This design optimizes feature learning and enhances model performance. Experimental validation using GaoFen-2 (GF-2) imagery in Xifeng County, Guizhou Province, China, involved a dataset of 2000 image patches (256 × 256 pixels) spanning seven categories. The method achieved corn and rice classification accuracies of 89.16% and 88.32%, respectively, with a mean intersection over union (mIoU) of 87.04%, outperforming comparative models (U-Net, DeeplabV3+, and CSCCU). This research paves the way for staple crop classification in complex land surfaces using high-resolution imagery, enabling accurate crop mapping and providing robust data support for smart agricultural applications.