Identification of soybean variety based on spectral data and RGB image fusion combined with deep learning method

The identification of soybean variety is of great importance for crop quality, stabilizing yield, and regulating agricultural markets. This study presents a novel, high performance approach for soybean variety identification by innovatively fusing hyperspectral data with RGB images and introducing an optimized deep learning framework. YOLO (You Only Look Once), a widely-used real-time object detection algorithm, was adapted and improved in this work. Different spectral preprocessing strategies were evaluated to enhance data quality, then a novelty linear interpolation method was proposed to integrate spectral and visual information to enrich feature inputs. YOLOv11 architecture was improved by optimizing parameter efficiency and strengthening the spatial attention mechanism, boosting model representational capacity. Results demonstrate that the combined preprocessing method of moving average and multiple scattering correction achieved optimal performance. The proposed enhanced YOLOv11 model with multi-modal data fusion at trained a test set accuracy of 98.8%, significantly outperforming conventional single modal approaches. These findings validate that multi-modal fusion of spectral and imaging data is a highly effective strategy for crop variety identification, with promising translational potential for broader applications in agricultural product classification and quality control.