Synergistic multi-level fusion framework of VNIR and SWIR hyperspectral data for soybean fungal contamination...

Current methods for detecting soybean fungal contamination are often destructive, time-consuming, and labor-intensive. This study proposed an efficient approach by fusing visible and near-infrared (VNIR) and short-wave infrared (SWIR) hyperspectral imaging (HSI) data for non-destructive classification of moldy soybeans. A multi-level framework was developed using low-level and intermediate-level data fusion strategies, combined with feature selection algorithms, namely competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA), and uninformative variable elimination (UVE). The results demonstrated that the low-level fusion (LLF) strategy with one-dimensional convolutional neural network (1D-CNN) achieved accuracy of 97.52%, while intermediate-level (ILF) with Transformer reached accuracy of 95.71%. The proposed framework enhances accuracy and robustness, offering a promising tool for non-destructive agricultural quality monitoring.