Collaborative discrimination of Cadmium and Lead contamination types and levels in Bok Choy based on multi-temporal...

The type and degree of heavy metal contamination in green vegetables are key factors for evaluating food safety. Current evaluation methods based on Hyperspectral Imaging (HSI) are mostly constrained by the single-task learning paradigm. Not only is it difficult to mine the correlated information between tasks, but it also faces the challenges of jointly modeling the high-dimensional redundancy, non-linear features, and dynamic responses of HSI data. To address this deficiency, this study proposes a novel RCBA hybrid deep network, which integrates Residual Convolutional Neural Network (Res-CNN), Bidirectional Long Short-Term Memory (BiLSTM), and Multi-Head Self-Attention (MHSA) mechanisms within the Hierarchical Multi-Task Learning (HMTL) framework. Through a cascaded approach, the RCBA-HMTL architecture effectively degrades high-dimensional spectral redundancy while achieving the synergistic discrimination of contamination types and degrees. The model further introduces Competitive Adaptive Reweighted Sampling (CARS) feature optimization and multi-temporal scale fusion strategies, overcoming the challenges of temporal dynamics. Experimental results demonstrate that on the test set, the model combining CARS feature optimization and the multi-temporal scale strategy achieved classification accuracies (Accuracy) of 0.962, 0.987, 0.985, and 0.981 in identifying the contamination type and the contamination levels of Cadmium (Cd), lead (Pb), and composite contamination, respectively, significantly outperforming traditional machine learning. This provides a highly efficient and reliable new paradigm for the fine-grained characterization and high-precision dynamic monitoring of heavy metal contamination in crops throughout their entire growth cycle.