CWBLS network and its application in portable spectral measurement.

The research presents a novel approach called the D-CWBLS network to address the challenges of poor accuracy and stability in regression models caused by low-signal-to-noise-ratio and low reproducibility data in portable near-infrared spectroscopy. The D-CWBLS network improves upon the BLS network in three key aspects. Firstly, it expands the network structure by incorporating Near-Infrared characteristic spectral band data, thereby emphasizing important information and enhancing accuracy. Secondly, it deepens the network by adding a Dropout layer vertically, optimizing the structure, eliminating redundant information, and improving robustness. Lastly, it combines optimized feature node weight matrices and enhanced node weight matrices to eliminate uncertainty resulting from randomness during network training, subsequently improving robustness. In tests examining model reproducibility, accuracy, and robustness, the D-CWBLS model demonstrated superior performance compared to traditional machine learning models (PLSR, BP-ANN, and ELM), as well as deep learning models (MLP, CNN, and RNN), and even basic BLS and CWBLS models. This highlights the significant progress made by the D-CWBLS model in addressing the challenges associated with using portable near-infrared spectroscopy devices in outdoor settings, exhibiting higher reliability and applicability.