Deep multi-block convolutional neural network for quantitative near-infrared spectral analysis

Abstract

Near-infrared (NIR) spectroscopy is a powerful analytical technique, but its effectiveness is often hindered by instrumental noise and a dependency on expert-driven preprocessing and calibration. This study introduces a novel, intelligent deep learning architecture, a deep multi-block convolutional neural network (CNN) that addresses these challenges by performing end-to-end quantitative analysis directly on raw spectral data. By segmenting the NIR spectrum into multiple blocks and processing them in parallel, the model mimics multi-scale feature learning, automatically identifying and leveraging localized spectral regions relevant to target properties. This approach eliminates the need for manual preprocessing, a significant step toward creating autonomous analytical systems. This study evaluated the model on four rice quality datasets (moisture, protein, fatty acids, and total fat) and benchmarked it against traditional partial least squares regression (PLSR) and other deep learning models. The proposed multi-block CNN consistently achieved superior performance, demonstrating its ability to handle raw spectral data robustly. For instance, it achieved coefficient of determination (R2) values of 0.996 ±0.003 for total fat and 0.997 ± 0.001 for fatty acids. Statistical tests confirm that the model’s performance on raw data is not significantly different from its performance with optimized preprocessing, validating its potential for real-world artificial intelligence applications in agriculture, food quality monitoring, and biosystems engineering where rapid, automated analysis is
critical.