Wheat flour, a critical component in a myriad of nutritional applications, is both extensively cultivated and diversely utilized across the globe. However, accurately distinguishing between different types of wheat flour remains a significant challenge. This study presents an innovative identification method marrying near-infrared spectroscopy (NIRS) with ensemble learning. We collected spectral data from four unique types of wheat flour using NIRS, subjecting them to a series of preprocessing steps to optimize our analysis. The Competitive Adaptive Re-weighted Sampling (CARS) algorithm facilitated the extraction of pivotal spectral features, thereby enhancing the generalizability of our ensemble learning model. Diverging from conventional meta-learner designs, we forwent traditional logistic regression in favor of constructing a neural network-based Stacking integrated learning model, the SLNIR-Net. This model stands as a seminal tool for the precise identification of wheat flour varieties. Comparative results highlight SLNIR-Net’s superior performance over established methodologies, including Partial Least Squares Discriminant Analysis (PLS-DA) and four other notable ensemble learning methods, with our model achieving the highest recognition accuracy. Notably, the fine-tuned SLNIR-Net model attained a 100 % accuracy rate, eclipsing the benchmark Stacking model. This leap in performance is a testament to the enhancements made to circumvent overfitting and amplify the model’s generalization capacity. Our study marks a substantial leap forward in the swift and accurate classification of wheat flour types, charting a new course for broader product identification exploits using NIRS.
Read full abstract