Detection of Wheat Protein Content Based on Near-Infrared Spectroscopy with Autoencoder and Stacking

Accurately detecting the protein content in wheat is crucial for enhancing its nutritional value, optimizing food processing quality, and improving agricultural resource management. This study addresses the rapid and accurate detection of wheat protein content by constructing a regression model based on near-infrared (NIR) spectroscopy, combining an autoencoder with stacked ensemble learning. First, Savitzky-Golay filtering was employed to preprocess the raw spectral data, effectively removing noise and improving data quality. Subsequently, an autoencoder was introduced to efficiently reduce the dimensionality of the complex high-dimensional spectral features, preserving essential information while significantly simplifying model complexity and enhancing prediction efficiency and accuracy. Building on this, a stacked ensemble learning strategy was applied to integrate the predictive advantages of multiple base models, accurately capturing the nonlinear relationships within the spectral data. Experimental results demonstrated the superior performance of this model in predicting wheat protein content, achieving an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.9766 and an RMSE of just 0.2529. Compared to traditional algorithms, the proposed model exhibited stronger generalization ability and robustness, providing robust technical support for wheat quality detection.