Construction and optimization of rapid quantitative models for crop straw proximate compositions based on miniature portable NIRS devices and machine learning.

The construction of rapid quantitative models for crop straw proximate compositions based on miniature portable near-infrared spectroscopy (NIRS) devices combined with machine learning algorithms can effectively advance the cost-efficient utilization of crop straw in energy applications. In this study, 250 representative crop straw samples and their proximate composition information were collected, and the spectral data were acquired by a miniature portable NIRS device. Combining with 5 types of preprocessing algorithms-detrending, multivariate scattering correction (MSC), standard normal variate transformation (SNV), Savitzky-Golay convolutional smoothing and its first-order derivatives (SGD1)-and their combinations, the partial least squares (PLS), support vector regression (SVR), and eXtreme gradient boosting (XGB) were employed to construct and optimize over 300 quantitative models for crop straw proximate compositions. The similarities and differences among the different preprocessing and their combinations, and machine learning algorithms were further investigated from the perspective of model interpretability. Results demonstrated that PLS models were suitable for predicting moisture and ash content, while XGB and SVR achieved optimal performance for volatile matter and fixed carbon prediction, respectively. Machine learning hyperparameters were significantly influenced by preprocessing algorithms. Additionally, models built using PLS, SVR, and XGB exhibited structural similarities in variable importance (VI) patterns for characterizing crop straw algorithms. This research provides robust model frameworks and data support for the application of miniature portable NIRS devices and rapid quantitative of crop straw proximate compositions.