Classification of Pesticide Residues in Sorghum Based on Hyperspectral and Gradient Boosting Decision Trees

To address the challenges posed by chemical methods for detecting pesticide residues in sorghum, such as complicated sample preparation and prolonged detection periods, this study presents a rapid and nondestructive detection approach based on hyperspectral imaging (HSI) technology. A group of sorghum without pesticide residues and three groups uniformly sprayed with pesticides were used in this study. Firstly, support vector machine (SVM) classification models were built using spectral data preprocessed with Savitzky–Golay (SG), discrete wavelet transform (DWT), and standard normal variate (SNV) methods, respectively, and SNV was determined to be the best preprocessing method. Secondly, the gradient boosting decision tree (GBDT) algorithm, principal component analysis (PCA), and the successive projections algorithm (SPA) were respectively used to extract feature wavelengths. Pesticide residue identification models based on full and feature wavelengths were then respectively established using backpropagation neural network (BPNN), SVM, and partial least squares discriminant analysis (PLS-DA). The results show that the BPNN model developed using the feature wavelengths obtained from GBDT was the best for identification of pesticide residues, with an accuracy of 97.8% for both the training and testing sets. Finally, visualization of pesticide residue species in sorghum was achieved using the optimal model. This study demonstrates that utilizing HSI in conjunction with the GBDT-BPNN model is an effective, rapid, and nondestructive method for identifying pesticide residues in sorghum.