Abstract:To address the problem that solid waste entrainment in imported iron ore is difficult to identify rapidly and overcome the low efficiency of traditional detection methods, this study conducted identification research using hyperspectral imaging technology combined with machine learning methods. Six types of iron ore and five typical smelting solid wastes were selected as research objects. After sequential standardized treatments including quartering, grinding, sieving through a 200-mesh standard sieve, and drying, various solid waste samples were separately blended with the six iron ore matrices at gradient ratios of 10%, 20%, 30%, 40%, and 50% to prepare a series of experimental samples. Subsequently, visible-near infrared (VIS-NIR) band hyperspectral images were collected. The average spectrum within the region of interest (ROI) was extracted to represent the spectrum of the corresponding sample, and a classification dataset was constructed. After preprocessing via Savitzky-Golay (S-G) smoothing and continuum removal, the spectral data were divided into five folds using the SPXY algorithm for five-fold cross-validation. Four methods, namely genetic algorithm (GA), competitive adaptive reweighted sampling (CARS), uninformative variable elimination (UVE), and successive projections algorithm (SPA), were separately employed to extract characteristic variables from the preprocessed data. Six-class classification models for iron ore-solid waste were established using back propagation neural network (BPNN), radial basis function (RBF) support vector machine (SVM), K-nearest neighbor (KNN), and random forest (RF), respectively. The results demonstrated that the classification model established by RBF-SVM based on the characteristic bands selected by SPA achieved the optimal performance after preprocessing, with an average cross-validation accuracy of 91.67%. Accurate prediction can be realized within the doping ratio range of 20%–50%. Furthermore, the model exhibited excellent applicability to different iron ore matrices, strong generalization ability and satisfactory stability, which can meet the application requirements for rapid identification of solid waste in imported iron ore. The classification model constructed using this modeling route for actual sample detection attained an accuracy of 92.41%, and can generate pixel-level visual maps of solid waste distribution combined with hyperspectral images. This method provides an efficient technical pathway for rapid screening of solid waste in imported iron ore at ports.