基于多种机器学习方法的SASI航空高光谱数据蚀变矿物填图研究OA

Alteration mineral mapping is an important application of hyperspectral remote sensing in geological exploration.Rapid and accurate completion of alteration mapping for ultra large volume hyperspectral datasets using machine learning models currently presents a major technical challenge.This study addresses the issues of insufficient reliability in supervised classification sample labels and unclear performance differences among various models and uses SASI airborne hyperspectral data from the Liuyuan area to develop a semi-automatic mineral labeling approach based on expert knowledge,and conducts a comparative analysis of multiple machine learning methods.Candidate mineral spectra are first derived through endmember extraction and matching with a standard spectral library,and the initial mapping results are further refined using mixture tuned matched filtering in combination with field verification data,yielding a relatively reliable mineral distribution dataset.Based on this dataset,several classifiers,including support vector machine(SVM),Random Forest,Decision Tree,Gradient Boosting,AdaBoost,Multi-layer Perceptron(MLP),and a Three-dimensional Convolutional Neural-network(3D-CNN),are evaluated using overall accuracy(OA),average accuracy(AA),and the Kappa coefficient.The results show that SVM achieves the best overall performance(OA=96.5％,Kappa=0.956),while AdaBoost performs comparatively poorly.Medium-Al muscovite is best identified by the gradient boosting model,whereas epidote exhibits notable confusion across all classifiers.The 3D-CNN achieves an OA of 95.89％,slightly lower than that of SVM,but shows improved class balance as reflected by higher AA.Overall,the results indicate that classification performance is largely governed by spectral separability,while the incorporation of spatial information contributes to improved class balance but does not consistently enhance overall accuracy.These findings provide useful insights for sample construction and model selection in hyperspectral alteration mineral mapping.