Impact of Training Sample Label Error on Hyperspectral Remote Sensing Image Classification
In the classification of images,especially hyperspectral images,insufficient attention has been paid to the impact of label quality for learning and training on classification performance.Therefore,based on the PyTorch framework and using the Indian Pines hyperspectral dataset,this article explores the impact of features with high spectral similarity on classification results under RF,BP,CNN,and SSConvNeXt models under different proportions of human error labeling.The results show that under the same mislabeling situation,SSConvNeXt and CNN exhibit a classification accuracy advantage of over 20％compared to RF and BP models,and the classification accuracy of SSConvNeXt and CNN models is above 96％when there are no human error labels,10 error noise labels,and 15％and 25％error labels,reflecting the model's fault tolerance and stability.However,in traditional RF and BP models,error labels have a significant and discrete impact on classification.Finally,the mechanism advantages of SSConvNeXt model in classification are mainly analyzed.This study can provide a certain method selection and quantitative analysis basis for the accuracy of remote sensing classification from the perspective of training samples.
hyperspectral remote sensingsample label qualitydeep l earningclassification accuracyclassification mechanism
万方数据
维普
