Research On Robust Method For Land Cover Classification Of High-Resolution Remote Sensing Image With Noisy Labels

In recent years,supervised learning methods based on deep learning(DL)have made breakthroughs in the land cover classification of high-resolution remote sensing images(RSI)due to their powerful feature learning capabilities.The success of this method heavily relies on largescale and high-quality labeled data.However,affected by data quality,complexity of ground objects,or professional level of labelers,the samples used to train deep learning models always inevitably contains a lot of noisy labels.How to achieve robust high-score remote sensing land cover classification with noisy labels is a key problem that cannot be avoided in theoretical research and practical application.In response to this problem,this paper has carried out research work from following three points:sensitivity analysis of semantic segmentation model to label noise,automatic detection of noisy labels,and improvement of robustness of semantic segmentation model under noisy labels.The main work and contributions of this paper are as follows:(1)We summarized the main types of label noise in land cover classification and pointed out their corresponding sources.Through the experiments of semantic segmentation under simulated noise conditions with satellite remote sensing images and UAV data,we analyze the effect to the performance of segmentation model of different proportions and types of label noise.(2)A confidence learning-driven intelligent detection method for noise labels in land cover sample datasets is proposed.The method first uses the noisy label dataset to train the semantic segmentation model,and then combines the pseudo labels predicted by the model and the artificial noisy labels to calculate the noise transition matrix,so as to obtain the confidence of whether each sample is correctly labelled as the basis for subsequent label noise detection.Using satellite remote sensing images and UAV datasets to conduct experiments,the experimental results show that the method can achieve stable label noise detection results under different noise ratios.(3)A robust high-score remote sensing land cover classification method with tolerance to label noise is proposed.The method first separates clean samples and noisy samples based on noise detection method,and then noise samples regarded as unlabeled samples are trained together with clean samples under the framework of semi-supervised learning to obtain a land cover semantic segmentation model.The results on satellite remote sensing images and UAV images both show that this method is more robust compared to the method without adding any skills.