Multi-label remote sensing image classification method based on semi-supervised learning

YANG Qiuyong; YANG Chun

doi:10.7688/j.issn.1000-1646.2025.03.12

Abstract：[Objective]Remote sensing images,as an important means of Earth observation,are widely used in various fields such as environmental monitoring,resource exploration,and disaster warning.However,remote sensing images are easily affected by sensor noise,atmospheric interference,and other factors during the acquisition process,which leads to a decrease in image quality and blurry details and thus poses significant challenges to subsequent image analysis and target classification.In the task of multi-label remote sensing image classification,traditional supervised learning methods are inadequate with significant classification errors as multiple categories of targets exist in the image,and there may be complex correlations and dependencies between these targets.[Methods]Therefore,to effectively address the impact of remote sensing image noise,accurately capture image features,and improve classification accuracy,a multi-label remote sensing image classification method based on semi-supervised learning was proposed.The remote sensing images were preprocessed using the perceptual loss function.By searching for pixel positions with missing details and blurs in the images,the signal-to-noise ratio residuals of the original and defective images were calculated,and the degree of degradation in remote sensing image quality was determined.A residual mapping based image denoising algorithm was designed,which adjusted the spectral values of noise positions according to the residual mapping values.By adjusting the relationship between high and low frequencies of pixels,the signal-to-noise ratio was improved,and the detailed information in the image was restored.The semi-supervised learning method was used to update and improve the image classifier,which improved the processing efficiency and classification accuracy of remote sensing images,thus achieving the classification of multi-label remote sensing images.[Results]To verify the effectiveness of the proposed method,image classification experiments were conducted at different resolutions and principal component numbers,and classification experiments were designed for different types of remote sensing images.The test results show that the proposed method performs well in denoising and image detail restoration and can clearly distinguish the color blocks in each region,restoring key detail information in the image.In terms of landform feature extraction,its result has a high degree of consistency with the actual landform distribution with only small errors,which proves its advantages in remote sensing image feature extraction.In terms of image classification accuracy,the proposed method achieves a classification accuracy of 0.88 at an image resolution of 70 ×80 and the principal component number of 12,demonstrating high classification accuracy.Meanwhile,when classifying different types of remote sensing images,the proposed method has a classification accuracy above 0.9 with a maximum of 0.98,which fully verifies its wide applicability and high classification accuracy.[Conclusion]The above results indicate that the proposed method achieves multi-label remote sensing image classification by utilizing an image denoising algorithm that combines the perceptual loss function and residual mapping and a semi-supervised learning method.It not only improves the efficiency and accuracy of remote sensing image classification but also provides new ideas and technical supports for the field of remote sensing image processing,which has higher theoretical significance and practical application value.