Research On Classification Of High Resolution Remote Sensing Images Method Based On Deep Learning

The classification of remote sensing images stands as one of the most important issues in the analysis of remote sensing images.The method of machine learning traditionally adopted by remote sensing images requires a lot of burdensome work in data feature analysis and data extraction,which is not only highly demanding in data representation,but also apparently confined to special or limited applications.However,the deeplearning technology developed recently is capable of making breakthroughs in this to an extreme extent.Therefore,this paper intends to explore the classification of remote sensing images based on deep leaning.Focusing on FCN-modify,this paper carries out contrastive experiments between FCN-modify + Focal loss function referred to therein and FCN-8s,Seg Net and U-net,and analyzes the results of these experiments.The experimental results show that the proposed scheme achieves better classification results than similar studies.With regard to the focus above,this paper is to:1.Add the structures accessible to different receptive fields to the traditional FCN model to get features of objects in different sizes for the purpose of FCN-modify;replace traditional cross-entropy cost function with Focal loss function,which is more inclined to classify hard examples;do contrastive analyses of the FCN-modify between the two different loss functions above.2.To solve the problem of limited sample data resources,this paper proposes an active learning method to optimize FCN-modify,and compares it with FCN-modify.3.Contrast and analyze the method referred to in the paper with traditional FCN,U-net and Seg Net models in classification experimental results,in order to verify the effectiveness of the former.According to the tests on FCN,U-net,Seg Net and FCNmodify with data-intensive orthophoto images accessed from Vaihingen,Stuttgart in Germany provided by ISPRS,the overall classification accuracies turn out to be 83.5%,83.8%,83%,86%and 90.9% respectively.We can draw a conclusion that FCN-modify boasts the highest overall classification accuracy.Moreover,in terms of smallsized ground objects like vehicles as well as hard examples like background,it also proves a high classification accuracy.In general,compared with other models,FCNmodify has made great improvement in the classification of remote sensing images.This paper is applicable to the classification of high-resolution remote sensing images based on deep learning.By adding the structure of different receptive fields in the FCN model,the Focal loss function is used to improve the classification accuracy.The research in this paper is on the classification of remote sensing images.It has made contributions and greatly improved the classification accuracy.This method has certain application value in the field of remote sensing.