High Resolution Remote Sensing Image Classification Based On Multi-class Features Deep Learning

With the start of high-resolution earth observation systems in China,the high spatial resolution(hereinafter referred to as "high resolution")remote sensing image data also show a huge growth.High resolution remote sensing image has the characteristics of high spatial resolution,high image complexity,lack of spectral information,large intra-class differences and rich geometric texture.The traditional classification method of remote sensing image classification is not ideal,so it is urgent to find a high-precision,fast and efficient image classification method.In recent years,deep learning technology has developed rapidly.To solve this problem,this paper applies deep learning technology to high-resolution remote sensing image classification.Considering the lack of spectral information of high-resolution remote sensing images,so many types of image features have been constructed as a complement to it in this paper.And a high-resolution remote sensing image classification method based on multi-class feature deep learning is proposed.The following tasks have been accomplished:Firstly,the research status of deep learning technology and remote sensing image classification at home and abroad is expounded.The commonly used remote sensing image classification methods are introduced,and the advantages of deep learning methods in high resolution remote sensing image classification are pointed out.Secondly,the deep learning classification method of high-resolution remote sensing image is introduced.And several common image classification deep models are expounded.Through detailed comparison and analysis,the better U-Net model is selected for subsequent high-resolution remote sensing image classification research.Thirdly,the multi-class feature construction of high-resolution remote sensing images is introduced.Considering that the spectral information of high-resolution remote sensing images is deficient,but it contains rich information such as space and texture,Therefore,four image features are constructed in this paper,namely,image spatial feature,image contrast feature,image vegetation feature and image texture feature,which are used as supplements to the original spectral information for subsequent deep learning training.In this paper,the digital surface model is used as the image spatial feature;the HC algorithm based on global contrast and the AC algorithm based on local contrast are compared and analyzed in detail,and the AC algorithm based on local contrast is selected to extract image contrast feature combined with the practical application;the normalized vegetation index is constructed to extract the image vegetation feature;the round LBP algorithm is used to extract image texturefeature.Fourthly,a high-resolution remote sensing image classification method based on multi-class feature deep learning is proposed.The U-Net model is used as the basic deep learning model,and the constructed image multi-class features are input into the model for training.Then the classification map is obtained by the prediction of test set.The conditional random field method is used to post-process the classification map.Because of the construction of four types of image features,four classification maps will be obtained relatively.According to certain fusion rules,several classification maps are fused as the final classification results.Fifthly,the experimental results are summarized and analyzed.Quantitative accuracy evaluation of the classification results of this method is carried out by using multiple accuracy evaluation indexes.The effects of image multi-class features and classification map fusion on classification results are analyzed in detail.Then the method is compared with several classic deep learning models,namely FCN-32 s,FCN-16 s,FCN-8s,SegNet and original U-Net models.The experimental results show that this method has high accuracy and good effect,and can classify high-resolution remote sensing images very well.