Intelligent Classification Of Typical Landcover In High-resolution Optical Remote Sensing Images Based On Deep Learning Methods

The typical landcover classification in high-resolution optical remote sensing images is a fundamental and important task in the field of remote sensing image processing.The high precision of typical landcover classification,such as waters,roads and residential areas,has important application value in topographic mapping,intelligence surveillance,urban planning and disaster emergency response.In recent years,the amount and type of remote sensing data have been increasing rapidly,so the traditional classification methods cannot meet the growing application requirements in precision and efficiency.The rise of deep learning technology provides a new technical approach for the landcover classification in remote sensing images and promotes the development of intelligent classification technology.Therefore,based on the theory of deep learning,this thesis focuses on the research of the intelligent landcover classification technology.The main work and innovations are as follows:1.The research status of landcover classification in high-resolution optical remote sensing images and the progress of semantic segmentation technology is summarized.In addition,the relevant basic theories and non-local structures are introduced.The thesis also analyzes the problems,which exist in the landcover classification methods that based on deep learning.2.A novel architecture of the network called double vision fully convolutional network is designed for object extraction in optical remote sensing images.By combining with different visions,the confidence of the predicted label map in the central region is improved,and the convergence rate of network training is accelerated.Experimental results show that the training time and computation cost are greatly saved without the decline in accuracy.Furthermore,the architecture provides a new idea for the design of classification network.3.Aiming at the classification for multi-class in remote sensing images,the multilevel attention fusion U-Net is proposed.This network combines the advantage of feature fusion of u-shaped convolutional network,and the attention mechanism structure is added to deal with the deep features.The experimental results on remote sensing image datasets of different scales show that the proposed network can achieve the higher classification accuracy than the other semantic segmentation networks,which means the network is highly practical.4.This thesis analyzes the generalization of convolutional neural network in the landcover classification of optical remote sensing images and puts forward the corresponding solution of adding texture features into the spectral features.The results of different remote sensing image datasets verify that the texture features can strengthen the ability of learning and improve the generalization of feature classification.Meanwhile,the scheme verifies the effectiveness of manually designed features for the improvement of classification.