High-Resolution Remote Sensing Image Land Cover Classification Based On Deep Learning And Multi-Scale And Multi-Feature Fusion

Remote sensing image land cover classification is an important research issue,and its scientific significance lies in how to further improve the accuracy of classification.It is significant for land use planning,urban management and environmental monitoring.With the development of high spatial resolution satellite sensors,high-resolution remote sensing images are becoming more and more extensive.The clear object features and complex spatial characteristics further increase the difficulty of land cover classification.Therefore,it is important to develop new classification methods and improve the ability of feature extraction and object recognition.In order to solve the problems of insufficient effect analysis and fusion depth and lack of contribution analysis in the existing remote sensing image classification methods with deep learning,in this study we combine superpixel segmentation and deep learning models together.We design a more comprehensive scale combining solution and propose a multi-scale classification method with deep network fusion and a multi-scale and multifeature classification method with attention mechanism.We comprehensively analyze the fusion effect and contribution difference of multi-scale and multi-type features and improve the accuracy of land cover classification,to form a high-resolution remote sensing land cover classification method system based on deep learning and multi-scale and multi-feature fusion.Taking the land cover classification of Gao Fen-2 remote sensing images in different areas of Zhejiang Province as an example,the proposed method is tested,applied and verified.The research content of this article is summarized as follows:(1)Focusing on the basic problem of effectively applying superpixel segmentation and deep learning models to remote sensing image classification,a single-scale land cover classification method for superpixels is proposed.We study the superpixel segmentation method based on false color composition and image enhancement and design the superpixel single-scale convolutional network using the original and transformed images,respectively,to explore the effect of various scales.The proposed method is fully demonstrated through experimental analysis.(2)Aiming at the key question of comprehensively exploring the multi-scale effect and achieving deep feature fusion,a multi-scale land cover classification method with deep network fusion is proposed.We study the complementary effects of different multiscale combining solutions,and design deep fusion with one-dimensional convolutional networks and recurrent networks for multi-scale land cover classification.In experiment,the accuracy of proposed method is superior to that of feature merging method.(3)Aiming at the key problem of realizing the multi-feature fusion with attention difference,a multi-scale and multi-feature land cover classification method with attention mechanism is proposed.We study complementary effects of different multi-feature combining solutions,and design deep fusion with attention mechanism and deep network for multi-scale and multi-feature land cover classification.In experiment,the accuracy and boundaries of proposed method is superior to that of other comparison methods.(4)Focusing on the key issue of improving the accuracy of remote sensing image transfer classification based on comprehensive feature fusion,a multi-scale and multifeature transfer learning method for land cover classification is proposed.We study the effects of different feature combining solutions in transfer learning,and design a multiscale and multi-feature deep transferable classification network with model weights.In experiment,the multi-scale and multi-feature transfer learning achieves better results.This paper is expected to solve the problem of deep fusion of multi-scale and multifeature information in remote sensing images,thereby improving the accuracy and consistency of land cover classification,promoting the research development of remote sensing classification,and promoting the application of remote sensing in industries such as geographic national census.