基于残差U-Net的遥感影像土地覆盖类型自动分割

研究目的:土地覆盖的准确分割对于土地调查和规划具有重要意义。针对传统方法对于高分辨率遥感影像分割存在精度和效率较低等问题,提出了深度学习遥感影像分割方法。研究方法:以2 m高分辨率遥感影像为数据源,选用一种加入残差块的U-Net模型(ResU_Net),对目标区域进行基于深度学习的土地覆盖分割,并与SVM、PSPNet、U-Net分割方法进行对比。研究结果:ResU_Net能够更加准确地表达高分辨率遥感影像的地物信息,该方法总体分割精度达到85.50%,Kappa系数为0.7603,总体精度和Kappa系数均高于SVM、PSPNet和U-Net分割方法(总体精度:ResU_Net(85.50%)>U-Net(79.44%)>PSPNet(78.90%)>SVM(66.80%))。研究结论:ResU_Net模型对高分辨率遥感影像的土地覆盖分割效果更优。

Automatic Segmentation of Land Cover Types in Remote Sensing Image Based on Residual U--Net

The purpose of this paper is that the accurate segmentation of land cover is of great significance to land survey and planning.A method of deep learning remote sensing image segmentation is to be proposed to solve the problems of low precision and low efficiency of traditional segmentation methods for high resolution remote sensing images.The research methods are as follows.This paper takes 2 m high-resolution remote sensing images as the data source,selects a U-Net model(ResU_Net)with residual block,conducts land cover segmentation based on deep learning in the target area,and compares it with SVM,PSPNet and U-Net segmentation.The results indicate that ResU_Net can more accurately express the ground object information of high-resolution remote sensing images.The overall segmentation accuracy of this method is up to 85.50%,and the Kappa coefficient is 0.7603.Both the overall accuracy and Kappa coefficient are higher than those of SVN,PSPNet and U-Net segmentation(OA:ResU_Net(85.50%)>U-Net(79.44%)>PSPNet(78.90%)>SVM(66.80%)).In conclusion,the U-Net model with residual block is more effective in land cover segmentation of highresolution remote sensing images.