Semantic segmentation of remote sensing images based on deep learning methods

Abstract

Remote sensing image segmentation has always been an important research direction in the field of remote sensing image processing, and it is a key step in the further understanding and analysis of remote sensing images. Image semantic segmentation is the process of classifying each pixel to form several sub-regions with respective characteristics, and extracting the objects of interest among them. However, due to the complex boundary and scale difference of the remote sensing image, the traditional algorithm can not meet the actual needs well, resulting in low segmentation accuracy. In order to further improve the accuracy of remote sensing image segmentation, this paper combines deep convolutional neural network with remote sensing image, based on the U-Net, firstly compares the model's segmentation accuracy under different learning strategies, and introduces a new learning strategy to improve the learning effect of the model; secondly, in the loss function part of the model, a new compound loss function is proposed to speed up the convergence of the network and improve the segmentation accuracy. Based on full experimental research on the WHDLD remote sensing image dataset, the results show that the improved method has 1.5% accuracy improvement compare to the U-Net.