An Improved Semantic Segmentation Method for Remote Sensing Images Based on Neural Network

Received: 10 November 2019 Accepted: 17 February 2020 Traditional semantic segmentation methods cannot accurately classify high-resolution remote sensing images, due to the difficulty in acquiring the correlations between geophysical objects in these images. To solve the problem, this paper proposes an improved semantic segmentation method for remote sensing images based on neural network. Based on residual network, the proposed algorithm changes the dilated convolution kernels in the dilated spatial pyramid pooling (SPP) module before extracting the correlations between geophysical objects, thus improving the accuracy of segmentation. Next, the high resolution of the input image was maintained through deconvolution, and the semantic segmentation was realized by the pixel-level method. To enhance the robustness of our algorithm, the dataset was expanded through random cropping and stitching of images. Finally, our algorithm was trained and tested on the Potsdam dataset provided by the International Society for Photogrammetry and Remote Sensing (ISPRS). The results show that our algorithm was 1.4% more accurate than the DeepLab v3 Plus. The research results shed new light on the semantic segmentation of high-resolution remote sensing images.