Semantic Segmentation for High Spatial Resolution Remote Sensing Images Based on Convolution Neural Network and Pyramid Pooling Module

Abstract

Semantic segmentation provides a practical way to segment remotely sensed images into multiple ground objects simultaneously, which can be potentially applied to multiple remote sensed related aspects. Current classification algorithms in remotely sensed images are mostly limited by different imaging conditions, the multiple ground objects are difficult to be separated from each other due to high intraclass spectral variances and interclass spectral similarities. In this study, we propose an end-to-end framework to semantically segment high-resolution aerial images without postprocessing to refine the segmentation results. The framework provides a pixel-wise segmentation result, comprising convolutional neural network structure and pyramid pooling module, which aims to extract feature maps at multiple scales. The proposed model is applied to the ISPRS Vaihingen benchmark dataset from the ISPRS 2D Semantic Labeling Challenge. Its segmentation results are compared with previous state-of-the-art method UZ_1, UPB and three other methods that segment images into objects of all the classes (including clutter/background) based on true orthophoto tiles, and achieve the highest overall accuracy of 87.8% over the published performances, to the best of our knowledge. The results validate the efficiency of the proposed model in segmenting multiple ground objects from remotely sensed images simultaneously.