Extraction of built-up areas from nighttime light images based on improved DeepLabV3+ network

Abstract

The extraction of urban built-up areas based on nighttime light images by deep learning algorithms is a new exploration in remote sensing research in recent years. An improved DeepLabV3+ network is proposed to address the phenomenon that ordinary convolutional neural networks processing remote sensing images will lose a large amount of detail information in the coded feature extraction stage, which in turn leads to poor edge segmentation and low accuracy. The network performs 2D decomposition of the asymmetric convolution in the ADSPP convolution layer, and then combines it with the null convolution to form an asymmetric null convolution for feature extraction, capturing more features by enhancing the skeleton part of the convolution kernel to improve the classification accuracy of urban built-up areas without increasing the computing time. This paper shows that the improved DeepLabV3+ network is more objective in characterizing urbanization development than the original DeepLabV3+ network in terms of the extent of built-up areas extracted from night-time light images.