Automatic Extraction of Urban Impervious Surface Based on SAH-Unet

Abstract

Increases in the area of impervious surfaces have occurred with urbanization. Such surfaces are an important indicator of urban expansion and the natural environment. The automatic extraction of impervious surface data can provide useful information for urban and regional management and planning and can contribute to the realization of the United Nations Sustainable Development Goal 11—Sustainable Cities and Communities. This paper uses Google Earth Engine (GEE) high-resolution remote sensing images and OpenStreetMap (OSM) data for Chengdu, a typical city in China, to establish an impervious surface dataset for deep learning. To improve the extraction accuracy, the Small Attention Hybrid Unet (SAH-Unet) model is proposed. It is based on the Unet architecture but with attention modules and a multi-scale feature fusion mechanism. Finally, depthwise-separable convolutions are used to reduce the number of model parameters. The results show that, compared with other classical semantic segmentation networks, the SAH-Unet network has superior precision and accuracy. The final scores on the test set were as follows: Accuracy = 0.9159, MIOU = 0.8467, F-score = 0.9117, Recall = 0.9199, Precision = 0.9042. This study provides support for urban sustainable development by improving the extraction of impervious surface information from remote sensing images.

Keywords

• impervious surface
• remote sensing
• deep learning
• SAH-Unet
• Google Earth images