Automatic impervious surface mapping in subtropical China via a terrain-guided gated fusion network

Large-scale, high-resolution and multi-temporal impervious surface maps from remote sensing are essential for socioeconomic and environmental studies. However, in complex regions like subtropical China, climate and terrain severely contaminate optical and synthetic aperture radar (SAR) imagery, resulting in relatively poor accuracy here. Therefore, this paper presents a novel network for accurately extracting impervious surfaces, in which a terrain-guided gated fusion module is proposed to adaptively fuse Sentinel-2 optical and Sentinel-1 SAR imagery. To eliminate the tedious manual annotation over vast and mixed regions, an automatic samples generation and weighting strategy is further introduced. Then, the framework monitors impervious surfaces in subtropical China from 2016 to 2021 with the support of Google Earth Engine. The final impervious surface map scores an overall accuracy and Kappa of 93.19-94.04% and 0.843-0.859, respectively. Moreover, its consistent performance in various scenarios and superior visual details than most products evidence the feasibility and reliability of the proposed framework. This map provides precise dynamics of impervious surfaces that will help to inform urban studies.