Deep learning is an important method for extracting impervious surfaces (IS), which has the advantages of high accuracy and generalization. However, the training of the models relies on a huge of train samples. Especially in large-scale and high-resolution IS mapping, it is time-consuming and laborious to obtain sufficient and high-quality training samples. Therefore, this study proposes a method to automatically extract IS based on multi-source remote sensing images and open-source data. Firstly, training samples are automatically obtained from crowdsourced OpenStreetMap data, and then the noise samples are weighted with open-source IS maps to reduce the negative influence of label noise on model training. Moreover, an ultra-lightweight CNN model with three branches was constructed to generate 10 m IS products by integrating optical, SAR and terrain data. In this paper, the method was validated in Vietnam. The results show that the overall accuracy and Kappa coefficient of the method proposed are 91.01% and 0.82, respectively, which are better than the currently released IS products. The research results of this paper can provide basic technology and data support for the sustainable development and ecological environment protection of tropical and subtropical cities in the Lancing-Mekong River basin. © 2023 SinoMaps Press. All rights reserved.
