Impervious surfaces caused by rapid urbanization affect the environment and increase the disaster risk. Currently, most studies have extracted impervious surfaces by manual participation for training samples with medium-and-high spatial resolution remote sensing images. Therefore, it is necessary to develop a new method for improving the efficiency of training sample acquisition and the accuracy of impervious surface extraction. In this paper, a novel impervious surface extraction method is proposed based on automatically generating training samples from multisource remote sensing products. First, the preliminary sample area was constructed through the overlay analysis of the classification consistency area of three remote sensing products and homogeneous area detection based on Sentinel-2 images. Second, four spectral indices and digital surface model (DSM) elevation data were used for sample selection, and the pixels were further purified by variance purification calculation. Finally, by sample migration and random forest model training, impervious surfaces were extracted for other years with limited data. Wuhan city in China was selected as the study area due to a large number of interior objects for impervious surfaces. Sentinel-2 images from 2018 to 2020, three 30 m-resolution products, and DSM data in 2018 were used. The proposed method's extraction accuracies of impervious surfaces for Wuhan in 2018, 2019, and 2020 are 94.02%, 94.45%, and 93.87%, respectively. Additionally, with the resolution improved up to 10 m, the method is more conducive to distinguishing the boundary between impervious surfaces and pervious surfaces.
Read full abstract