Supervised classification is a vital approach to extract impervious surface areas (ISA) from satellite images, but the training samples need to be provided through heavy manual work. To address it, this study proposed an automatic method to generate training samples from high-resolution night light data, considering that nighttime lights generated by human activities is strongly correlated with impervious surface. First, positive and negative samples for ISA were located according to the distribution of nighttime lights. Second, the feature sets were constructed by calculating the spectral and texture feature from the OLI images. Third, an ensemble ELM classifier was selected for ISA classification and extraction. Four large cities were selected as study areas to examine the performance of the proposed method in different environment. The results show that the proposed method can automatically and accurately acquire ISA with an overall accuracy higher than 93% and Kappa coefficient higher than 0.87. Furthermore, comparative experiments by biophysical composition index (BCI)and classification by manual sample were conducted to evaluate its superiority. The results show that our method has better separability for ISA and soil than the BCI. In general, the proposed method is superior to manual methods, except Harbin mostly because some impervious surfaces with weak light intensity are selected as negative samples.