Abstract: Research on the identification and classification of urban underlying surface elements based on optical remote sensing images has attracted significant attention.However, the currently available optical remote sensing image datasets for underlying surface classification suffer from issues such as low data source accuracy, limited classification categories, and lack of standardization.These problems make it difficult to meet the research needs for land and sea use classification and sponge city construction.To address this, this paper aims to meet the research needs for fine classification of underlying surface elements for urban hydrological simulation.We construct a high-resolution remote sensing image urban underlying surface fine classification dataset (DUSC-7)based on aerial images with a resolution of 0.1 meters.We extract the underlying surface elements from the images to create sample slices, and perform semi-automatic annotation with reference to the results of the third national land survey and topographic maps.This results in a classification dataset of urban underlying surface elements containing 7 categories and 8859 instances.The images for each category in the dataset are randomly divided into test and training sets in a 3∶7 ratio, and validation experiments are conducted.The experimental results show that, in the verification of the effectiveness of general classification models, the overall test accuracy of the current mIoU models achieve more than 0.648 8.The constructed DUSC-7 dataset can effectively meet the verification requirements for urban underlying surface element classification algorithms.

Key words: underlying surface, classification, remote sensing images, dataset, sponge city