Automated Extraction of Impervious Surface Area Using Hyperlocal Samples From Multisource Data Fusion Across Economic–Geographic Zones

Due to their complex spatial structure and spectral heterogeneity, mapping impervious surface area (ISA) at large scales with high resolution remains challenging. Key issues include labor-intensive sample collection, suboptimal feature selection that often fails to account for regional climate and urban development differences, and the lack of up-to-date ISA products at 10 m resolution. To address these limitations, we propose an automated ISA extraction method based on economic–geographic zoning and multi-source data fusion using hyperlocal samples. The study area is first divided into zones based on climate zones and urban development levels. Within each zone, hexagonal units are defined to enable localized sample selection and feature optimization. Training samples are automatically generated using rule-based and threshold-based methods, incorporating land use, land cover, and human activity information from OpenStreetMap, Sentinel-2 imagery, and Black Marble nighttime light (NTL) imagery. A total of 45 features are constructed, with NTL features playing an innovative role in improving classification performance. The Recursive Feature Elimination with Cross-Validation (RFECV) algorithm is applied to select the most relevant features for each zone. Based on the Google Earth Engine (GEE) platform, a regionally adaptive random forest model is implemented to produce a 10-meter-resolution ISA map of the study area. The results show that the proposed hyperlocal sampling strategy significantly improves sample quality. Zonal feature selection not only simplifies the model but also enhances interpretability and classification accuracy. Quantitative analysis confirms that NTL features are among the most significant inputs. The proposed method achieves an overall accuracy of 90.72%, outperforming existing ISA products GAIA, FROM_GLC10, GHSL, and ESRI_LandCover by 7.12%, 3.02%, 5.07%, and 2.47%, respectively, demonstrating its effectiveness.