Improving subpixel impervious surface estimation based on point of interest (POI) data

Accurate estimation of impervious surface area (ISA) at the subpixel level is essential for understanding urbanization and its environmental impacts. In recent years, point-of-interest (POI) data has demonstrated unique value for urban studies. However, its potential for improving subpixel ISA estimation has yet to be fully realized. This research seeks to overcome the challenges of fusing POI data with remote sensing imagery and improve subpixel ISA estimation. To form an integrated sample dataset for subpixel ISA estimation, POI data were processed using kernel density analysis and transformed into continuous raster layers compatible with remote sensing imageries. The proposed method was tested in two study areas with distinctly different urban land patterns: Shenzhen, China, and Chicago, USA. Two widely used machine learning models, Classification and Regression Tree (CART) and Convolutional Neural Network (CNN), were developed based on the integrated sample dataset. The results show POI data significantly improved both models. Incorporating POI data reduced MAE by 52.75% for CART and 56.68% for CNN, and RMSE by 45.39% and 48.54%, respectively, compared to models without POI data. The fully trained POI-integrated CNN achieved the highest accuracy (MAE: 2.95, RMSE: 5.12, R²: 0.99). By achieving accurate subpixel ISA estimation with minimal additional procedures, the proposed method is expected to offer an objective and repeatable approach, providing reliable basic data for urban environmental research and planning.