DCD-Net: a DenseASPP‑enhanced network with strip pooling and attention for impervious surface extraction from 10 m remote sensing imagery

Rapid urbanization has intensified the need for accurate impervious surface data to support urban planning and environmental management. (1)Background: While remote sensing imagery is widely applied for impervious surface extraction, existing methods still face challenges in preserving spatial details and refining complex boundaries. (2)Methods: To address these limitations, we developed a Densely Connected Deep Convolutional Impervious Surface Extraction Model (DCD-Net), which replaces the Atrous Spatial Pyramid Pooling (ASPP) module in DeepLabv3+ with DenseASPP to capture multi-scale features through dense connections of dilated convolutions, enhancing contextual information utilization. Strip pooling and attention mechanisms are integrated to improve long-range dependency modelling and feature representation, enabling sharper delineation of impervious surface boundaries. (3) Results: Using the Henan section of the Yellow River Basin with Esri_Land_Cover_2022_10m labelled data, DCD-Net achieved an overall accuracy of 94.03% and a mean intersection over union (MIoU) of 84.03%, outperforming MA-Net, UNet and DeepLabv3+. (4) Conclusions: These findings demonstrate the effectiveness of DCD-Net in enhancing impervious surface extraction accuracy and indicate its potential as a reliable tool for urban ecological monitoring and sustainable urban development.