Extraction of impervious surface from high-resolution remote sensing image based on deep learning

The traditional impervious surface extraction algorithm is difficult to keep balance on accuracy and efficiency. Aim to solve this problem, We propose a semantic segmentation model based on deep learning (called Dense-SegUNet). Our contributions are as follows: (1)The efficient encoding and decoding structure of UNet is introduced as the skeleton of Dense-SegUNet to realize semantic segmentation. (2)The advantages of DenseNet are refined, and four Denseblock with different depths and three Transition layer are used as encoding parts for feature extraction. This structure can realize feature reuse, strengthen the transmission of deep information and avoid gradient disappearance. (3)The hard version of swish (h-swish) activation function and test-time enhancement algorithm (TTA) are introduced to improve the accuracy of network model. We compared Dense-SegUNet with traditional classification models and other deep learning models. The results show that the Kappa coefficient of the new model is 15.2% higher than the RF model (0.9530vs0.8272), and 2.7% higher than U-Net (0.9530vs0.9278).