Search

Targeting the subtropical climate characteristics of the Guangdong - Hong Kong - Macao Greater Bay Area, this study acquired the images of the experimental area from the TerraSAR - X Radar remote sensing satellite. Given the varying scale of the surface feature targets in the Radar satellite observation scenes, this study proposed an ENet convolution spatial pyramid pooling module ( ENet - CSPP) model for surface feature classification. Since ordinary convolution can more effectively maintain domain information than atrous convolution, this study proposed a multi - scale feature fusion module based on convolution spatial pyramid pooling. Since there were a few training samples in the SAR remote sensing image dataset, this study combined the multi - scale feature fusion module with the lightweight convolutional neural network. The encoder of the ENet - CSPP network consisted of an improved ENet network and the convolution spatial pyramid pooling module. The decoder output surface feature classification images after the fusion of deep and shalow features. The quantitative comparison experiments were conducted on the GDUT - Nansha dataset. The ENet - CSPP model outperformed other models in three performance indices, namely pixel accuracy, average pixel accuracy, and mean intersection over union. This result indicates that the multi - scale lightweight model effectively improved the accuracy of surface feature classification. [ABSTRACT FROM AUTHOR]