Built-up area extraction in PolSAR imagery using real-complex polarimetric features and feature fusion classification network

Extraction of built-up areas from polarimetric synthetic aperture radar (PolSAR) images plays a crucial role in disaster management. The polarimetric orientation angles (POAs) of built-up areas exhibit diversity, and built-up areas with POA close to 45 degrees are often misclassified as vegetation. To address this problem, a polarimetric feature suitable for the extraction of built-up areas with large POAs is first designed, and a mixed real-complex-valued polarimetric feature combination is constructed. Then, a real-complex and spatial feature fusion classification network (RCSFFCNet) is designed. In which the proposed mixed real-complex-valued residual structure can efficiently extract mixed numerical features. Additionally, a multi-local spatial convolutional attention module is designed and embedded to efficiently fuse mixed numerical features, as well as superpixel multi-local spatial features. Experiments were conducted using PolSAR images from Gaofen-3, Radarsat-2, and ALOS-2/PALSAR-2. The experimental results show that the feature combination proposed in this paper increases the F1 score of builtup areas by approximately 2%-3%, and the F1 score of built-up areas extracted using the RCSFFCNet also improves by about 2%-3%, with F1 scores exceeding 95%. On all three datasets, the proposed method achieves the best performance in extracting built-up areas with various POAs, indicating overall superiority from feature selection to model implementation.