Urban area identification from Polarimetric Synthetic Aperture Radar data using a deep convolutional neural network

Identification and extraction of urban built-up areas from Polarimetric Synthetic Aperture Radar (PolSAR) data remain challenging due to complex scattering mechanisms and sensor-dependent characteristics, which often degrade classification accuracy. To address this issue, this paper proposes a novel deep learning framework specifically designed for PolSAR urban area detection. The proposed method is based on a dual-branch convolutional neural network architecture that jointly exploits complementary polarimetric information. The first branch processes roll-invariant elements derived from the circular covariance matrix, while the second branch extracts scattering-related features obtained from modified four-component and Cloude decompositions. The features learned by both branches are concatenated and adaptively recalibrated using a squeeze-and-excitation (SE) block, followed by residual learning blocks to enhance discriminative representation and training stability. The effectiveness of the proposed framework is evaluated for binary urban and non-urban classification using PolSAR datasets acquired by different sensors, including RADARSAT-2 (C-band), ALOS-2/PALSAR-2 (L-band), and AirSAR (C-band). Experimental results demonstrate that the proposed method achieves an overall accuracy of 97.27% with a kappa coefficient of 93.09% on the ALOS-2/PALSAR-2 dataset, and 95.86% overall accuracy with a kappa coefficient of 89.89% on the AirSAR dataset, showing competitive or superior performance compared to widely used CNN-based baselines. These results confirm the robustness and effectiveness of the proposed feature fusion and channel recalibration strategy for PolSAR urban area detection through diverse sensor configurations.