Exploring CNN for urban area Extraction from Polarimetric SAR data

The convolutional neural networks CNNs have been widely used in pattern recognition and image processing in the last ten years. This paper intends to identify the urban area from polarimetric Synthetic aperture radar (PolSAR) data using the CNN model. This model has shown its effectiveness in classification. For this purpose, we propose to use a feature vector that contains the elements of the circular matrix and the calculated parameters from Cloude decomposition (entropy, alpha angle and anisotropy). The used elements are roll invariants, which make them not sensitive to the orientation angle, which can create ambiguity between oriented buildings and vegetation. To test the performance of the proposed method, ALOS-2/PALSAR-2L-band and AIRSAR C-band over the San-Francisco, USA were used. The results show that man-made targets could be effectively discriminated from natural scenes using the proposed method.