Polarimetric SAR Landcover Classification Based on CNN with Dimension Reduction of Feature

To overcome the dimensionality disaster and feature redundancy problem, caused the Polarization feature dimensions increasing, in the deep learning network training for landcover classification using polarimetric SAR image, a novel polarimetric SAR image classification architecture using DeepLabV3 after twice dimensionality reduction with wavelet fusion and PCA is proposed. The three-layer wavelet decomposition and feature fusion and PCA processing are implemented before DeepLabV3 training to reduce the feature dimension. The full-polarization SAR data of Gaofen-3 in Baihu Farm area was used for verification. Compared with traditional method, the method proposed in this paper improved significantly in classification accuracy, which proved the effectiveness of the new method.