Densely Connected Convolutional Neural Network Based Polarimetric SAR Image Classification

With the development of representation learning, deep learning based methods have become the state-of-the-art method in some related fields of pattern recognition. This phenomenon brings new challenges and opportunities to polarimetric SAR image interpretation. In this paper, we propose a novel classification method for polarimetric SAR image based on a fresh technique in deep learning: DenseNet. A 20-layers (with 3 dense block and 2 transition layers) DenseNet is built to implement polarimetric SAR image classification. The proposed method effectively prevents gradient vanish and overfitting by feature reuse while automatically extracting high-level features and performing pixel-wise multi-class classification. Last but not least, the proposed method achieves the state-of-the-art experimental result on PolSAR Flevoland 15-class benchmark dataset.