PolSAR image classification with small sample learning based on CNN and CRF

Convolutional neural networks (CNN) have achieved great success in the optical image processing field. Hence, methods based on CNN are introduced into PolSAR image classification. Usually CNN needs a lot of training samples, but the cost of collecting ground truth data and making labels is very high. Our goal is to increase training samples by repeating learning processes with small sample learning technique. The proposed method used in this study is CNN and conditional random fields(CRF), which combines the structured modeling ability of CRF and the feature extraction advantage of CNN. On base of CNN and CRF, the framework of small sample learning is developed. The experimental data are two AIRSAR datasets. The paper will analyze the appropriate ratio of samples for small sample learning in the whole dataset. The results show that for these two data sets, when the ratio is 0.5%, small sample learning can achieve very high classification accuracy. It is similar to the accuracy of other methods which need at least 3% samples for training.