Metric-based Meta-Learning Model for Few-Shot PolSAR Image Terrain Classification

Abstract

Deep learning models have been successfully applied to polarimetric synthetic aperture radar (PolSAR) image terrain classification. However, they usually need many labeled samples from the current PolSAR image to be classified to train the model parameters, which fundamentally limits their application. In this paper, we propose a metric-based meta-learning framework for few-shot terrain classification (MML-FSTC) of PolSAR image. The proposed MML-FSTC firstly utilizes the collected base labeled samples (BLS) from typical PolSAR systems to learn an effective transferable three-dimensional convolutional neural network (3DCNN) to extract features for terrain classification. Then, by mimicking the few-shot learning of human, MML-FSTC integrates additional samples, whose label spaces are probably different from those of BLS, and proposes a meta-learning strategy for terrain classification that consists of multiple few-shot training episodes, each of which has very few labeled terrain samples and potentially different labeled space. By minimizing the cosine distance of the support and query deep features, MML-FSTC finetunes the pre-trained 3DCNN in the meta-learning stage, and utilizes a metric-based cosine distance function to predict the classification. The classification results demonstrate the superiority of MML-FSTC for few-shot PolSAR image terrain classification.