PolSAR Image Classifiacation Based on Deep CNN and Adaboost

Polarimetric Synthetic Aperture Radar (PolSAR) image classification aims at classifying resolution cells into homogeneous groups according to physical property. Deep learning is developing rapidly in the recent years and has been utilized in PolSAR classification. However, the utilization of the deep features is still limited, inadequate use of which makes convolutional neural networks (CNNs) suffer from over-fitting problem in small training dataset and produce unsatisfactory results. In this paper, a novel classification algorithm is proposed based on an Adaptive boost (Adaboost) deep learning method. The ensemble learning schedule is introduced to CNN, which enhances the robustness and adaptability of the deep mode. Besides, the novel model has less parameters, thereby equipping with better performance compared with the CNNs. The experimental results on real PolSAR dataset verify the effectiveness and the superiority of the proposed method, and demonstrate that it can provide stronger noise immunity and obtain smoother homogeneous areas in classification.