Remote sensing image classification based on convolutional neural networks with two-fold sparse regularization

Abstract

Convolutional neural networks (CNNs) have shown great potential for remote sensing image classification. As the features obtained from a deep CNN generally exhibit high generalization capacity, the subsequent classifier is normally able to provide good results without the need for careful optimization. However it is well-known that, in the pursuit of high classification results, it is generally difficult to acquire a large number of training samples for the learning stage of the CNN. Therefore, it is important to exploit not only the feature generalization ability, but also the machine generalization ability for accurate classification with limited training samples. In this work, we introduce a new two-fold sparse regularization method (based on the deep CNN learning framework) for remote sensing image classification. Our proposed method, called TFCNN (for two-fold CNN), exploits the fact that the sparseness of features leads to increased linear class-separability. It also relies on the fact that the sparseness of the activation function collaborates with those features in a straightforward manner. As a result, the proposed TFCNN naturally achieves (for the first time in the literature) both feature and machine generalization, in complementary fashion. Our experimental results are conducted with three types of remote sensing data: hyperspectral images, multispectral images and synthetic aperture radar (SAR), suggesting that the proposed framework achieves excellent classification performance in all cases.