LAND USE CLASSIFICATION BASED ON MULTI-STRUCTURE CONVOLUTION NEURAL NETWORK FEATURES CASCADING

Abstract

Learning efficient image representations is at the core of the classification task of remote sensing imagery. The existing methods for solving image classification task, based on either feature coding approaches extracted from convolution neural networks(CNNs) or training new CNNs, can only generate image features with limited representative ability, which essentially prevents them from achieving better performance. In this paper, we investigate how to transfer features from these successfully pre-trained CNNs for classification. We propose a scenario for generating image features via cascading features extracted from different CNNs. First, pre-trained CNNs, like CaffeNet, VGG-S and VGG-F, are used as feature extractor since their different structures help extract richer information of images. Then the fully-connected layers of the pre-trained CNNs are fine-tuned with UC Merced land use dataset. Finally, the image features generating from cascading the outputs of three networks above, are fed into multi-class Optimal Margin Distribution Machine (mcODM) to obtain the final classification results. Extensive experiments on public land use classification dataset demonstrates that the image features obtained by the proposed scenario can result in remarkable performance and improve the state-of-the-art by a significant margin. The results reveal that the features from pre-trained CNNs generalize well to land use dataset and are more expressive than features from single CNN.