Analysis of the inter-dataset representation ability of deep features for high spatial resolution remote sensing image scene classification

摘要

Recently, scene based classification has become a new trend for very high spatial resolution remote sensing image interpretation. With the advent of deep learning, the pretrained convolutional neural networks (CNNs) have been proved effective as feature extractors for scene classification tasks in the remote sensing domain, but the potential characteristics and capabilities of such deep features have not been sufficiently analyzed and fully understood. Facing with complex remote sensing scenes with huge intra-class variations, it is still not clear about the limitation of these powerful deep features in exploring essential invariant attributes of remote sensing scenes of the same kind but, in most cases, from separate sources. Therefore, this paper makes an intensive investigation in the feature representation ability of such deep features from the aspect of inter-dataset scene classification of remote sensing images. Four well-known pretrained CNN models and three different commonly used datasets are selected and summarized. Firstly, deep features extracted from various intermediate layers of these models are compared. Then, the inter-dataset feature representation ability is evaluated using cross-classification of different datasets and discussed in terms of imaging spatial resolution, image size, model structure, and time efficiency. Finally, several instructive findings are revealed and conclusions are drawn regarding the strength and weakness of the CNN features in the application of remote sensing image scene classification.