Deep Sparse Representations for Land-Use Scene Classification in Remote Sensing Images

Land-use scene classification which provides high-level interpretation for remote sensing images is a challenging task, and the performance of this task mainly lies in the representative ability of extracted scene-level features. Motivated by the good generalization ability of deep convolutional neural network (CNN), we design a novel method, called deep sparse representation (DSR), building scene-level features via directly transferring CNN models pre-trained on general-purpose image dataset for land-use scene classification. We first extract deep CNN features in a feedforward way, and then compute sparse codes for the deep features through optimizing a sparse coding formulation; we finally max-pool the obtained sparse codes to form the image representations. Extensive experiments on a public land-use scene benchmark demonstrate that the image features generated by the proposed DSR, even with a simple linear classifier, can result in remarkable classification accuracies, which now become the new state-of-the-art performance on the benchmark.