Unsupervised hierarchical convolutional sparse auto-encoder for high spatial resolution imagery scene classification

Recently, efficiently representing the scenes from a large volume of high spatial resolution (HSR) images is a critical problem to be solved. Traditional scene classification problems were solved by utilizing the spatial, spectral and structural features of the HSR images separately or jointly, which lacks considering all those features of the images integrally and automatically. In this paper, we propose an efficient hierarchical convolutional sparse auto-encoder (HCSAE) algorithm considering all the features of the images integrally for scene classification, which adopts the unsupervised hierarchical idea based on the single-hierarchy convolutional sparse auto-encoder (CSAE). Compared with the single-hierarchy CSAE, HCSAE can extract more robust and efficient features containing abundant detail and structural information in the higher hierarchy for scene classification. To further improve the calculation performance and reduce the over-fitting of the network, a “dropout” strategy is adopted in this paper. The experimental results were confirmed by the UC Merced dataset consisting of 21 land-use categories, and showed that HCSAE performs better than the traditional scene classification methods and the single-hierarchy CSAE algorithm.