A Height Attention Network for Satellite Video Classification

With video satellites coming into service, it is convenient to obtain video satellite imagery with continuous angles for dynamic observation of the Earth. Compared to traditional re-mote sensing data, satellite video data shows advantages over land cover classification because of extra angular information. However, most of existing deep learning based remote sensing classification methods were proposed for hyperspectral image classification. These models do not perform as good as expected in satellite video data classification because they do not take the extra dimension into consideration. To tackle this problem, a novel height attention network is proposed to improve the satellite video data classification performance. In particular, the 4D convolution is employed to jointly capture angular-spectral-spatial features. Furthermore, a height attention module is used to extract the discriminative height features after principal component analysis. The experiment results on real video data sets demonstrate that the proposed model perform competitive classification results with shorter training time.