Abstract
In this work, a discriminatively learned CNN embedding is proposed for remotesensing image scene classification. Our proposed siamese network simultaneouslycomputes the classification loss function and the metric learning loss functionof the two input images. Specifically, for the classification loss, we use thestandard cross-entropy loss function to predict the classes of the images. Forthe metric learning loss, our siamese network learns to map the intra-class andinter-class input pairs to a feature space where intra-class inputs are closeand inter-class inputs are separated by a margin. Concretely, for remotesensing image scene classification, we would like to map images from the samescene to feature vectors that are close, and map images from different scenesto feature vectors that are widely separated. Experiments are conducted onthree different remote sensing image datasets to evaluate the effectiveness ofour proposed approach. The results demonstrate that the proposed methodachieves an excellent classification performance.