The deep learning models especially the CNN have achieved amazing performance on natural image retrieval. However, remote sensing images captured with anamorphic lens are still retrieved via manual selection or traditional SIFT-based methods. How to leverage the advanced CNN models for remote sensing image retrieval is a new task of significance. This paper focuses on the aurora images captured with all-sky-imagers (ASI). By analyzing the imaging principle of ASI and characteristics of aurora, a salient region determination (SRD) scheme is proposed and embedded into the Mask R-CNN framework. Thus, we can regard an image as a "bag" of salient regions (BoSR). In practice, each salient region is represented with a CNN feature extracted from the SRD embedded Mask R-CNN. After clustered to generate a visual vocabulary, each CNN feature is quantized to its nearest center for indexing. In the stage of online retrieval, by computing the similarity scores between query image and all images in the dataset, ranking results can be obtained and image with the highest value is exported as the top rank. Extensive experiments are conducted on the big aurora data, and the results demonstrate that the proposed method improves the retrieval accuracy and efficiency.