MRSSC: A BENCHMARK DATASET FOR MULTIMODAL REMOTE SENSING SCENE CLASSIFICATION

Abstract. Scene classification based on multi-source remote sensing image is important for interpretation, and has many applications, such as change detection, visual navigation retrieval. Deep learning become a research hotspot in the field of scene classification, dataset an driving force to promote its development. Most datasets are optical images, multimodal relatively rare. Existing that contain both SAR data, SARptical WHU-SEN-City, which mainly focused urban area without wide variety categories. This largely limits development domain adaptive algorithms classification. In this paper, we proposed multi-modal (MRSSC) Tiangong-2, Chinese manned spacecraft can acquire images at same time. The contains 12167 (optical 6155 6012 SAR, resp.) seven typical scenes, namely city, farmland, mountain, desert, coast, lake river. Our evaluated by state-of-theart adaptation methods establish baseline with average accuracy 79.2%. MRSSC will be released freely educational purpose found China Manned Space Engineering data service website (http://www.msadc.cn). fill gap between different sources, paves way generalized model earth observation data.