"Multimodal remote sensing dataset MRSD120k"

"MRSD120K is a multimodal remote sensing dataset that is used to enhance the model\u2019s capacity to understand and execute heterogeneous subtasks with diverse requirements. MRSD120K comprises approximately 120,000 multimodal image pairs  (optical\u2013SAR and optical\u2013DSM), along with correspond-ing image metadata and task instructions. The images are collected from various benchmark datasets, such as BigEarthNet-MM3, So2Sat-LCZ42, ISPRS Potsdam, WHU-OPT-SAR, Hi-BCD, and CAU-Flood. To implement an efficient training process, the original multimodal imagery is processed through unified cropping and format normalization. The task instructions constructed based on the task types and objectives of the source datasets; and a suite of interrelated subtask sequences automatically generated by a vision-language model (VLM) using customized spatiotemporal and functional prompts."