Remote Sensing Image Change Instruction dataset (RSICI); Remote Sensing Image Change Preference dataset (RSICP)
收藏资源简介:
RSICI与RSICP是由中国科学院计算技术研究所等机构联合构建的遥感图像变化描述专用指令与偏好数据集。该数据集规模约4万条样本,其内容基于LEVIR-CD、SYSU-CD、S2Looking和CDD等多源遥感变化检测数据,通过Qwen-VL-Max大模型结合二值掩膜几何先验自动生成精细化文本描述。数据构建过程采用算法化指令生成范式,并经过关键词过滤与专家校验确保质量,旨在为遥感图像变化描述任务提供大规模、高质量的训练与评估基准,推动大模型在环境监测、灾害评估等领域的细粒度时空理解能力突破。
RSICI and RSICP are specialized instruction and preference datasets for remote sensing image change description, jointly constructed by the Institute of Computing Technology of the Chinese Academy of Sciences and other institutions. The dataset contains approximately 40,000 samples, with its content sourced from multi-source remote sensing change detection datasets including LEVIR-CD, SYSU-CD, S2Looking, and CDD. Fine-grained textual descriptions are automatically generated using the Qwen-VL-Max Large Language Model (LLM) combined with binary mask geometric priors. The dataset construction adopts an algorithmic instruction generation paradigm, and undergoes keyword filtering and expert verification to ensure data quality. It aims to provide a large-scale, high-quality training and evaluation benchmark for remote sensing image change description tasks, and promote breakthroughs in the fine-grained spatio-temporal understanding capabilities of LLMs in fields such as environmental monitoring and disaster assessment.
数据集概述
RSICCLLM 是一个用于遥感图像变化描述(Remote Sensing Image Change Captioning)的多模态大语言模型数据集。
核心任务
- 输入:双时相遥感图像(bi-temporal remote sensing images)
- 输出:对场景变化的自然语言描述
数据集状态
- 当前正在整理中,即将公开发布
代码与配套资源
该数据集对应的模型实现、训练与推理代码、评估协议以及预训练检查点将一并公开。
引用信息
bibtex @inproceedings{wang2026rsiccllm, title = {RSICCLLM: A Multimodal Large Language Model for Remote Sensing Image Change Captioning}, author = {Wang, Yelin and others}, booktitle = {European Conference on Computer Vision}, year = {2026} }

- 1RSICCLLM: A Multimodal Large Language Model for Remote Sensing Image Change Captioning中国科学院·计算技术研究所人工智能安全国家重点实验室; 大湾区大学; 哈尔滨工业大学·深圳; 东莞市智能与信息技术重点实验室 · 2026年



