JL1-CC&QA
收藏资源简介:
JL1-CC&QA是由清华大学联合产业界构建的遥感变化理解基准数据集,旨在通过自然语言描述和交互式问答深化对地表变化的语义理解。该数据集基于5,000对吉林一号卫星获取的双时相高分辨率图像,共计包含17,021条质量验证的变化描述和20,060组问答对,覆盖建筑、道路、水体等多种地物变化类型。数据通过多模态大语言模型生成、视觉增强的LLM评判和专家验证的三阶段流程构建,确保了注释的准确性与多样性。该数据集主要应用于遥感变化检测、变化描述生成和视觉问答等任务,致力于解决传统像素级变化检测中语义信息缺失的问题,推动多任务协同的智能遥感分析发展。
JL1-CC&QA is a benchmark dataset for remote sensing change understanding, jointly developed by Tsinghua University and industry partners. It aims to deepen the semantic understanding of land surface changes via natural language descriptions and interactive question-answering. Built upon 5,000 pairs of high-resolution bi-temporal remote sensing images captured by the Jilin-1 satellite constellation, this dataset contains a total of 17,021 quality-validated change descriptions and 20,060 question-answering pairs, covering various land feature change types such as buildings, roads, and water bodies. This dataset is constructed via a three-stage pipeline: generation by multimodal large language models, evaluation by vision-enhanced LLMs, and expert validation, which ensures the accuracy and diversity of annotations. Primarily applied to tasks including remote sensing change detection, change description generation, and visual question answering, this dataset aims to address the issue of missing semantic information in traditional pixel-level change detection, and promote the development of multi-task collaborative intelligent remote sensing analysis.
数据集概述
数据集名称:JL1-CD
核心用途:遥感变化检测(Remote Sensing Change Detection)
相关论文:JL1-CD: A New Benchmark for Remote Sensing Change Detection and a Robust Multi-Teacher Knowledge Distillation Framework(arXiv:2502.13407)
发布时间:2025年2月23日
开源状态:已开源
下载地址
- Google Drive:https://drive.google.com/drive/folders/1ELoqx7J3GrEFMX5_rRynMjW9-Poxz3Uu?usp=sharing
- 百度网盘:https://pan.baidu.com/s/1_vcO4c5DM5LDuOqLwLrWJg?pwd=5byn
- Hugging Face:https://huggingface.co/datasets/circleLZY/JL1-CD
配套框架
- 基于 OpenCD 工具箱实现:https://github.com/likyoo/open-cd
训练与测试
环境安装
- 按照 OpenCD 仓库的安装说明进行配置。
训练流程(以 Changer-MiT-b0 模型为例)
-
训练原始模型: bash python tools/train.py configs/changer/changer_ex_mit-b0_512x512_200k_cgwx.py --work-dir /path/to/save/models/Changer-mit-b0/initial
-
训练教师模型(CAR分区,如3个分区): bash python tools/train.py configs/distill-changer/changer_ex_mit-b0_512x512_200k_cgwx-small.py --work-dir /path/to/save/models/Changer-mit-b0/small python tools/train.py configs/distill-changer/changer_ex_mit-b0_512x512_200k_cgwx-medium.py --work-dir /path/to/save/models/Changer-mit-b0/medium python tools/train.py configs/distill-changer/changer_ex_mit-b0_512x512_200k_cgwx-large.py --work-dir /path/to/save/models/Changer-mit-b0/large
- 完成上述两步后,获得原始模型及三个教师模型(小、中、大),此时可应用 O-P 策略。
-
训练学生模型:
-
在配置文件中初始化检查点路径:
checkpoint_student、checkpoint_teacher_l、checkpoint_teacher_m、checkpoint_teacher_s。 -
运行命令: bash python tools/train.py configs/distill-changer/distill-changer_ex_mit-b0_512x512_200k_cgwx.py --work-dir /path/to/save/models/Changer-mit-b0/distill
-
获得 MTKD 框架下的学生模型。
-
测试
-
学生模型(MTKD训练后): bash python test.py <config-file> <checkpoint>
-
O-P 策略测试:参考脚本
tools/test_pipline/single-partition-3-test.py。
预训练模型(Checkpoints)
- 百度网盘:https://pan.baidu.com/s/1F5MIGCCiNHFifNl_kDiklA?pwd=4tid
- Hugging Face:https://huggingface.co/circleLZY/MTKD
引用
bibtex @article{liu2025jl1, title={JL1-CD: A New Benchmark for Remote Sensing Change Detection and a Robust Multi-Teacher Knowledge Distillation Framework}, author={Liu, Ziyuan and Zhu, Ruifei and Gao, Long and Zhou, Yuanxiu and Ma, Jingyu and Gu, Yuantao}, journal={arXiv preprint arXiv:2502.13407}, year={2025} }




