LandGPT：支持土地属性推理的多模态大语言模型

简介 (introduction)

LandGPT是中国地质大学（武汉）UrbanComp团队使用8张A6000训练的多模态大语言模型（非任何第三方GPT模型驱动），支持高分辨遥感影像、社交媒体数据和多模态时空数据库作为输入，开展精细地块尺度下的土地的社会经济、多级土地利用分类和人群流量预测等多方面的下游任务。
LandGPT is a multimodal large language model developed by the UrbanComp team at China University of Geosciences (Wuhan) using eight A6000 GPUs for training. Any third-party GPT models do not drive it. LandGPT supports high-resolution remote sensing imagery, social media data, and multimodal spatiotemporal databases as inputs, enabling it to perform a variety of downstream tasks at a fine-grained parcel scale, such as socioeconomic analysis of land, multi-level land use classification, and population flow prediction.

演示视频 (Demo Video)

视频原链接 (Video Link): https://www.bilibili.com/video/BV1CRBDYQEzt

演示是输入高分辨率遥感影像或地理坐标后，获取数据库中的开放地图、手机信令、用户签到、行为轨迹等多模态、多源时空大数据，识别并给出地块的多级土地利用类型。该Demo的模型测试精度（Kappa系数）可达到0.85以上，达到人类专家水平。
The demonstration involves inputting high-resolution remote sensing imagery or geographic coordinates to retrieve multimodal, multi-source spatiotemporal big data from the database, including open maps, mobile signaling, user check-ins, and behavior trajectories. The system then identifies and provides multi-level land use classifications for the specified parcels. The model used in this demo achieves a testing accuracy (Kappa coefficient) of over 0.85, which is comparable to the level of human experts.

训练数据介绍（Training Data）

训练数据来自于我们和阿里巴巴集团基于位置服务（LBS）团队合作获得的CN-MSLU-100K数据集，是一套“以数据为中心”为策略，通过人机协作构建的多源时空数据土地利用分类数据集。
The training data comes from the CN-MSLU-100K dataset, which we developed in collaboration with Alibaba Group's Location-Based Services (LBS) team. This dataset employs a "data-centric" strategy and is constructed through human-machine collaboration to classify land use using multi-source spatiotemporal data.

CN-MSLU-100K共有10万数据量，7个一级类别和28个二级类别的土地利用类型，数据集包括高分遥感地块、GPS位置和兴趣点数据等，是目前最大的地块级土地利用数据集。
The CN-MSLU-100K dataset contains more than 100,000 data points, featuring seven primary categories and 28 secondary categories of land use types. It includes high-resolution remote sensing parcels, GPS locations, and points of interest data, making it the largest parcel-level land use dataset currently available.