---
tags:
- mllm
- land-use
- LLM
---
# LandGPT

**This repository contains LandGPT-8B and the corresponding model inference code.**

**If you use this model, please pay attention to our work: LandGPT: A Multimodal Large Language Model for Parcel-Level Land Use Classification with Multi-Source Data**

To reproduce the main contributions of the paper, please download this model and configure the model runtime library according to `pip install lmdeploy>=0.5.3`.

In view of the rapid development of large language model-related technologies, some existing libraries have encountered conflicts. We recommend using 

```
torch                     2.1.0
transformers              4.37.2
flash-attn                2.3.6
lmdeploy                  0.5.3
```
to avoid related dependency conflicts.

In addition, a proper JSONL file is required as an image index to feed into the model. The code for creating this JSONL file has been open-sourced at `https://doi.org/10.6084/m9.figshare.28143191`

Assuming you have set up the runtime environment and created the JSONL file, please run:

```
python true_pre_2level.py
```

This will launch the model inference.

We recommend using a computing card with at least 24GB of VRAM for model inference, as insufficient VRAM may affect the inference quality.

We will introduce a more comprehensive and convenient model reproduction method in the next model release. Please stay tuned to our work! For demonstration videos related to LandGPT, please refer to: https://urbancomp.net/archives/landgpt-demonstration

# LandGPT

**此仓库为LandGPT-8B和相应的模型推理代码**

**如使用此模型，请关注我们的工作：LandGPT: A Multimodal Large Language Model for Parcel-Level Land Use Classification with Multi-Source Data**

如复现文章主要工作，请下载此模型，并根据` pip install lmdeploy>=0.5.3`配置模型运行库。

鉴于当前大语言模型相关技术发展迅速，现有部分库出现了冲突。我们建议使用
```
torch                     2.1.0
transformers              4.37.2
flash-attn                2.3.6
lmdeploy                  0.5.3
```
来避免相关依赖冲突。

此外，还需要合适的jsonl文件作为图像索引送入模型，制作此jsonl文件的代码已经开源在`https://doi.org/10.6084/m9.figshare.28143191`。

假设你已经配置好了运行环境，并创建好了jsonl文件，请运行：

```
python true_pre_2level.py
```

即可启动模型推理。

我们建议使用至少24GB显存的计算卡来进行模型推理，过低的显存可能会影响推理质量。

我们会在下个模型推出更加全面和更加便捷的模型复现方式，请持续关注我们的工作！LandGPT相关的演示视频可参见：https://urbancomp.net/archives/landgpt-demonstration

如果出现通信问题，可以设置`export HF_ENDPOINT=https://hf-mirror.com`来获取镜像模型资源。

欢迎关注我们的工作！下一代模型即将到来~
```
@article{zhu2025landgpt,
  title={LandGPT: a multimodal large language model for parcel-level land use classification with multi-source data},
  author={Zhu, Geyuan and Tang, Mi and Ma, Yueheng and Hu, Zhihui and Yu, Chenglong and Zhang, Xiang and Hu, Huanjun and Guan, Qingfeng and Yao, Yao},
  journal={International Journal of Geographical Information Science},
  pages={1--24},
  year={2025},
  publisher={Taylor \& Francis}
}
```