README.md

July 3, 2025 · View on GitHub

SegChange-R1: LLM-Augmented Remote Sensing Change Detection

📄 This is the official implementation of the paper:
SegChange-R1: LLM-Augmented Remote Sensing Change Detection

Fei Zhou

Neusoft Institute Guangdong, China & Airace Technology Co.,Ltd., China

If you like SegChange-R1, please give us a ⭐!

Remote sensing change detection is used in urban planning, terrain analysis, and environmental monitoring by analyzing feature changes in the same area over time. In this paper, we propose a large language model (LLM) augmented inference approach (SegChange-R1), which enhances the detection capability by integrating textual descriptive information and guides the model to focus on relevant change regions, accelerating convergence. We designed a linear attention-based spatial transformation module (BEV) to address modal misalignment by unifying features from different times into a BEV space. Furthermore, we introduce DVCD, a novel dataset for building change detection from UAV viewpoints. Experiments on four widely-used datasets demonstrate significant improvements over existing method.

Baseline

🚀 Updates

✅ [2024.06.01] Open source code
✅ [2025.06.22] Upload to arXiv。

Model Zoo

SOTA

Quick Start

System Requirements

Python 3.12
CUDA + PyTorch
HuggingFace
Stable network connection
High-quality proxy IPs (important)

Installation

1. Create a virtual environment

conda create -n segchange python=3.12 -y
conda activate segchange

2. Install dependencies

pip install -r requirements.txt

3. Configuring HuggingFace Images

vim ~/.bashrc

export HF_ENDPOINT="https://hf-mirror.com"

source ~/.bashrc

Data Preparation

Two dataset structure formats are supported：

1. Default

The structure of the dataset is as follows:

data/
  ├── train/
  │   ├── A/                  # First phase training image
  │   ├── B/                  # Second phase training image
  │   ├── label/              # Training Label (Change Mask)
  │   └── prompts.txt         # The training set text describes the prompt
  ├── val/
  │   ├── A/              
  │   ├── B/             
  │   ├── label/         
  │   └── prompts.txt 
  └── test/
      ├── A/            
      ├── B/             
      ├── label/      
      └── prompts.txt

Change the data_format parameter file configs to default.

2. Custom

The structure of the dataset is as follows:

data/
  ├── A/                      # First phase training image
  │── B/                      # Second phase training image
  │── label/                  # Label (Change Mask)
  │── list                    # List file
  │   ├── train.txt           # A list of training sets
  │   ├── val.txt             # A list of validation sets
  │   └── test.txt            # A list of test sets
  └── prompts.txt             # Text description prompts

Change the data_format parameter file configs to custom.

Training

Generate word embedding files

Use Text Generation and change the 'desc_embs' parameter file configs to 'None' to execute the script.

python ./examples/text_gen.py -c ./configs/config.yaml

If you want to detect changes in multiple categories, you need to manually label the category description text.

Command-line training

python train.py -c ./configs/config.yaml

testing

python test.py -c ./configs/config.yaml

Inference TIF

python infer.py -c ./configs/config.yaml

app demo

cd examples/gradio_app
chmod +x ./run.sh
bash run.sh

@article{zhou2025segchange-r1,
  title={SegChange-R1: LLM-Augmented Remote Sensing Change Detection},
  author={Zhou, Fei},
  journal={arXiv preprint arXiv:2506.17944},
  year={2025},
  eprint={/2506.17944},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}