This repository hosts the code for the SFMT-UDA model which is a novel two-stage Source-free Multi-target Unsupervised Domain Adaptation framework for cross-city LCZ classification.

• python 3.9
• pytorch 2.3.1+cu121
• torchvision 0.18.1
• others:opencv-python, tqdm, scipy, sklearn, matplotlib, seaborn

• Prepare dataset

  • Download VHRLCZ. Samples in VHRLCZ are organized in the ImageNet format (The data is stored in a root directory, with each category having a separate folder).
  • We generate some .txt files containing paths and labels using /SFMTDA/Stage1/Generate_list/readfile_LCZ.py.
    1. Replace the folder path in readfile_LCZ.py with the path where you placed the dataset.
    2. then run the script:

    python readfile_LCZ.py

    3. Copy the generated .txt files to /SFMTDA/Stage1/code/data/{dataset_name}.

• For Stage 1

  1. Change the current directory to SFMTDA/Stage1/code/uda.
  2. Run the main script using:

  sh run.sh

• For Stage 2,

  • Prepare pseudo labels for Stage 2

    1. Copy SFMTDA/Stage1/code/uda/ckps/target/uda/{dataset_name} to the /SFMTDA/Stage2/STDA_weights/STDA folder
    2. Change the current directory to /SFMTDA/Stage2
    3. Run the script:

    sh ./brige.sh

    4. The generated .csv is in ./csv_pseudo_labels

  • Run the main script using:

    sh ./MH_MTDA.sh

• LCZ Mapping

  • Run the script:

    sh ./mapping.sh

The VHRLCZ dataset consists of two components: Jilin-1 satellite imagery and Google Earth imagery. It is primarily designed for fine-grained LCZ classification and cross-city domain adaptation research. The dataset can be accessed via Google Drive files.

1. Google Earth Imagery: The Google Earth imagery in VHRLCZ is sourced from the LCZC-GES2 dataset (refer to our previous work). We selected Google Earth RGB images from LCZC-GES2 that cover seven cities in southeastern China: Guangzhou, Hefei, Hong Kong, Nanchang, Nanjing, Shanghai, and Wuhan. In total, there are 18,936 image patches, each measuring 320 × 320 pixels with a spatial resolution of 1 meter.

2. Jilin-1 Imagery: We acquired Jilin-1 scene-framed products from Chang Guang Satellite Technology Co., Ltd. These images have a spatial resolution of 0.5 meters and cover areas in two Chinese cities: Guangzhou and Changsha. The imagery was manually labeled using digital polygons, following the detailed decision-rule workflow proposed by Kim et al. (2021) and our newly developed workflow.

   Our workflow is:
   

   Based on these labels, we extracted 5,667 RGB image patches, each measuring 640 × 640 pixels.

• [2025/03/07]: Uploaded a batch of code, including the implementation of SFMT-UDA.
• [2025/07/03]: Uploaded some checkpoints. Please refer to link.

If you use this code in your research, please consider citing the following paper:

@ARTICLE{11185121,
  author={Wu, Qianqian and Liu, Yinhe and Zhong, Yanfei and Lin, Kexin and Ma, Xianping and Pun, Man-On},
  journal={IEEE Transactions on Geoscience and Remote Sensing}, 
  title={Source-free Multi-target Unsupervised Domain Adaptation for Cross-City Local Climate Zone Classification}, 
  year={2025},
  volume={},
  number={},
  pages={1-1},
  keywords={Adaptation models;Urban areas;Data models;Remote sensing;Meteorology;Training;Labeling;Supervised learning;Feature extraction;Land surface;Local Climate Zone classification;multi-target domain adaptation;source-free domain adaptation;Very High Resolution imagery},
  doi={10.1109/TGRS.2025.3615979}}

This project is inspired by and builds upon the work from:

• SHOT-plus
• CoNMix

We sincerely appreciate their contributions to this research area.

For any inquiries or further information, please contact me.

There are some other works in our group:

SSRS