Home

Awesome

Paraformer: Updating Large-scale High-resolution Land Cover Maps from Low-resolution Historical Labels

Large-scale high-resolution (HR) land-cover mapping is a vital task to survey the Earth's surface and resolve many challenges facing humanity. However, it is still a non-trivial task hindered by complex ground details, various landforms, and the scarcity of accurate training labels over a wide-span geographic area. To address these limitations, we propose an efficient, weakly supervised framework (Paraformer), a.k.a Low-to-High Network (L2HNet) v2, to guide large-scale HR land-cover mapping with easy-access historical land-cover data of low resolution (LR).

The Paraformer is accepted by CVPR 2024 :rocket:Highlight:rocket: with a score of 5/5/4. See you in VALSE (Chongqing, 5 May) and CVPR (Seattle, 17 June)!

Contact me at ashelee@whu.edu.cn

Our previous works:

Training Instructions

  1. Download the imagenet21k ViT pre-train model at Pre-train ViT and put it at "./networks/pre-train_model/imagenet21k"

  2. If you want to run the code with the default Chesapeake dataset, we provide example data for the state of New York. Download the dataset at Baidu cloud and put them at "./dataset/Chesapeake_NewYork_dataset".

  3. Run the "Train" command:

    python train.py --dataset Chesapeake --batch_size 10 --max_epochs 100 --savepath *save path of your folder* --gpu 0

  4. After training, run the "Test" command:

    python test.py --dataset Chesapeake --model_path *The path of trained .pth file* --save_path *To save the inferred results* --gpu 0

  1. Generate a train and test lists (.csv) of your dataset (an example is in the "dataset" folder).

  2. Change the label class and colormap in the "utils.py" file.

  3. Add your dataset_config in the "train.py" and "test.py" files.

  4. Run the command above by changing the dataset name.

The Chesapeake Dataset

<img src="https://github.com/LiZhuoHong/Paraformer/blob/main/The%20Chesapeake%20Dataset.png" width="70%"> The Chesapeake Bay dataset, grouped by Microsoft Research, contains 1-meter resolution images and a 30-meter resolution land-cover product as the training data pairs and also contains a 1-meter resolution ground reference for assessment. The figure illustrates the location, Digital Elevation Model (DEM), numbers of the tiles, and data samples of the Chesapeake Bay dataset.

The data can be downloaded at Microsoft's website: Chesapeake dataset

The Poland Dataset

<img src="https://github.com/LiZhuoHong/Paraformer/blob/main/The%20Poland%20dataset.png" width="70%"> The Poland dataset contained 14 Provinces of Poland, including the Provinces of Pomorskie, Lódzkie, Lubuskie, Dolnoslaskie, etc. The figure demonstrates the location, DEM, numbers of the tiles, and data samples of the Poland dataset.

The data can be downloaded at Poland dataset

The SinoLC-1 Dataset

<img src="https://github.com/LiZhuoHong/Paraformer/blob/main/The%20SinoLC-1%20dataset.png" width="70%"> Based on our previous work on SinoLC-1 (i.e., the first 1-m land-cover map of China), we regard the intersected results of three 10-m land-cover products (ESA_GLC10, Esri_GLC10, and FROM_GLC10) as the LR training labels of 1-m Google Earth images. The Paraformer refines a more accurate urban pattern. For the whole of Wuhan City, the reported overall accuracy (OA) of SinoLC-1 is 72.40%. The updated results of the proposed Paraformer reach 74.98% with a 2.58% improvement.

The data can be downloaded at SinoLC-1 dataset

Citation

@article{li2022breaking,
title={Breaking the resolution barrier: A low-to-high network for large-scale high-resolution land-cover mapping using low-resolution labels},
author={Li, Zhuohong and Zhang, Hongyan and Lu, Fangxiao and Xue, Ruoyao and Yang, Guangyi and Zhang, Liangpei},
journal={ISPRS Journal of Photogrammetry and Remote Sensing},
volume={192},
pages={244--267},
year={2022},
publisher={Elsevier}
}

@InProceedings{Li_2024_CVPR,
 author    = {Li, Zhuohong and He, Wei and Li, Jiepan and Lu, Fangxiao and Zhang, Hongyan},
 title     = {Learning without Exact Guidance: Updating Large-scale High-resolution Land Cover Maps from Low-resolution Historical Labels},
 booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
 month     = {June},
 year      = {2024},
 pages     = {27717-27727}
}