An Efficient Urban Flood Mapping Framework Towards Disaster Response Driven by Weakly Supervised Semantic Segmentation with Decoupled Training Samples

Introduction

This is the official PyTorch implementation of paper: An Efficient Urban Flood Mapping Framework Towards Disaster Response Driven by Weakly Supervised Semantic Segmentation with Decoupled Training Samples, which propopsed a novel weakly supervised semantic segmentation framework and an efficient weak training data generation approach for rapid flood mapping from high-resolution aerial imagery in urban areas.

Weak training data generation

• Weak data from pre-flood river
• Weak data from SAM (Segment Anything Model)

Weakly supervised semantic segmentation framework (TFCSD)

Datasets

The datasets used in this paper can be accessed from [GoogleDirve] or [BaiduDisk].

Dataset	Fully Labeled Data	Flood Patches	Non-Flood Patches	Spatial Resolution(m)	Patch Size	Band	Platform
Calgary-Flood	2862	254	800	0.2	512 × 512	RGB	Airplane
Huston-Flood	2560	305	800	0.3	512 × 512	RGB	Airplane

Calgary-Flood dataset Huston-Flood dataset

Directory Structure

To utilize the code, need to first generate lists of image/label files and place as the structure shown below. Every txt file contains the full absolute path of the files (each file per line). For your convenience, we provide exampled list files of training project for weakly supervised learning in the related datasets [GoogleDirve] or [BaiduDisk]. The image file is the RGB-band image and the label file is the 256-bit image with value from 0 to 255 (0: non-flood pixel, 1:food pixel, 255: unknown pixel).

/root
    /train_image.txt       # the list of training images with absolute path 
    /train_label.txt       # the list of training labels with absolute path
    /test_image.txt        # the list of testing images with absolute path
    /test_label.txt        # the list of testing labels with absolute path
    /val_image.txt         # the list of validation images with absolute path
    /val_label.txt         # the list of validation labels with absolute path

Code

Requirements

• The code is developed using Python 3.7 with PyTorch 1.9.1.
• Anaconda is used to manage the environment and package.

Installation

• git clone https://github.com/YJ-He/Flood_Mapping_WSSS
• cd Flood_Mapping_WSSS
• conda env create -f environment.yaml
• conda activate flood_mapping_wsss
• cd utils/lib_tree_filter then, python setup.py build develop

Training

• set root_dir and hyper-parameters configuration in ./configs/config_WSL.cfg.
• run python train_WSL.py.

Evaludation

• set root_dir and hyper-parameters configuration in ./configs/config_WSL.cfg.
• set pathCkpt in test_WSL.py to indicate the model checkpoint file.
• run python test_WSL.py.

Acknowledgment

• Natural Resources Canada
• City of Calgary
• National Oceanic and Atmospheric Administration (NOAA)
• Megvii Research

Citation

If this repo is useful in your research, please kindly consider citing our papers as follows.

@article{he2024efficient,
  title={An efficient urban flood mapping framework towards disaster response driven by weakly supervised semantic segmentation with decoupled training samples},
  author={He, Yongjun and Wang, Jinfei and Zhang, Ying and Liao, Chunhua},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={207},
  pages={338--358},
  year={2024},
  publisher={Elsevier}
}
@article{he2022enhancement,
  title={Enhancement of Urban Floodwater Mapping From Aerial Imagery With Dense Shadows via Semi-Supervised Learning},
  author={He, Yongjun and Wang, Jinfei and Zhang, Ying and Liao, Chunhua},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  year={2022},
  publisher={IEEE}
}

If our work give you some insights and hints, star me please! Thank you~