Flood Detector

A flood detector using effientnet as semantic segmentation model.

Dataset

Sen1Floods11: a georeferenced dataset to train and test deep learning flood algorithms for Sentinel-1 (Example). This data was generated by Cloud to Street, a Public Benefit Corporation: https://www.cloudtostreet.info/. For questions about this dataset or code please email support@cloudtostreet.info.

Paper: Bonafilia, D., Tellman, B., Anderson, T., Issenberg, E. 2020. Sen1Floods11: a georeferenced dataset to train and test deep learning flood algorithms for Sentinel-1. The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 210-211.

Training and Testing

Requirements

Python environment

It is highly recomomended to run the training scrips inside of python environment, in case the system does not have any use the following procedure (check Virtualenvwrapper Installation for more details).

$ pip install --user virtualenvwrapper virtualenv

Add three lines to your shell startup file (.bashrc, .profile, etc.) to set the location where the virtual environments should live:

export WORKON_HOME=$HOME/.virtualenvs
export PROJECT_HOME=$HOME/Devel
source /usr/local/bin/virtualenvwrapper.sh

After editing it, reload the startup file (e.g., run source ~/.bashrc) and create a python environment:

$ mkvirtualenv venv_py 
$ workon venv_py

Installing the package

Once the python environment was configured, run the following procedure to install all the required packages

(venv_py)$ git clone https://github.com/ahestevenz/flood-detection-segmentation
(venv_py)$ cd flood-detection-segmentation
(venv_py)$ pip install .

Training

The following script will be used to train the model.

(venv_py)$ bn-run-train --help                                                          
usage: bn-run-train [-h] [-j JSON_FILE] [-v] [-p output.prof]

Welcome to the Bushfires Detector training script

optional arguments:
  -h, --help            show this help message and exit
  -j JSON_FILE, --json_file JSON_FILE
                        JSON configuration (default: "/Users/ahestevenz/Desktop/tech-projects/1_code/flood-detection-segmentation/config.json")
  -v, --verbose         Increase logging output (default: INFO)(can be specified several times)
  -p output.prof, --profile output.prof
                        Run with profiling and store output in given file

Run the above command to start the training (check the data and artefacts paths in config.json first)

(venv_py)$ bn-run-train --json_file config.json   

Testing

To evaluate the model indicated in config.json file, run this command:

(venv_py)$ bn-run-test --json_file config.json 

Results

Qualitative results using the validation dataset

In the plot below, you can observe quantitative results using a similarity metric between ground truth and prediction masks.

TODO List

• Add Docker container option to run/train models
• Add release revision for NASA_IIAC_IEEE_GRSS_ESITC_Flood_Data dataset from the NASA Interagency Implementation and Advanced Concepts Team
• Notebook example