gistam2023

Repository for the paper Using deep learning and radar backscatter for mapping river water surface

Installation

Clone this repository

git clone https://github.com/galatolofederico/gistam2023.git
cd gistam2023

Create a virtualenv and install the requirements

virtualenv --python=python3.8 env
. ./env/bin/activate
pip install -r requirements.txt

If you are interested in the dataset please contact us

Usage

To build the datasets run

python -m scripts.build-dataset -cn IDAN
python -m scripts.build-dataset -cn leesigma
python -m scripts.build-dataset -cn intensity

To train a model run

python train.py \ 
    -cn <CONFIG> \
    wandb.log=<True/False> \
    wandb.tag=<WANDB_TAG> \
    train.steps=<STEPS> \
    train.save_file="$(pwd)/model-<CONFIG>-<STEPS>.pth"

Where

<CONFIG> can be: IDAN leesigma or itensity
<STEPS> has been set to 30000 in the paper

This script will save the trained model in $(pwd)/model-<CONFIG>-<STEPS>.pth

To run the inference with a model run

python predict-folder.py \
    predict_folder.device=cuda:0 \
    predict_folder.folder=<DATA_FOLDER> \
    predict_folder.river=<RIVER_MASK_TIF> \
    predict_folder.model=<MODEL> \
    predict_folder.output=<OUTPUT_FOLDER>

Scripts

To train all the models from the paper run

./train_all.sh

To run the inference with the trained models run

./predict-all.sh

To compute the performance metrics for all the models run

./evaluate-all.sh

Contributions and license

The code is released as Free Software under the GNU/GPLv3 license. Copying, adapting and republishing it is not only allowed but also encouraged.

For any further question feel free to reach me at federico.galatolo@ing.unipi.it or on Telegram @galatolo