Forecasting Wildfire Danger Using Deep Learning

• Introduction
• TL; DR
• Getting Started
  • Using Binder
  • Clone this repo
  • Using conda
  • Using Docker
• Running Inference
• Implementation overview
• Documentation
• Acknowledgements

Introduction

The Global ECMWF Fire Forecasting (GEFF) system, implemented in Fortran 90, is based on empirical models conceptualised several decades back. Recent GIS & Machine Learning advances could, theoretically, be used to boost these models' performance or completely replace the current forecasting system. However thorough benchmarking is needed to compare GEFF to Deep Learning based prediction techniques.

The project intends to reproduce the Fire Forecasting capabilities of GEFF using Deep Learning and develop further improvements in accuracy, geography and time scale through inclusion of additional variables or optimisation of model architecture & hyperparameters. Finally, a preliminary fire spread prediction tool is proposed to allow monitoring activities.

TL; DR

This codebase (and this README) is a work-in-progress. The master is a stable release and we aim to address issues and introduce enhancements on a rolling basis. If you encounter a bug, please file an issue. Here are a quick few pointers that just work to get you going with the project:

• Clone & navigate into the repo and create a conda environment using environment.yml on Ubuntu 18.04 and 20.04 only.
• All EDA and Inference notebooks must be run within this environment. Use conda activate wildfire-dl
• Check out the EDA notebooks titled EDA_X_mini_sample.ipynb. We recommend jupyterlab.
• Check out the Inference notebooks for 1 day, 10 day, 14 day and 21 day predictions.
• The notebooks also include code to download a small sample dataset.

Next:

• See Getting Started for how to set up your local environment for training or inference
• For a detailed description of the project codebase, check out the Code_Structure_Overview
• Read the Running Inference section for testing pre-trained models on sample data.
• See Implementation Overview for details on tools & frameworks and how to retrain the model.

The work-in-progress documentation can be viewed online on wildfire-forecasting.readthedocs.io.

Getting Started

Using Binder

While we have included support for launching the repository in , the limited memory offered by Binder means that you might end up with crashed/dead kernels while trying to test the Inference or the Forecast notebooks. At this point, we don't have a workaround for this issue.

Clone this repo

git clone https://github.com/esowc/wildfire-forecasting.git
cd wildfire-forecasting

Once you have cloned and navigated into the repository, you can set up a development environment using either conda or docker. Refer to the relevant instructions below and then skip to the next section on Running Inference

Using conda

To create the environment, run:

conda env create -f environment.yml
conda clean -a
conda activate wildfire-dl

The setup is tested on Ubuntu 18.04, 20.04 and Windows 10 only. On systems with CUDA supported GPU and CUDA drivers set up, the conda environment and the code ensure that GPUs are used by default for training and inference. If there isn't sufficient GPU memory, this will typically lead to Out of Memory Runtime Errors. As a rule of thumb, around 4 GiB GPU memory is needed for inference and around 12 GiB for training.

Using Docker

We include a Dockerfile & docker-compose.yml and provide detailed instructions for setting up your development environment using Docker for training on both CPUs and GPUs. Please head over to the Docker README for more details.

Running Inference

• Examples:
  The Inference_2_1.ipynb, Inference_4_10.ipynb, Inference_4_14.ipynb, Inference_7_21.ipynb notebooks demonstrate the end-to-end procedure of loading data, creating model from saved checkpoint, and getting the predictions for 2 day input, 1 day output; and 4 day input, 10 day output, 4 day input, 14 day output and 7 day input, 21 day output experiments respectively.

• Testing data:
  Ensure the access to fwi-forcings and fwi-reanalysis data. Limited sample data is available at gs://deepgeff-data-v0 (Released for educational purposes only).

• Pre-trained model:
  All previously trained models are listed in pre-trained_models.md with associated metadata. Select and download the desired pre-trained model checkpoint file via gsutil from gs://deepgeff-models-v0, set the $CHECKPOINT_FILE, $FORCINGS_DIR and $REANALYSIS_DIR directory paths through the flags while running testing or inference.

  • Example usage:
    python src/test.py -in-days=2 -out-days=1 -forcings-dir=${FORCINGS_DIR} -reanalysis-dir=${REANALYSIS_DIR} -checkpoint-file='path/to/checkpoint'

Implementation overview

We implement a modified U-Net style Deep Learning architecture using PyTorch 1.6. We use PyTorch Lightning for code organisation and reducing boilerplate. The mammoth size of the total original dataset (~1TB) means we use extensive GPU acceleration in the code using NVIDIA CUDA Toolkit. For a GeForce RTX 2080 with 12GB memory and 40 vCPUs with 110 GB RAM, this translates to a 25x speedup over using only 8 vCPUs with 52GB RAM.

For reading geospatial datasets, we use xarray and netcdf4. The imbalanced-learn library is useful for Undersampling to tackle the high data skew. Code-linting and formatting is done using black and flake8.

• The entry point for training is src/train.py. Input variables used for training the model, by default, as configured in the master branch, are Temperature, Precipitation, Windspeed and Relative Humidity. Support for additional variables Leaf Area Index, Volumetric Soil Water Level 1 and Land Skin Temperature and implemented in the respective branches:

  • For training with input variables t2, tp, wspeed and rh + additionally lai, switch to the lai branch. Note: You will additionally require require the data for precisely these 5 variables in the /data dir to perform the training/inference for this combination of inputs.

  • For training with input variables t2, tp, wspeed and rh + additionally swvl1, switch to the swvl1 branch. Note: You will additionally require require the data for precisely these 5 variables in the /data dir to perform the training/inference for this combination of inputs.

  • For training with input variables t2, tp, wspeed and rh + additionally skt, switch to the skt branch. Note: You will additionally require require the data for precisely these 5 variables in the /data dir to perform the training/inference for this combination of inputs.

  • For training with input variables t2, tp, wspeed and rh + additionally skt as well as swvl1, switch to the skt+swvl1 branch. Note: You will additionally require the data for precisely these 6 variables in the /data dir to perform the training/inference for this combination of inputs.

    • Example Usage: python src/train.py [-in-days 4] [-out-days 1] [-forcings-dir ${FORCINGS_DIR}] [-reanalysis-dir ${REANALYSIS_DIR}]

• Dataset: We train our model on 1 year of global data. The gs://deepgeff-data-v0 dataset demonstrated in the various EDA and Inference notebooks are not intended for use with src/train.py. The scripts will fail if used with those small datasets. If you intend to re-run the training, reach out to us for access to a bigger dataset necessary for the scripts.

• Logging: We use Weights & Biases for logging our training. When running the training script, you can either provide a wandb API key or choose to skip logging altogether. W&B logging is free and lets you monitor your training remotely. You can sign up for an account and then use wandb login from inside the environment to supply the key.

• Visualizing Results: Upon completion of training, the results summary json from wandb can be visualized in terms of Accuracy %, MSE % and MAE % using the plotting module.

  • Example Usage:
    python src/plot.py -f <file> -i <in-days> -o <out-days>

• The entry point for inference is src/test.py. Note: When performing inference for a model trained with an additional variable in any of the branches, ensure access to the respective variables in the /data dir.

  • Example Usage: python src/test.py [-in-days 4] [-out-days 1] [-forcings-dir ${FORCINGS_DIR}] [-reanalysis-dir ${REANALYSIS_DIR}] [-checkpoint-file ${CHECKPOINT_FILE}]

• Configuration Details:

Optional arguments (default values indicated below):

`  -h, --help show this help message and exit`

• The src/ directory contains the architecture implementation.

  • The src/dataloader/ directory contains the implementation specific to the training data.
  • The src/model/ directory contains the model implementation.
  • The src/model/base_model.py script has the common implementation used by every model.
  • The src/config/ directory stores the config files generated via training.

• The data/EDA/ directory contains the Exploratory Data Analysis and Preprocessing required for forcings data demonstrated via Jupyter Notebooks.

  • Forcings: data/EDA/EDA_forcings_mini_sample.ipynb (Resolution: 0.07 deg x 0.07 deg, 10 days)
  • FWI-Reanalysis: data/EDA/EDA_reanalysis_mini_sample.ipynb (Resolution: 0.1 deg x 0.1 deg, 1 day)
  • FWI-Forecast: data/EDA/EDA_forecast_mini_sample.ipynb (Resolution: 0.1 deg x 0.1 deg, 10 days)

• A walk-through of the codebase is in the Code_Structure_Overview.md.

Documentation

We use Sphinx for building our docs and host them on Readthedocs. The latest build of the docs can be accessed online here. In order to build the docs from source, you will need sphinx and sphinx-autoapi. Follow the instructions below:

cd docs
make html

Once the docs get built, you can access them inside docs/build/html/.

Acknowledgements

This project tackles Challenge #26 from Stream 2: Machine Learning and Artificial Intelligence, as part of the ECMWF Summer of Weather Code 2020 Program.

Team: Roshni Biswas, Anurag Saha Roy, Tejasvi S Tomar.