/smoke-detection

Primary LanguageJupyter Notebook

Smoke Detection Using Deep Learning

Successful wildfire intervention is dependent on the early detection of wildfire smoke; mere minutes can drastically alter the extent of societal and ecological damage from wildfires.

When using deep learning for detecting wildfire smoke, past research has largely focused on CNNs. Here, we explore using a combined CNN-LSTM to capture both spatial and temporal information about the smoke.

Getting started

All packages and dependencies are included in the Dockerfile at the root of this repo.

To build the Docker container, run the following from the root of this repository:

docker build -t <desired_name_of_image> .

To run the container:

docker run -it -p 9000:9000 <desired_name_of_image>

The choice of ports (in this case, 9000) is arbitrary, replace it with whatever ports you have open.

To run the container, and be able to save changes made to the Jupyter notebook:

docker run -it -p 9000:9000 -v <local_project_path>/smoke-detection/src:/userdata/kerasData <desired_container_name>

The -v option treats the /src folder as a Docker volume for persistent data storage. Changes made in this folder persist past the life of the Docker container.

Once the container is running, you should see output similar to the following. Open the link your browser to use the interactive Jupyter notebook.

output of running Docker container

Code maintenance

Parent repositories

This repository is forked from i3perez/keras-smoke-detection, which is forked from ar-noc/smoke-detection

Both may be set as upstream remotes for this repository. For example, once established, you can view your remotes via git remote -v

output of git remote -v

Working with Docker

If you've accumulated a bunch of dangling images (denoted by the <none> name and tag) then you can clear them with:

docker system prune

Working with AWS

Set everything up

  1. Set up EC2 G or P instance with at least 75GB of EBS storage.
  2. Create an IAM Role that allows for all S3 policies. Assign the role to the EC2 instance from Step 1.
  3. Create an S3 bucket.
  4. ssh into EC2 instance. Pull down appropriate images from HPWREN.

ssh -i <path to private key you generated> ec2-user@<DNS iPV4 name of instance>

  1. Untar the file.
  2. Copy over images to S3 bucket.

aws s3 cp my_image_folder s3://my_bucket/dest_folder_name --recursive

Once the S3 bucket has been populated

  1. If the EC2 bucket doesn't already have the files from the S3 bucket, cp them over.
  2. Build the Docker container
  3. Run the container (see command above)

Current TODOs

NOTE: This repo is a WIP. Upcoming improvements include:

  • full information about the fork history of the repo
  • instructions on how to fetch updates from various upstream remote repos
  • how to run this on Chameleon
  • how to set up and use the Docker containers given recent improvements
  • any alteration made to the Waggle setup process
  • code restructuring and changes pulled in from ar-noc, which are in local branches but have yet to be pushed
  • better documentation and cleaning up deprecated files