Wildfire has become one of the most devastating disasters that not only causes huge loss to human lives and properties, but also emits enormous CO2 into the environment. The 2018 California Camp fire alone has caused $16.5billion loss and emitted a year’s worth of California power pollution!

Currently, there’re over a thousand of earth observation(EO) satellites that are orbiting us, however, only less than 10 of them can monitor wildfire, such as Sentinel and Landsat. These satellites either only track small portions of the land, or require extensive specialties and dedicated preprocessing skills to process, which greatly limit the capability of real-time monitoring of the wildfire across the globe. Therefore, it's important to find a way that may utilize many more of the EO satellite imagery and improve the effectiveness of wildfire monitoring.

With the increasing development of deep learning technologies, convolutional neural network (CNN) has become one of the most powerful tools in image processing. In this work, we trained a U-net based CNN deep learning model on Databricks, it takes raw imagery from different satellites as the input, and is able to quickly detect the wildfire and estimate the area of the burning scar.

What it does

This is a user-friendly application. It takes imagery from different satellites resources as input, and then quickly predicts the forest fire probability and segments the burning scar zones.

In addition, given the image resolution and the forest type, it can calculate the total area of the burnt zone of a wildfire, and estimate the total CO2 emission from this fire.

How I built it

  1. Download the satellite imagery using Google image API to create the training dataset.
  2. Manually contour the burning scar zones as the label.
  3. Build a training pipeline embedded with U-net model.
  4. Train the model on Databricks with 2-instance CPU, and save the trained model to AWS S3.
  5. Create the application using Streamlit framework, and deploy it as a Docker project.

Challenges I ran into

  1. It's challenging to find a data source with good quality because the data format, size, resolution, and origins of the images are very different. Therefore we spent quite some time building the images library and pre-processing the images.
  2. It's hard to define the metric that can properly evaluate the model. We spent a lot of time on training and fine-tuning the model to achieve reasonable accuracy.
  3. When using the application, the satellite imagery used for prediction may be dramatically different in sizes and resolution compared to the one we used while training the model. Therefore the input imagery requires significant pre-processing.
  4. Our team had a critical lack of domain knowledge in satellite imagery, however, through hard work, perseverance, and ingenuity, we acquired the domain knowledge necessary to solve the challenge.

Accomplishments that I'm proud of

  1. Real-time wildfire monitoring is very difficult, our application solves that problem with rapid predictions and a high degree of accuracy.
  2. Existing wildfire monitoring only focuses on a restricted area due to very limited satellite coverage, our application solves that problem by taking various satellite images to reconstruct a much larger monitoring area, greatly expanding the effectiveness.
  3. Our real-time application has a very important attribute: simplicity. We use readily available public resources along with a simple model to facilitate local governments in their efforts to monitor wildfires in real-time.

What I learned

  1. The existing earth satellite monitoring systems are very limited.
  2. Deep learning is a critical tool for image processing.
  3. There are readily available public resources, you just have to look for it.

What's next for Wild Fire Real Time Detection using Satellite Imagery

  1. Connect our application with the public satellite API or website, and create a real-time monitoring system.
  2. Create an API that publishes the wildfire alert to local governments and corresponding organizations based on geographical information.
  3. Create an API that publishes the wildfire information to social media to increase public awareness of natural disasters.

Built With

Share this project: