Design Problem

How can I utilize machine learning and data science to predict forest fires?

Design Process

This project started as a simple grid automaton that gradually became more complicated as my programming skills grew. I started with the idea of a grid that could contain forest fire data from a top down view. This is actually a very common fire simulation method that is used worldwide.

There are two main types of fire prediction services: simulations and machine learning. Simulations are good because of control over design process, but are very limited in development, and often can’t model patterns more complex than what is coded into it. To contrast this, ML based approaches are very diverse and adaptable, but present a different issue: data availability. To run a ML model, you need tons of data to train the model.

I started with simulations, first trying grid based modeling and then moving to point based modeling. The development of these can be seen in the legacy portion of my website. Then, I decided to make the jump and attempt a machine learning technique, which requires a lot of data. My design process is below:

1. Research

2. Data sourcing

3. Model set up

4. Train model.

5. Evaluate model

Flowchart of data and training pathway:

To get an in depth dive into my process, please head to this site:

Link to Process Journal and Final Reflection Video

https://sites.google.com/isk.ac.ke/at-forest-fire-simulations/home

The results of the forest fire modeling are also in the process journal in depth. Here are a few predictions:

Here is one fire example after my first set of training. The first row is the actual data, the second is the model prediction, and the third is the overlay of the two.

This is the same fire but after the model has trained for 68 iterations. It can be seen that this is much fuzzier, and that the extent of confidence is lower. This is very interesting, and will be commented on in the reflection video.

Menu