Random Foresters| Spot That Fire V3.0
Spot That Fire V3.0
Engine for Fast and Scalable Fire Prediction (EFSFP)
Summary
Our project consists of a lightweight implementation of fire data analysis that can track points most susceptible to lead to larger fires, thus enabling a more agile response by authorities.
How We Addressed This Challenge
The problem
Fire detection software cannot afford to do one thing: waste time. A few extra hours of data processing with an overly complicated algorithm can mean the loss of thousands of human and animal lives, besides often irreparable damages to the environment. Many systems fall prey to the urge to use bleeding-edge algorithms: Recurrent Convolutional Neural Networks, Attention Models and so on. While these are great for certain applications, they are simply too heavy to run in real time. That is why we have aimed to develop the most lightweight possible approach to fire detection and prediction.
Our solution
Our model takes as input the Near-Real-Time (NRT) data published by NASA via FIRMS and runs a fast clustering algorithm on it to group fires across the globe. Then, it compares the current data to the historical series and looks for fire spots that are historically correlated to larger disasters. Through a simple modelling of how fires usually spread from those points, we can then alert the relevant authorities of the danger, so that they can contain the flames before they destroy more lives.
Our model is so lightweight it can be run on any smartphone. This enables it to work not as a substitute for the more sophisticated (and also more expensive) models out there, but instead as a complement to them. Our approach can detect dangerous fires fast, and thus can act as a starting point for investigation via more complex methods. Having this preliminary layer of analysis can accelerate goverment response times by a lot (so much so that each inndividual can run the algorithm on their own machines, independently).
The simplicity of our implementation also means it is especially scalable and easy to build upon. It can, thus, be incorporated into larger systems by independent developers, and act as a building block to other, more robust systems both in the private and public sectors. The algorithm amounts to a few lines of code, and can be debugged and expanded much faster than advanced Deep Learning methods would permit.
How We Developed This Project
Our implementation
Our approach was inspired by Artés, T., Oom, D., de Rigo, D. et al. A global wildfire dataset for the analysis of fire regimes and fire behaviour. Sci Data 6, 296 (2019). Like them, we applied the DBSCAN clustering algorithm to the fire spot data, but added a further simplifying step. We turned a 2D, temporal series clustering problem into a 3D clustering problem by treating the time dimension as a third spatial dimension. With this approach, fires become a point cloud in 3D space, which allowed for much faster and simpler processing of the data, and produced clusters that reflect the time evolution of fires.
We used Python 3.8 to develop our solutions, and employed the sklearn implementation of the DBSCAN Algorithm.
How We Used Space Agency Data in This Project
We downloaded the data from https://nrt3.modaps.eosdis.nasa.gov/archive/FIRMS/c6/Global using a MAP KEY. The data contains fire spots detected by NASA Satellites, organized in CSV format. We put that data into tensors, and then ran clustering algorithms on them to find fire clusters, and then predict the most critical fire spots based on the historical series.
Project Demo
https://docs.google.com/presentation/d/15XQynTcG86oMsrwRdaGZCFIbYMI62TDTmGb6q4y02fk/edit?usp=sharing
Project Code
Data & Resources
Artés, T., Oom, D., de Rigo, D. et al. A global wildfire dataset for the analysis of fire regimes and fire behaviour. Sci Data 6, 296 (2019).
Scikit-learn: Machine Learning in Python, Pedregosa et al., JMLR 12, pp. 2825-2830, 2011.