Forest Fire Prediction using Earth Observation and AI

The Problem

Forest fires are major natural disasters that are only getting worse as climate change progresses. Sadly the prediction, prevention and alert mechanisms available in most countries are quite old and were developed using fairly small datasets from a time when weather patterns were more stable. Existing systems are simply no longer fit for purpose.

My Plan

Whilst a student in 2017, I put together a 3 person team to take part in the NASA Europa Challenge and the RCUK Smart Cities Challenge representing the University of York. The goal was to use our spare time to develop a user friendly and relatively accurate estimate of forest fire likelihood in a an area over the next 7 days with the highest granularity feasible and then also estimate the direction and speed at which the fire would travel if it were to start in an area.

We judged that if we utilised high resolution historical satellite and forest fire occurrence data along with information such as soil moisture, temperature, precipitation amounts, precipitation types, precipitation volumes, humidity, cloud cover, UV radiation levels amongst other parameters, we could likely develop a good solution.

The main problem was that we were inexperienced, had no funding and no backing to represent the University of York in these challenges. So we had to improvise...

2017 NASA Europa Challenge and RCUK Smart Cities Challenge

Representing the University

I was entering these competitions just as I was completing my second year of university and entering my year in industry (internship) at Amadeus IT. This left my team with 1 month to work on the project full-time and 2 months part-time alongside internship work prior to competition submission deadlines. Most students entering competitions representing the university in the lead-up to internships historically tend gave up a few weeks after entering, causing a fair bit of embarrassment. This made the university hesitant to support us, however we needed to formally represent a university to enter the competitions.

As course representative, I had good relations with most Computer Science department staff. Following rejections from most staff including the head of board of studies, a lecturer agreed to formally back us with the expectation of advice from him but no further support from the university.

Getting the Data

The lack of funding meant that we weren't able to purchase historical weather or forest fire data so we resorted to web scraping this information. Unfortunately the web scraped nature of the data alongside the fact that it had been collected from a wide array of different sources meant that data formats, granularity, units and other details had to be standardised.

Limitations of the Above

This approach meant our machine learning model was only able to reach an accuracy of 72% when estimating forest fire likelihood in the next 7 days in California, accuracy fell a bit further in the rest of the world.

Our system was also quite amateurish, we were reading and writing data from spreadsheets rather than a database. There was also very little multi-threading.

Finals

We were invited to the competition finals in Finland where we came in 4th place in the NASA Europa Challenge, we later came 1st place in the UK Smart Cities Challenge.

2018 UN World Challenge and Microsoft Azure Challenge

We used all the prize money from previous competitions to improve the system in our spare time over the period of a year. By the time we started our final year of university we had obtained higher quality data, experimented with different feature extraction and model training techniques, reduced latency through multi-threading and proper database utilisation and finally, upgraded the UI. Our upgraded system reached a prediction accuracy of 76% and could also estimate the direction a forest fire would spread should one occur.

We entered the upgraded system into the UN World Challenge and a Microsoft competition, leading to first place in both. Our final solution utilised gradient boosted decision trees, implemented through Microsoft's lightgbm.

ESA Phi Conference

Our system caught the eye of the ESA and were were invited to speak at the prestigious ESA Phi Conference where we made valuable contacts.

How Things Ended Up

Our goal was to improve the world's fire fighting infrastructure hence leading to our focus on sales to governments. These proved more beaurocratic than expected leading to us running out of funding. I learnt about the world of venture capital soon after.

https://web.archive.org/web/20200426145405/http://wildfireaware.co.uk/about/