My secondment adapting a Southern European ML wildfire prediction model for Scotland

When most people think of countries affected by wildfires, Scotland is usually at the bottom of the list. Despite these preconceptions however, wildfire activity in Scotland has increased in recent decades, driven by shifting climate patterns, evolving land‑use practices, and the growing frequency of extreme weather events. This relative lack of wildfire occurrences compared to other warmer European regions is one of the reasons why wildfire research in Scotland has historically been a low-priority area.

In 2025, I undertook a secondment with the National Centre for Atmospheric Science at the University of St Andrews, with my goal being to use machine learning techniques to gain a deeper understanding of wildfire activity in Scotland and provide a method for modelling and predicting wildfire occurrences. With only 3-months to tackle this project, I quickly realised that building an AI model from scratch wouldn’t be feasible. 

Through collaboration with researchers who had designed a wildfire prediction model for Southern European regions, my project quickly shifted to adapting and tailoring this model for Scotland. This involved finding & integrating Scottish environmental data into the model, running it, then evaluating the results to assess its applicability. The results revealed that whilst atmospheric weather variables are usually the most important factor in wildfire occurrences, Scotland’s more temperate climate means that the weather holds much less significance compared to other countries. Instead, physical features like landcover type become a lot more impactful in the model, reflecting both the unique vegetation present in Scotland and the common land management practice - muirburning, which can intensify and spin out of control.

I tailored a machine‑learning framework for Scotland using atmospheric, land‑cover, topographical, and human‑activity datasets spanning a 15 year period to create an AI-ready dataset that provides a great launchpad for analysis with machine learning algorithms such as support vector machines and random forests. Developing these methods not only provides new insights into Scottish wildfires, but it also lays out a roadmap that someone looking to analyse wildfires in their local region could follow in the future.

Considering the challenge of interpretability and trustworthiness in ML and AI, I used SHAP values to quantify the contribution of each predictor to model outputs, which provides a unique insight into the AI ‘black box’. These values quantify the impact different features have on wildfire prediction and are also a mechanism for explainable AI, showcasing the reasoning and weights the model uses when identifying the strongest drivers of wildfire likelihood.

Using the trained model, I created a national‑scale wildfire risk map which displayed spatial patterns of wildfire susceptibility and demonstrated how integrated modelling outputs can support risk‑informed decision‑making for land managers, emergency response planners, and climate‑risk practitioners. The groundwork also provides the ability to predict short-term wildfire likelihood across Scotland in the short-term by inputting forecasted weather variables or outlining future trends and patterns by utilising longer-term climate projections. Highlighting my full process and the model used was vital to ensure transparency, reproducibility, and community reuse.