A number of people have been kind enough to ask me about my research, so I thought I'd share a non-technical summary. The challenge with the topic is that it's intersectional, and for the quantitative research aspect, it's very difficult to give an accessibly short answer that doesn't start to talk about convolutional neural networks, knowledge distillation and the properties of edge computing and spectral imaging.
What follows is my pre-research and ethics approval summary designed for a non-specialist audience. It has been included in a scientific poster for use at conferences before results are available.
As an aside, the purple image is the backdrop of a large forest fire in the Pacific North West, colour adjusted to fit UCL branding and used as a backdrop on the scientific poster. The contouring is visual adornment for interest, not an actual cartographic treatment.
Why does wildfire matter?
Climate change has fundamentally altered fire regimes globally, transitioning wildfire from a localised hazard to a systemic global risk.
Emergent next-generation fire-focused satellites now deliver wildfire data at resolutions, revisit intervals, and times of day that were previously impossible. Onboard AI eliminates detection latency by processing wildfire data directly on the satellite, rather than sending it back to Earth.
But machine learning models applied to African savanna fire were built elsewhere. Calibrated on different ecologies, fuels and fire intensities - and built, above all, for suppression. A detection system trained to recognise the spectral signature of Californian or Mediterranean fire does not merely underperform in African contexts, it misinterprets them.
In Sub-Saharan Africa (SSA), fire is not uniformly a hazard. It is an ecological process integral to landscape management and cultural practice evolved over centuries.
Miscalibration doesn’t just produce the wrong data - it produces the wrong governance response. The problem compounds at institutional level. Fire governance across SSA was not designed to integrate AI tools and has historically marginalised the local knowledge most relevant to managing fire effectively.
