SilvIA M2: our predictive model, validated as an MVP

Predicting a major wildfire 10 days ahead is no longer theory: it is applied data engineering.

Our SilvIA M2 model has been validated as our predictive MVP. In the visualisation you can see the system assessing historical risk in Spain throughout 2023 and 2024: the green dots represent fires the model was able to anticipate before they happened; the red ones, those that slipped past the radar.

Technical results of the validation

The metrics mark a clear difference from current standards:

• Hyperlocal accuracy: we operate at a resolution of 1 km², compared with the 9 km of the current European standard. We don't predict risk for a whole valley, but for a specific hillside.
• Efficiency in the south: in the southern regions of the peninsula, the model anticipates more than 98% of critical fires (>5 hectares) while flagging just 10% of the territory as high-risk. It is pure optimisation of prevention resources.
• Domain knowledge: in areas such as Asturias (north), where most fires are due to scheduled pasture burns rather than purely climatic factors, the model keeps a hit rate above 53%, showing that our separation of regional architectures works.

We achieved these metrics without burning tens of thousands of euros on brute-force computing, relying on tree-based models (ExtraTrees), smart sampling and regional data routing.

Climate-risk management needs less reaction and more predictive anticipation.

Great work by Rafael Dias Ribeiro de Almeida and Omar Bouaziz, who have achieved an enormous step forward in our development.