Fulfilling the Initial Spread Index: A Physically-Informed Model for Predicting Fire Occurrence Using Latent Cluster Structure and Correlation Dynamics
This project investigates the mechanisms of forest fire occurrence using a physics-aligned machine learning approach. By leveraging physical variables such as ISI, FFMC, DMC, DC, BUI and FWI, we construct a neural network that respects the principles of fire propagation.
The model introduces two latent dimensions—Positive_Cluster and Negative_Cluster—generated through unsupervised KMeans clustering. These clusters represent aggregated fire-spread accelerators and suppressors. Variables are filtered based on their correlation to fire occurrence (|r| > 0.3), and the model is trained on this reduced feature set for better interpretability and physical alignment.
This model achieves perfect classification (1.00 precision/recall) when applied to the UCI Cleaned_Algerian_Forest_Fire_Data.csv dataset, the model maintains high fire recall despite noisier inputs, underscoring the importance of structural physical data in fire modeling.
The presented architecture successfully fulfills and extends the utility of the Initial Spread Index (ISI) by embedding it within a physically interpretable neural framework. The model proves that fire occurrence can be accurately predicted when physical variables and latent propagation structure are respected.
- Scientific relevance: Demonstrates how real-world operational metrics like ISI can be completed and clarified through machine learning.
- Practical value: Offers high recall fire detection in structured environments and resilient early-warning behavior in low-structure data.
- Next steps: Extend modeling to simulate multi-stage propagation and fire spread rate based on geographic and temporal features.
This work sets the stage for physically honest fire modeling using interpretable machine learning.