Nationwide Forest Fire Risk Projection in Nepal Using Earth Observation Data and Explainable Machine Learning

This started as a nationwide research project to predict forest fire risk using a meta-model framework built from an ensemble of ten base models:

1. Random Forest (RF)
2. Boosted Regression Tree (BRT)
3. Decision Tree (DT)
4. eXtreme Gradient Boosting (XGBoost)
5. Adaptive Boosting (AdaBoost)
6. Artificial Neural Networks (ANNs)
7. Support Vector Machines (SVMs)
8. classification and regression tree (CART)
9. Adaptive Boosting (AdaBoost)
10. multivariate adaptive regression splines (MARS)

Also, the initial idea was to evaluate forest fire risk across three climate regimes: past (2001–2012), present (2012–2025), and future (2025–2050), by integrating historical data with climate projections based on SSPs and RCPs. That's still the direction, but I've since stepped back to test the methodology itself before committing to it.
Right now, I'm comparing three modeling approaches: a conjoined ensemble, a multi-model framework, and a single explainable model to determine which one actually fits this problem best, rather than defaulting to the most complex option. Once that question is settled, I'll move forward with feeding the prepared data through the chosen framework.

The methodological foundation for this comes from my earlier research work in Rasuwa District, where the core approach was first developed and validated. Combining ensemble or explainable learning techniques with high-resolution climate and land-use data should eventually produce a clearer picture of forest fire risk hotspots and trends across Nepal.
I hope the findings will help policymakers, forest managers, and local communities better understand evolving fire regimes and strengthen their preparedness and adaptive management strategies in the face of climate change.