# A combined statistical and machine learning approach for spatial prediction of extreme wildfire frequencies and sizes

arXiv: https://arxiv.org/abs/2112.14920 · HTML: https://mlsanity.com/paper/2112.14920/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 30 Dec 2021.

## Built on by (most cited first, 5 of 5)

- [Regression modelling of spatiotemporal extreme U.S. wildfires via partially-interpretable neural networks](https://mlsanity.com/paper/2208.07581/) (2022)
- [Deep graphical regression for jointly moderate and extreme Australian wildfires](https://mlsanity.com/paper/2308.14547/) (2023)
- [Insights into the drivers and spatio-temporal trends of extreme Mediterranean wildfires with statistical deep-learning](https://mlsanity.com/paper/2212.01796/) (2022)
- [A Review of Statistical and Machine Learning Approaches for Coral Bleaching Assessment](https://mlsanity.com/paper/2511.12234/) (2025)
- [Neural Architectures for Amortized Bayesian Inference: Statistical Foundations and Empirical Assessments](https://mlsanity.com/paper/2601.07944/) (2026)

## Abstract (arXiv)

Motivated by the Extreme Value Analysis 2021 (EVA 2021) data challenge we propose a method based on statistics and machine learning for the spatial prediction of extreme wildfire frequencies and sizes. This method is tailored to handle large datasets, including missing observations. Our approach relies on a four-stage high-dimensional bivariate sparse spatial model for zero-inflated data, which is developed using stochastic partial differential equations(SPDE). In Stage 1, the observations are categorized in zero/nonzero categories and are modeled using a two-layered hierarchical Bayesian sparse spatial model to estimate the probabilities of these two categories. In Stage 2, before modeling the positive observations using spatially-varying coefficients, smoothed parameter surfaces are obtained from empirical estimates using fixed rank kriging. This approximate Bayesian method inference was employed to avoid the high computational burden of large spatial data modeling using spatially-varying coefficients. In Stage 3, the standardized log-transformed positive observations from the second stage are further modeled using a sparse bivariate spatial Gaussian process. The Gaussian distribution assumption for wildfire counts developed in the third stage is computationally effective but erroneous. Thus in Stage 4, the predicted values are rectified using Random Forests. The posterior inference is drawn for Stages 1 and 3 using Markov chain Monte Carlo (MCMC) sampling. A cross-validation scheme is then created for the artificially generated gaps, and the EVA 2021 prediction scores of the proposed model are compared to those obtained using certain natural competitors.