An Approach to Reducing the Dimensionality of a Wildfire Machine Learning Application

A major challenge in a project to produce adaptive, ensemble forecasts of 6 fire weather variables on NOAA’s National Blend of Models (NBM) grid is to find a way to train at over 3 million grid points and 100 lead times for ensembles of size 100. This requires modification of the point-wise adaptive method previously developed by the author.

A first step in this process is to use K-means clustering to reduce the grid to locations where the fire weather climatology is similar. This was achieved by calculating the Hot-Dry-Windy-Index from 2019-2023 ERA5 Reanalysis data for analysis times of 1200, 1800, and 0000 UTC and interpolating those results to the NBM grid. Clustering was accomplished for a range of values of K based on the following variables: the number of times at a grid point that the HDWI ≥ 70 at the same time that RH ≤ 25%, the x- and y-distance of the grid point from a location in the center of the NBM domain, and the grid point elevation. Based on combined consideration of “finding the bend” in total squared error versus number of clusters and subjective geographic similarity, a value of K=100 was adopted. Training can then be done by combining all points within each cluster, thus reducing the dimensionality by almost 10⁴. A second reduction, which also is consistent with the physics, consists of considering each of the 6 variables as a cohort rather than training each variable individually. Even if we consider this cohort as more than a partial training reduction, though, the dimensionality still consists of 100 leads times 100 ensemble members times 100 clusters or 10⁶. Another reduction consists of collapsing the 100 ensemble members by fitting cohort Weibull distributions using a neural network. Open questions to be discussed include how best to produce forecasts for the 100 forecast ranges and the possible application of a “forest-type” ecosystem adaptive approach.