Bar Plot from Scientific Research

Mean absolute SHAP values from an XGBoost classifier (n = 47,742 structures) trained on merged WUI data from five fires. Bars show the average |SHAP| for each predictor, ranked by contributions: structure separation distance (SSD), exterior siding, year built, flame length, ember deposition, roof construction, vent screen, vegetation separation distance (VSD), window pane and eaves. Higher |SHAP| indicates greater contribution to the model’s prediction of structure destruction. The vertical dotted line marks the clustering cutoff (0.5) used to identify redundant features. SSD is the single strongest driver of predicted loss, followed by exterior siding and year built.

DOI