Effects of temporal baseline choice on wildfire severity mapping with 3 m PlanetScope imagery

Selecting pre-fire imagery for early-spring wildfires involves a trade-off between temporal proximity and seasonal comparability. We evaluated how the choice of temporal baseline affected severity mapping using 3m PlanetScope imagery for the Uljin–Samcheok and Gangneung–Donghae fires in South Korea in March 2022. A common spring 2022 post-fire window was paired with spring 2021 (season-matched), winter 2022 (same-year), and winter 2021 (previous-year off-season) baselines. Maxar imagery viewed in Google Earth Pro supplied reference polygons and unburned, low-, and high-severity labels; nine Composite Burn Index observations corroborated these labels. PlanetScope independently supplied ten spectral-change predictors, processed with min–max scaling. Random Forest, Gradient Tree Boosting, and k-Nearest Neighbors (KNN) were assessed within each fire using a 300 m train–test exclusion buffer and 1000 repeated sampling and partitioning runs per fire–baseline–classifier combination. Pooled mean test accuracy was 0.963 for the season-matched baseline, exceeding same-year winter (0.938) and previous-year off-season (0.928) by 2.57 and 3.53 percentage points, respectively. Season matching yielded the highest mean accuracy in five of six fire–classifier combinations. Gains were greatest for KNN, particularly for low and high severity. Shapley additive explanations (SHAP), computed with PermutationExplainer, identified changes in the Burn Area Index (BAI) as leading predictors in selected Gangneung–Donghae fits. Interannual variation and pair-specific validity limit the attribution of these gains to phenology alone. These within-fire results support the evaluation of season-matched imagery alongside recent pre-fire observations for the intended classifier.