Assessment of Spatial Autocorrelation and Scalability in Fine-Scale Wildfire Random Forest Prediction Models

Wildfire prediction models that can be applied across diverse regions at fine scales (< 100 m) are critical for wildfire management. Remote sensing offers a path forward by providing heterogeneous and dynamic measurements of fuel load, type, and flammability. Machine learning methods such as random forests provide an empirical framework that are high-accuracy, computationally efficient, interpretable and able to model complex ecological relationships. Here we use high resolution (70 m, every 3 to 5 days) remote sensing observations of evapotranspiration and evaporative stress index, which represent plant water stress, from Ecosystem Spaceborne Thermal Radiometer on Space Station (ECOSTRESS), as well as topography and weather data, to predict burn severity and occurrence for 8 large wildfires that burned 3,715 km2 from 2021 and 2022 in New Mexico, USA. These fires ranged from low to high burn intensity, and covered a diverse range of ecoregions (deserts, grasslands, forests), plant species, and topographies. We used a single model to predict the burn severity of all wildfires one week before occurrence. The prediction accuracy was greatest when using all predictors (ECOSTRESS, weather, topography) (R2 = 0.77). We assessed the role of spatial autocorrelation in driving model performance by: 1) increasing the sample spacing of our dataset, 2) introducing new predictors that represent spatial structure in the data, and 3) training our model on half the fires and predicting the other half of the fires. We found that after increasing sample spacing, model accuracy declined. However, we found declines in model accuracy were more impacted by decreased training set size compared to the distance spacing – indicating that the models are likely accurately capturing fine-scale processes. Scalability of random forest models was also found to be more challenging for regression problems but was accurate for classification of burned pixel occurrence (total pixel accuracy of 67 %). These results provide promising results for application of random forest models to predict fine-scale fire severity and occurrence with applications for fire management.