Optimizing Stacked Ensemble Machine Learning Models for Accurate Wildfire Severity Mapping

Wildfires are increasingly frequent and severe, posing substantial risks to ecosystems, communities, and infrastructure. Accurately mapping wildfire severity (WSM) using remote sensing and machine learning (ML) is critical for evaluating damages, informing recovery efforts, and guiding long-term mitigation strategies. Stacking ensemble ML (SEML) enhances predictive accuracy and robustness by combining multiple diverse models into a single meta-learned predictor. This approach leverages the complementary strengths of individual base learners while reducing variance, ultimately improving model reliability. This study aims to optimize a SEML framework to (1) identify the most effective ML models for use as base learners and meta-learners, and (2) determine the optimal number of base models needed for robust and accurate wildfire severity predictions. The study utilizes six ML models—Random Forests (RF), Support Vector Machines (SVM), k-Nearest Neighbors (KNN), Linear Regression (LR), Adaptive Boosting (AB), and Multilayer Perceptron (MLP)—to construct an SEML. To quantify wildfire impacts, we extracted 118 spectral indices from post-fire Landsat-8 data and incorporated four additional predictors (land cover, elevation, slope, and aspect). A dataset of 911 CBI observations from 18 wildfire events was used for training, and models were validated through cross-validation and bootstrapping to ensure robustness. To address multicollinearity and reduce computational complexity, we applied Linear Discriminant Analysis (LDA) and condensed the dataset into three primary components. Our results indicated that simpler models, notably LR and KNN, performed well as meta-learners, with LR achieving the highest predictive accuracy. Moreover, using only two base learners (RF and SVM) was sufficient to realize optimal SEML performance, with an overall accuracy and precision of 0.661, recall of 0.662, and F1-score of 0.656. These findings demonstrate that SEML can enhance wildfire severity mapping by improving prediction accuracy and supporting more informed resource allocation and management decisions. Future research should explore additional meta-learning approaches and incorporate emerging remote sensing data sources such as hyperspectral and LiDAR.