Mapping wildfire dynamics: GeoAI-driven comparative analysis of deep and machine learning ensembles for susceptibility prediction in California

Wildfire susceptibility mapping is a critical undertaking for prevention and management activities, particularly in regions like California, which faces increasing frequency and intensity of fires. Advanced GeoAI, Machine Learning (ML), and Deep Learning (DL) algorithms are indispensable tools in this field. While ensemble methods can enhance prediction accuracy by reducing variance and mitigating bias, few studies systematically compare ML and DL ensembles while also evaluating computational trade-offs and spatial realism, which are crucial for practical implementation. This research addresses these gaps by developing and comparing two GeoAI-driven frameworks in California: (1) an ML ensemble integrating Decision Tree (DT), Random Forest (RF), XGBoost, and LightGBM; and (2) a DL ensemble combining DNN, ANN, H20.ai, ResNet, and LSTM architectures. The models were trained and tested using wildfire occurrence data (2016–2024) and eight key environmental factors (topographic, climatic, and vegetation-related). To ensure spatial realism and avoid leakage, we evaluate all models with 5-fold spatial block cross-validation (K-means regions + GroupKFold) in addition to a conventional random split. Among single ML models under spatial CV, Random Forest and XGBoost provide the strongest discrimination (ROC-AUC ≈ 0.84 and 0.83, Accuracy ≈ 0.74), while LightGBM is competitive but modestly lower (ROC-AUC 0.82, Accuracy 0.73). A fold-safe stacking ensemble that combines DT, RF, XGBoost, SVM (RBF), and LightGBM attains the highest accuracy (≈ 0.75) with a ROC-AUC ≈ 0.84, indicating a small gain in overall correctness without a clear improvement in rank-ordering over the best single models. In contrast, the DL ensemble was highly resource-intensive (102.5 min training time, 6.1 GB memory) and achieved lower accuracy (77.3 %, R²: 0.39), failing to outperform its individual component DL models. These findings demonstrate that tree-based ensembles, such as RF, XGBoost, and the ML ensemble, currently outperform DL ensembles in terms of accuracy, efficiency, and stability for this domain. We recommend RF or XGBoost for operational wildfire susceptibility mapping using a single model and ML-ensemble using stacking models, as it produces the most reliable and spatially coherent assessments. • Developed and compared GeoAI-driven ML and DL ensemble frameworks for wildfire prediction in California. • LightGBM achieved the highest performance (Accuracy 87.6 %, ROC-AUC 0.94) with high computational efficiency. • ML ensembles provided robust accuracy (86.6 %) and realistic spatial patterns, outperforming DL ensembles. • DL ensembles were resource-intensive (6.1 GB, 102.5 min) without surpassing tree-based methods. • Tree-based ensembles offered the best balance of accuracy, efficiency, and stability for operational wildfire mapping.