Integrating CCD and Adaptive-dNBR With Metaheuristic-Optimized Hybrid Deep Learning for Wildfire Detection and Susceptibility Mapping in Los Angeles County

Wildfires are among the most destructive natural disasters and pose significant risks to ecosystems, infrastructure, and human lives. This study focuses on wildfire susceptibility mapping in Los Angeles County, a region prone to severe wildfires, including the 2025 Palisades and Eaton wildfires. To enhance wildfire inventory mapping, we developed an integrated approach that combines coherence change detection (CCD) from Sentinel-1 SAR data and the adaptive-differenced normalized burn ratio (adaptive-dNBR) from Sentinel-2 optical imagery. This integration improved precision and recall to 0.97 and 0.90 compared with standalone methods (CCD = 0.75 and 0.97, dNBR = 0.78 and 0.98, adaptive-dNBR = 0.96 and 0.91). Using this improved inventory map as input data, we generated wildfire susceptibility maps employing a tree-based machine learning models—XGBoost and LightGBM, and deep learning models—a convolutional neural network (CNN), long short-term memory (LSTM), and hybrid CNN-LSTM—optimized with gray wolf optimization (GWO). The novel approach of metaheuristic-optimized hybrid CNN-LSTM model (CNN-LSTM-GWO) achieved the highest evaluation metrics (AUC = 0.958, Precision = 0.919, F1-score = 0.920, and MCC = 0.807), outperforming the standalone, nonoptimized models and tree-based machine learning. Key wildfire-related factors, such as maximum temperature, land use, solar radiation, humidity, wind speed, and slope, were identified as the most influential drivers of wildfire occurrence based on information gain ratio and Shapley additive eXplanations. Validation against newer wildfires (Bridge and Hudges) demonstrated strong alignment between the predicted high-risk zones and actual fire perimeters, confirming the model’s robustness for future wildfire prediction. This study represents a significant advancement in remote sensing-based wildfire analysis by integrating multisource satellite data with metaheuristic-optimized hybrid deep learning techniques. The proposed framework offers a powerful tool for proactive wildfire risk assessment and mitigation planning in Los Angeles County and other fire-prone regions.