Predicting the Onset Day of California Wildfires Using Deep Learning Methods

Abstract

The increasing frequency and severity of wildfires in California, exacerbated by climate change and human activities, demand advanced predictive tools for effective mitigation. This study employs deep learning (DL) and machine learning (ML) models—convolutional neural networks (CNN), long short‐term memory (LSTM), random forest (RF), and decision trees (DT)—to predict wildfire onset days using a dataset of 14,989 data points (1984–2025) that incorporates historical and projected climate variables such as precipitation, temperature extremes, wind speed, and seasonal patterns. Among the tested models, CNN demonstrated the highest accuracy, achieving a mean absolute error (MAE) of 0.012, mean absolute relative error (MARE) of 0.597%, root mean squared error (RMSE) of 2.432, and an R2 of 0.991. The novelty of this study lies in the customized application of CNN for spatiotemporal wildfire prediction, where climate variables are treated as multi‐channel temporal–spatial tensors, enabling the model to learn both short‐term and long‐term dependencies within structured climate data. This approach goes beyond the conventional use of CNN in image processing by integrating dilated convolutions and optimized kernel architectures to detect rare, high‐impact events like heatwaves and wind bursts that often precede wildfire occurrences. These architectural enhancements allow CNN to extract deep, nonlinear patterns from interdependent climate features while maintaining parameter efficiency and reducing overfitting, marking a significant advancement over standard ML and DL approaches. Despite its strong performance, the model's reliance on projected climate data introduces inherent uncertainties, and the lack of real‐time human activity variables, such as land‐use changes and policy interventions, may limit operational applicability. Future improvements should focus on integrating real‐time sensor networks, refining climate projections, and validating across diverse geographic regions to strengthen the model's reliability and scalability. Ultimately, CNN‐based models have the potential to become crucial tools in proactive wildfire management, enabling timely resource allocation and reducing environmental and human impacts amid a rapidly changing climate.