Predicting the Containment Time of California Wildfires Using Machine Learning

California’s wildfire season keeps getting worse over the years, overwhelming the emergency response teams. These fires cause massive destruction to both property and human life. Because of these reasons, there’s a growing need for accurate and practical predictions that can help assist with resources allocation for the Wildfire managers or the response teams. In this research, we built machine learning models to predict the number of days it will require to fully contain a wildfire in California. Here, we addressed an important gap in the current literature. Most prior research has concentrated on wildfire risk or how fires spread, and the few that examine the duration typically predict it in broader categories rather than a continuous measure. This research treats the wildfire duration prediction as a regression task, which allows for more detailed and precise forecasts rather than just the broader categorical predictions used in prior work. We built the models by combining three publicly available datasets from California Department of Forestry and Fire Protection’s Fire and Resource Assessment Program (FRAP). This study compared the performance of baseline ensemble regressor, Random Forest and XGBoost, with a Long Short-Term Memory (LSTM) neural network. The results show that the XGBoost model slightly outperforms the Random Forest model, likely due to its superior handling of static features in the dataset. The LSTM model, on the other hand, performed worse than the ensemble models because the dataset lacked temporal features. Overall, this study shows that, depending on the feature availability, Wildfire managers or Fire management authorities can select the most appropriate model to accurately predict wildfire containment duration and allocate resources effectively.

💡 Research Summary

This research paper presents a machine learning approach to predict the exact number of days required to fully contain wildfires in California, addressing a critical operational need for fire management authorities.

Motivation and Problem Statement: California’s wildfire seasons are becoming longer and more severe, straining emergency response capabilities. Accurately forecasting containment time is essential for effective resource allocation (crews, engines, aircraft), evacuation planning, and public safety alerts. Current practice relies heavily on expert judgment, which can be subjective and inconsistent. While prior research has focused on predicting wildfire risk, spread, or classifying duration into broad categories (e.g., short/medium/long), this study identifies a gap: predicting containment time as a continuous, regression-based measure provides the granular, actionable forecasts needed for operational decision-making.

Methodology and Data: The study integrates three publicly available datasets: historical fire perimeter data from Cal Fire’s FRAP program, its corresponding data dictionary, and spatial shapefiles from the California Open Data Portal. After extensive data cleaning—removing records with missing, corrupted, or inconsistent dates (e.g., containment date earlier than alarm date)—and extracting centroid coordinates (latitude/longitude) for each fire, a final analytical dataset of 15,547 wildfire incidents was created. The target variable, Containment_Days, was derived as the difference between the fire’s alarm date and 100% containment date. Due to a heavy right-skew in the distribution, a log-transformed version (Log_Cont_Days) was used for modeling. A temporal train-test split was employed, with fires before 2018 used for training and fires from 2018 onward reserved for testing/validation, ensuring evaluation reflects generalization to future events.

Model Development and Comparison: Given the dataset’s composition of primarily static features (e.g., cause, location, final fire size in acres—also log-transformed), the study compared the performance of two tree-based ensemble models—Random Forest and XGBoost—against a Long Short-Term Memory (LSTM) neural network. The LSTM was trained on the same static features to evaluate its performance in the absence of explicit temporal sequences (e.g., daily weather variations).

Key Results and Insights:

1. The XGBoost model slightly outperformed the Random Forest model. This is attributed to XGBoost’s gradient boosting framework, which is often more effective at capturing complex, non-linear interactions among static features.
2. The LSTM model underperformed compared to both ensemble models. This result is interpreted as a direct consequence of the dataset’s lack of explicit temporal or sequential features (like daily weather time-series), which are necessary for LSTMs to leverage their strength in modeling time-dependent patterns.
3. The findings highlight a crucial model selection insight: there is no universally “best” model. The optimal choice depends heavily on feature availability. Tree-based ensembles like XGBoost are well-suited for datasets dominated by static features, while sequence models like LSTM become powerful contenders when rich temporal data is available.

Conclusion and Implications: This research successfully demonstrates the application of machine learning for continuous wildfire containment time prediction. It provides empirical evidence that model suitability is context-dependent, guided by the nature of the available data. For fire management authorities, this offers a practical framework: if detailed temporal data (e.g., meteorological time-series) can be integrated, sequence models should be explored; otherwise, advanced tree-based ensembles like XGBoost offer a robust and accurate solution. The study thus contributes a valuable, data-driven tool that can complement expert judgment and enhance strategic resource planning for wildfire containment operations.