Machine Learning Approach for Next-Day Wildfire Prediction: Challenges, Solutions, andInsights

Next-day wildfire prediction is a critical research problem with significant implications for the environment, society, and economy. This study addresses the challenges associated with accurately predicting fire occurrences and presents a machine learning methodology designed to achieve high sensitivity and specificity in predicting wildfires at a country-wide scale with high spatial granularity. The unique aspects of the problem, including extreme data imbalance, massive scale, heterogeneity, and absence of fire, are thoroughly examined.

The proposed methodology focuses on three key components:

• Feature Set Enhancement: An extended set of fire driving factors, encompassing topography, meteorology, Earth Observation data, and historical fire occurrence information, is utilized. This comprehensive feature set provides a holistic view of the factors influencing fire risk.
• State-of-the-Art Classification Algorithms: A set of well-established classification algorithms, including Random Forest, Extremely Randomized Trees, XGBoost, and shallow Neural Networks, for benchmarking is employed. These algorithms are carefully tuned and optimized to strike a balance between sensitivity and specificity. Furthermore, state-of-the-art Deep Learning Methodologies like Semantic Segmentation and Metric Learning are employed and tuned for this specific task.
• Effective Cross-Validation and Model Selection: Two alternative cross-validation schemes and custom validation measures are introduced to ensure optimal training of classification models. This allows for the selection of diverse models based on the desired trade-off between sensitivity and specificity.

The paper addresses specific challenges, such as extreme data imbalance, massive scale of data, heterogeneity, and absence of fire. The scale of the dataset, with over 830 million instances covering a 500m grid cell resolution for the entire Greek territory, necessitates careful undersampling for model training. Heterogeneity and concept drifts in different months are acknowledged, and the absence of fire instances is discussed in the context of unpredictable factors.

The study explores pitfalls, best practices, and directions for further investigation, providing valuable insights into the complexities of next-day wildfire prediction. The impact of class_weights hyperparameter in compensating for data imbalance is highlighted, emphasizing its significance in cost-sensitive learning.

In conclusion, the proposed machine learning methodology demonstrates effectiveness and efficiency in next-day fire prediction, aligning with real-world fire prediction system requirements. Further, our proposed methods achieve adequately high effectiveness scores (sensitivity > 90%, specificity > 80%) and are realized within a pre-operational environment that is continuously assessed on real-world conditions and also improved based on the feedback of the Greek Fire Service.  The study contributes insights that can guide future research in addressing the challenges associated with wildfire prediction, paving the way for more accurate and reliable models in the field.

Acknowledgement: "This work has been supported by the national research project PREFERRED, which is co-funded by Greece and the European Union through the Regional Operational Programme of Attiki, under the call "Research and Innovation Synergies in the Region of Attica” (Project code: ΑΤΤΡ4-0340489)"