Wildfire spreading prediction using multimodal data and deep neural network approach.

Area of Science:

• Environmental Science
• Geospatial Analysis
• Artificial Intelligence

Background:

• Wildfire spread prediction is crucial for large countries, but ground monitoring is impractical.
• Remote sensing data offers a viable solution for global wildfire monitoring.
• Existing methods often focus on short-term predictions using drone data.

Purpose of the Study:

• To develop an effective pipeline for large-scale wildfire spread prediction using geospatial data and machine learning.
• To forecast fire spread over a 1-to-5-day horizon.
• To identify significant features influencing wildfire behavior.

Main Methods:

• A neural network model based on the MA-Net architecture was trained.
• Environmental and climate data, including spatial distribution, were utilized.
• Feature importance analysis was conducted to understand contributing factors.

Main Results:

• The model achieved an F1-score of 0.64–0.68 for predicted burned areas over 1–5 days.
• Wind direction and land cover parameters were identified as the most significant features.
• The study was conducted in northern Russian regions.

Conclusions:

• Geospatial data-driven AI approaches can effectively predict large-scale wildfire spread.
• The MA-Net model shows promise for supporting emergency systems and decision-making.
• The approach is adaptable for use in other regions.