[Paper Review] Spain on Fire: A novel wildfire risk assessment model based on image satellite processing and atmospheric information

Helena Liz-López, Javier Huertas‐Tato+4 more|arXiv (Cornell University)|Jun 8, 2023

Fire effects on ecosystems4 citations

TL;DR

This paper proposes the Wildfire Assessment Model (WAM), a deep learning-based regression model that leverages satellite-derived atmospheric variables and the greenness index to predict wildfire resource needs, control/extinction times, and burnt area. Pretrained via self-supervised masked patch prediction on 100,000 unlabeled samples and fine-tuned on 311 labeled wildfires, WAM achieves 9% higher accuracy in aerial resource estimation and 21% improvement in extinction time prediction compared to baselines.

ABSTRACT

Each year, wildfires destroy larger areas of Spain, threatening numerous ecosystems. Humans cause 90% of them (negligence or provoked) and the behaviour of individuals is unpredictable. However, atmospheric and environmental variables affect the spread of wildfires, and they can be analysed by using deep learning. In order to mitigate the damage of these events we proposed the novel Wildfire Assessment Model (WAM). Our aim is to anticipate the economic and ecological impact of a wildfire, assisting managers resource allocation and decision making for dangerous regions in Spain, Castilla y León and Andalucía. The WAM uses a residual-style convolutional network architecture to perform regression over atmospheric variables and the greenness index, computing necessary resources, the control and extinction time, and the expected burnt surface area. It is first pre-trained with self-supervision over 100,000 examples of unlabelled data with a masked patch prediction objective and fine-tuned using 311 samples of wildfires. The pretraining allows the model to understand situations, outclassing baselines with a 1,4%, 3,7% and 9% improvement estimating human, heavy and aerial resources; 21% and 10,2% in expected extinction and control time; and 18,8% in expected burnt area. Using the WAM we provide an example assessment map of Castilla y León, visualizing the expected resources over an entire region.

Motivation & Objective

• To develop a novel wildfire risk assessment model that predicts resource needs, control time, extinction time, and burnt area using satellite and atmospheric data.
• To address the challenge of limited labeled wildfire data by employing self-supervised pretraining on large-scale unlabeled data.
• To improve wildfire management decision-making in high-risk regions like Castilla y León and Andalucía through data-driven forecasting.
• To evaluate the model’s generalization capability across different Spanish regions with varying meteorological conditions.

Proposed method

• The WAM model uses a residual-style convolutional neural network architecture to process georeferenced atmospheric and vegetation variables as image-like inputs.
• It is pre-trained using a self-supervised masked patch prediction objective on 100,000 unlabeled satellite and environmental data samples to learn spatial and temporal patterns.
• The encoder from the pre-trained autoencoder is fine-tuned on 311 labeled wildfire events from Castilla y León to perform regression on key fire management metrics.
• The model fuses atmospheric variables (e.g., temperature, humidity, wind speed) and the greenness index (NDVI) as input features.
• A regression head predicts four key outputs: required human, heavy, and aerial resources, control time, extinction time, and expected burnt area.
• The model generates regional risk assessment maps for Castilla y León, visualizing predicted resource needs across the autonomous region.

Figure 1: Visual representation of WAM methodology steps.

Experimental results

Research questions

• RQ1 Can a self-supervised pretraining strategy on large-scale unlabeled satellite data improve wildfire risk prediction in data-scarce scenarios?
• RQ2 How effective is a residual convolutional network fine-tuned on few labeled wildfires in predicting resource allocation and fire behavior metrics?
• RQ3 To what extent can a wildfire prediction model trained on one Spanish region (Castilla y León) generalize to another region with different meteorological conditions (Andalucía)?
• RQ4 What performance gains does the proposed WAM model achieve over traditional machine learning baselines in predicting wildfire management outcomes?

Key findings

• The WAM model achieved a 9% improvement in predicting aerial resources and a 3.7% improvement in heavy resource estimation compared to baseline models.
• It reduced prediction error by 21% for extinction time and 10.2% for control time, indicating strong performance in temporal forecasting.
• The model showed a 18.8% improvement in estimating the expected burnt area, demonstrating enhanced spatial impact prediction.
• Despite strong performance in Castilla y León, the model failed to generalize to Andalucía without retraining, highlighting limitations in cross-regional transferability.
• Retraining the encoder during fine-tuning significantly improved regression performance, suggesting that initial pretraining alone is insufficient for optimal task-specific learning.
• The use of a residual autoencoder architecture outperformed a sequential architecture due to better gradient flow and deeper feature learning.

Figure 2: Map of Spain showing the locations of all wildfires recorded between 2001 and 2012. Wildfires in Castilla y León are highlighted in blue and those in Andalucía in red.