Application of Deep Learning in Forest Fire Prediction: A Systematic Review

Forests are among the world's most valuable ecological resources. However, they face significant threats from Forest Fires (FFs), causing environmental damage and impacting wildlife and economies. The increasing global occurrence of FFs has created an urgent need for more accurate prediction methods. Traditional FF prediction approaches, reliant on meteorological data and human expertise, are often limited in accuracy and scalability. Deep Learning (DL) offers a promising solution for enhancing prediction capabilities. This systematic review evaluated various DL techniques for FF prediction, analyzing their methodologies, effectiveness, and challenges. Covering studies published between January 2017 and July 2024, 55 of 656 papers were selected for detailed analysis. The study revealed that Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) models are the most frequently used, with most datasets being publicly available. These findings indicate that classification models and simulation-based studies dominate the field. Commonly used metrics include accuracy, precision, recall, F1 score, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC). Key meteorological features, such as Temperature, Humidity, and Wind speed, have been extensively studied using the Normalized Difference Vegetation Index (NDVI), Land Surface Temperature (LST), and Normalized Difference Moisture Index (NDMI), the most commonly used satellite-derived features. However, integrating human activity data remains underexplored despite its potential to improve prediction accuracy significantly. Addressing this gap could enhance the practical applicability of DL models for FF predictions. This study provides insights into the most prevalent and effective DL techniques for FF prediction and highlights areas for future research.