Enhancing Satellite-Based Wildfire Monitoring: Advanced Contextual Model Using Environmental and Structural Information

Abstract

Satellite-based wildfire detection models provide real-time data that is essential for effective management of wildfires, whose frequency and intensity are increasing. However, conventional temperature anomaly-based contextual algorithms fail to incorporate environmental and structural information, resulting in poor generalization capabilities across various environmental conditions and wildfire intensities. This study developed a novel machine learning-based wildfire detection model incorporating advanced contextual features to fulfill the deficiencies of conventional contextual (CC) algorithms. Specifically, two advanced contextual features-considering the baseline environment (ACE) and structural similarity (ACS)-were introduced, and their effectiveness was assessed relative to CC features. The integration of ACE and ACS significantly enhanced the performance of all the evaluated machine learning algorithms, with the light gradient boosting machine (LGBM) model, considering ACE and ACS as input variables, achieving the highest accuracy (precision: 98.56%; recall: 97.78%; F1-score: 98.17%). Owing to its heightened detection sensitivity, the LGBM model identified additional wildfires that the GEO-KOMPSAT-2A (GK-2A) forest fire (FF) products had failed to detect. The variations in the environmental heterogeneity and wildfire intensity significantly diminished the effectiveness of conventional temperature anomaly-based algorithms. ACE and ACS effectively mitigated performance deterioration attributable to environmental heterogeneity and intensity variations, respectively. This investigation analyzed the potential error sources in conventional satellite-based wildfire detection models and applied a novel methodology to refine them, ultimately proposing a model that can be generalized across varied environmental conditions and wildfire intensities.