M-BurnScar: Temporal-Based Burnt Scar Profiling for Modelling Risk of Forest Fires in ASEAN Countries

The M-BurnScar platform is an AI-based system designed for the real-time detection of burnt scars and forest fire risk assessment using Sentinel-2 satellite imagery. This platform integrates advanced machine learning models, including Random Forest, LightGBM, K-Nearest Neighbors (KNN), Gaussian Naive Bayes, Support Vector Machines (SVM), Multi- Layer Perceptron (MLP), and Tabular Neural Networks (TabNN), to accurately detect burnt areas and predict fire risks across multiple regions in Southeast Asia. The Random Forest model achieved the highest performance with an accuracy of 97.24%, making it the most effective for burnt scar detection. The system processes large-scale satellite imagery to generate interactive maps that highlight burn severity levels, hotspot locations, and air quality (PM2.5) data. These features enable government agencies, environmental organizations, and local communities to make data-driven decisions regarding fire prevention and management. Additionally, the platform provides APIs for real-time data integration, allowing external systems to efficiently access and analyze the data. Training workshops were conducted with stakeholders to ensure effective usage, and feedback indicated strong support for the platform’s simplicity, real-time capabilities, and potential to enhance fire risk management strategies.