Burn Severity Estimation in Mendocino National Forest using BMA Model and Sentinel Images

Wildfires have devastating impacts on forest and its surrounding regions due to the intense heat they produce which further depends on the severity of the fire. Hence understanding the area affected by wildfire and thereof classifying the severity of the fire is an important aspect of monitoring affected regions after fire. For this purpose, we propose a study which involves prediction of burned area, due to wildfires, and based on this classifying burn severity. For validating our approach, we made use of Sentine1-2 satellite data, which provided us with imagery related to area affected by fire. The project’s methodology involves the extraction of relevant features from satellite imagery, weather data and historical fire records but due to its limited and imbalanced nature, oversampling and undersampling techniques are applied to enhance and improve the generalization capability of various machine learning models including CART, ExtraTrees Regressor which are employed to develop the Bayesian Model Averaging (BMA) technique used in this project. The implementation is tested and a comprehensive evaluation is followed and comparison of these models is done based on appropriate performance-metrics. Our model achieves an impressive accuracy of 96.35%. Overall, the findings and results of this project will provide important insights for forest fire management and prediction strategies, furthermore it can serve as a foundation for the development of early warning systems to minimize the devastating effects of fires it can assist authorities in proactive planning, resource allocation and development of mitigation-efforts.