Comparison of Geospatial foundation model based mapped burn scars with predicted forest fire susceptible locations using environmental factors

Abstract

The occurrence of natural disasters in various forms is a widespread phenomenon globally, and in recent years, their frequency and intensity have been on the rise. Forest fires are among the factors that significantly impact ecosystems and ecological balance in natural environments. This paper compares geospatial foundation models with traditional machine learning algorithms for burn scar mapping. Geospatial foundation models, pre-trained on diverse spatial data, are needed to efficiently handle the complexity and volume of satellite imagery. They enhance burn scar mapping by providing faster analysis, and reducing reliance on extensive manual labeling. These models enable scalable and automated processing of new fire events, thus advancing the state of the art in timely and accurate disaster response and ecological monitoring. There is growing interest in the scientific community to explore whether results from geospatial foundation models can be successfully compared with standard mapping and susceptibility methods, which include environmental factors. The study utilizes historical active fire data sourced from MODIS (Moderate Imaging Spectroradiometer) Aqua Terra satellites to model fire proneness using machine learning approach. Environmental parameters using Landsat data and Sentinel data were incorporated into the model. The parameters used for training the model were extracted and resulted in a good ROC for the testing data. The model predicted significant regions falling within the very high severity and moderate severity category. Additionally, the work is verified using planet imagery.