Satellite Imagery Data Curation Workflow for Wildfire Detection With Advanced Segmentation Modeling

Across the globe, wildfires are occurring at increased frequency, significantly impacting ecosystems and human civilizations. This research paper focuses on the efficacy of two of the most advanced semantic segmentation machine learning models, specifically U-Net based on convolutional neural network and SegFormer based on Vision Transformer network for wildfire detection utilizing Moderate Resolution Imaging Spectroradiometer (MODIS) satellite imagery. The dataset is assembled using a specially built pipeline composed of 1) a workflow to obtain the wildfire candidate location and date, and 2) a subsequent step to collect satellite imagery data utilizing Google Earth Engine image collection and image download service. Experimental evaluation on this dataset shows that both models demonstrate high predictive power of fire at specific geolocations, with ViT outperforming U-Net at the edges of fire regions.