Probabilistic Wildfire Segmentation Using Supervised Deep Generative Model from Satellite Imagery

Wildfires are one of the major disasters among many and responsible for more than 6 million acres burned in United States alone every year. Accurate, insightful, timely wildfire detection is needed to help authorities mitigate prevent further destruction. Uncertainty quantification always a crucial part natural disasters, such as wildfires, modeling products can be misinterpreted without proper uncertainty quantification. In this study, we propose supervised deep generative machine-learning model that generates stochastic detection, allowing fast comprehensive individual collective events. proposed approach, also aim address patchy discontinuous Moderate Resolution Imaging Spectroradiometer (MODIS) product by training with MODIS raw combined bands detect fire. This approach allows us generate diverse but plausible segmentations represent disagreements regarding delineation boundaries subject matter experts. The segmentation via two streams which learns meaningful latent distributions, other visual features. Two branches join eventually become image-to-image model. compared baseline models: (1) permanent dropout test phases (2) Stochastic ReLU activations. statistical metrics demonstrate better agreements between ground truth segmentations. Furthermore, used multiple scenarios evaluate comprehension, Probabilistic U-Net demonstrates understanding underlying physical dynamics wildfires baselines.