Probabilistic Wildfire Segmentation Using Supervised Deep Generative Model from Satellite Imagery

Abstract

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