Large-scale burn severity mapping in multispectral imagery using deep semantic segmentation models

The raw MTBS data were downloaded from the official website (https://www.mtbs.gov/direct-download), and we select the pre-fire images, post-fire images, and thematic burn severity layers from 2010 to 2019 (over 7000 fires) across the conterminous United States as the data source of Landsat-BSA dataset. The raw MTBS dataset is preprocessed, reformed, and relabeled to produce the Landsat-BSA dataset with a sample size of 256 × 256. In the Landsat-BSA dataset, the seven categories of raw MTBS burn severity, additionally including the class of unburned area outside the fire perimeter, were relabeled into five categories (i.e., unburned, low, moderate, high, and non-processing area/cloud, respectively). The relabeled unburned category combines the raw categories of the unburned area outside the perimeter, unburned to low, and increased greenness. The final Landsat-BSA dataset has 6000 samples, of which 70%, 20%, and 10% are used for training, validation, and testing, respectively, by using a single random shuffling to help alleviate overfitting and improve robustness. The severity level distributions are comparable and independent across training, validation, and testing sets.