Development of super-resolution based downscaling for wildfire risk

Abstract:

The series of recent wildfire events worldwide highlights climate driven fire risk as one of the urgent threats to humanity under global warming. In particular, successive wildfire in California in the year 2018 and 2019 caused unprecedented damage. In response to such events, detailed regional information about wildfire risk and its prediction is a key to emergency management and planning. Traditionally, dynamical downscaling based on Regional Climate Models (RCMs) and statistical downscaling based on historical relationship between lower and higher resolution datasets have been developed and widely used. However, each has its own drawback: the former still requires considerable amount of computing cost and time, and the latter is limited by amount of data available per target region and linear/stationary assumption used in most of statistical approaches. Furthermore, dynamical seasonal forecasting models contain their own internal bias, requiring additional correction. In this study, we apply various machine learning algorithms that are widely used in the super-resolution image process to produce high-resolution fire risk forecasting information using reanalysis and forecasting model data. We expect not only produce more accurate and efficient performance in higher resolution, but also to omit additional bias correction process for the model data. The spatial resolution is scaled up more than 16 times from 0.94 degree (100km, CFSv2) to 0.04 degree (4km, Prism) over California and Fire Weather Index (FWI) derived from the Canadian forest fire danger rating system is used to represent wildfire risk.