Future projection of fire weather and improvement of numerical fire weather forecasting using machine learning

Abstract

The series of recent extreme wildfire events worldwide highlights fire weather as a major concern under global warming. Fire weather, which builds on the warmer and drier weather, can influence the state of fuel to ignite more easily, to spread faster and to burn longer. Given that the ever-increasing threat of massive wildfire, understanding fire weather and its earlier warming contribute a large portion of forest management and the public preparedness to minimize potential fire damage. This thesis explores the trend of fire weather in terms of anthropogenic and natural influences. A set of multi-model large ensemble climate simulations from the Half a degree Additional warming, Prognosis and Projected Impacts (HAPPI) project, paints a picture on how the rising temperature jeopardizes nearly all the world by accelerating climate driven fire hazard. On top of global warming, the intensity and frequency of wildfire are closely related to nature variabilities of the climate system. Particularly, in the western United States, a five-to-seven-year loop of recurring extreme wildfire is identified as an interaction between ocean, atmosphere and vegetation possibly driven by El Niño-Southern Oscillation. Certainly, devastating fires are occurring almost every year in the western United States, and these ongoing issues urge the establishment of a more accurate and efficient fire weather outlook system. This dissertation documents a concept of hybrid fire weather forecasting model, called CFS-SR, constructed by integrating numerical climate model (The Climate Forecasting System version 2) with a machine learning method known as Single Image Super Resolution. The CFS-SR significantly improved accuracy at lead times of up to seven days with enhanced spatial resolution up to 4km. This efficient application presents a potential of using machine learning techniques in the field of fire weather forecasting.