A Deep Learning Approach to Building Real Time Diagnostics of Fire Weather Risk in Southern California

Southern California is prone to extreme fire weather conditions involving high winds, high temperatures and low humidity. Observations from weather stations provide useful measurements of such extreme conditions but the sparse spatial distribution of stations is inadequate, especially over the complex terrain of southern California. To “fill in the gaps” between stations, we present a Deep Learning (DL) model that uses a large set of sparsely distributed weather stations to produce diagnostic high-resolution gridded air temperature, relative humidity, dew point temperature, sustained winds and gusts. The DL model is trained on 10-day data with 10-min frequency. Once the model is trained, 500 m gridded maps are generated. At independent stations not used for training, the DL model is able to capture trends in the ramp up and ramp down of events and are associated with mean absolute, mean bias and RMS errors that fall within NOAA Real-time Mesoscale Analysis (RTMA) standards. The presentation will discuss the implementation of the DL diagnostic model in real time.