1-Step ahead 5-Day Forecast of Normalized Burn Ratio using a Combination of Sentinel-2 and Machine Learning

Wildfires have become one of the world’s most destructive extreme events. In the US west-coast, for example, wildfire has caused severe loss of property, lives, and vegetation. Timely burn severity estimation is useful for planning and management of after-fire rehabilitation.  This study investigates plausibility of forecasting the normalized burn ratio (NBR), using machine learning models  – recurrent neural network (RNN), long-short term memory (LSTM), and Gated Recurrent Unit (GRU) – and Sentinel-2 imagery, for Campfire, California, U.S. The Campfire is the deadliest and most destructive wildfire in history of California’s wildfires. Sentinel-2 is used in this study because it captures remotely sensed images at high spatial and temporal resolutions of 10 m and 5 days, respectively. The resulting NBR time-series has a 5-hour interval. One-interval look-back in the ML algorithm results in a one-step 5-day prediction. To estimate NBR at the current time step, the machine learning method uses output from previous time-step and input of the current time-step as input variables to the model.  Results of this study show that combining machine learning and Sentinel-2 images produces plausible NBR 1-step ahead 5-day forecasts.