Machine Learning Historical and Forecasted Gridded Grassland NDVI

Normalized difference vegetation index (NDVI) composite image data has been broadly utilized in identifying high wildfire risk areas over the past decades. In recent years, the need for daily resolution historical and forecasted grassland NDVI at high spatial resolutions within California utility service territories has emerged. To address these needs, the 16-day MODIS Aqua and Terra NDVI remotely-sensed data for 41 California grassland locations were smoothed and matched to dynamically downscaled high-resolution weather predictors using discrete Fourier transformation and linear interpolation techniques. A random forest machine learning model that adopted long term weather predictors including accumulated precipitation and short term weather predictors like daily maximum wind and temperature was built to formalize a gridded grassland NDVI forecast. Finally, our unique forecast framework was used to both generate a 30-year historical NDVI data set and provide an operational 7-day ahead forecast. We will present machine learning model construction and validation process, operational NDVI forecast maps and demonstrate how these analytics have been used to evaluate wildfire risk and to mitigate wildfire threat across California.