Keywords: drought risk modelling, data mining, vegetation outlook, VegOut, data mining ensemble algorithms, satellite vegetation data, vegetation condition prediction, National Drought Mitigation Center, USA, United States, water shortages
Improving drought risk modelling: using multiple periods of satellite data with ensembles of data mining algorithms
This paper describes improvements to the Vegetation Outlook (VegOut) model developed by the National Drought Mitigation Center for the USA. VegOut provides early drought warning by predicting general vegetation conditions at multiple time steps (e.g., 2–, 4– and 6–week outlooks). VegOut integrates climate, biophysical, oceanic, and satellite–based vegetation data with data mining algorithms to identify historical patterns and to predict future vegetation conditions based on these patterns and current conditions. This study investigated various algorithms and found that bagging ensembles of the linear regression algorithm outperformed other algorithms, including the currently used algorithm in VegOut. This study found that using several weeks of recent satellite vegetation data enhances prediction accuracy, especially for long–range outlooks. These discoveries illustrate that model development must carefully analyse both input data and algorithmic choice. The results of this study will be used to improve VegOut's prediction accuracy.