Page Banner

United States Department of Agriculture

Agricultural Research Service

Title: A Two-Tier Statistical Forecast Method for Agricultural and Resource Management Simulations

Authors
item Mauget, Steven
item Ko, Jonghan - TEXAS AGRILIFE EXTENSION

Submitted to: Journal of Applied Meteorology and Climatology
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: November 20, 2007
Publication Date: June 1, 2008
Citation: Mauget, S.A., Ko, J. 2008. A two-tier statistical forecast method for agricultural and resource management simulations. Journal of Applied Meteorology and Climatology. 47(6):1573-1589.

Interpretive Summary: The state of the El Niño-Southern Oscillation climate mechanism over the Pacific Ocean can be used to predict winter growing season rainfall over some parts of the U.S. Simple methods to predict growing season climate based on either Pacific sea-surface temperature (SST) or sea-level pressure (SLP) measurements can be used to study the effects of forecast information in agriculture, but may be limited in their accuracy. Here, three advanced statistical methods that use the combined state of Pacific SST and SLP conditions are tested as alternative prediction methods. These methods are: Discrimination Analysis, a Probabilistic Neural Network that uses a genetic algorithm to select smoothing parameters, and Support Vector Machines. Each provides some added skill in retrospective forecasts of winter (November-March) precipitation over a west Texas wheat growing region based on equatorial Pacific SSTA and SLP conditions before planting, compared to more simple schemes used by earlier researchers. While Quadratic Discrimination Analysis provided the best forecast accuracy as measured by the Gandin-Murphy skill score, it is emphasized that the performance of these methods will most likely vary between forecasting applications.

Technical Abstract: Simple conditional methods to predict growing season climate based on the state of leading ENSO sea-surface temperature (SSTA) and sea-level pressure anomaly (SLPA) indicators provide the means to model and study the effects of forecast information in agriculture, but may be limited in their accuracy. Here, three multivariate statistical classifiers as evaluated as alternative forecasting methods. These methods are: Discrimination Analysis, a Probabilistic Neural Network that uses a genetic algorithm to select smoothing parameters, and Support Vector Machines. Each provides some added skill in retrospective forecasts of winter (November-March) precipitation over a west Texas wheat growing region based on previous Spring-Summer equatorial Pacific SSTA and the Southern Oscillation Index conditions, relative to the simple 3-phase schemes used by earlier researchers. While Quadratic Discrimination Analysis provided the best hindcast skill as measured by the Gandin-Murphy skill score, it is emphasized that the performance of these methods will most likely vary between forecasting applications.

Last Modified: 12/25/2014
Footer Content Back to Top of Page