Predicting Drought Using Pattern Recognition

Abstract

Predicting droughts and their impacts upon overall agricultural production helps in drought management. Generally, statistical regression or time series techniques are employed to predict agricultural droughts quantitatively. Linear (error correction, linear discriminant) and nonlinear (k-Nearest Neighbor) techniques of pattern recognition were used for predicting agricultural droughts qualitatively. A total of five crop districts in the province of Saskatchewan in the Canadian prairies, were selected. Thirty two variables were derived for each district from the daily temperature and precipitation data for the period from 1975 to 2002 to develop pattern recognition models. The variables derived from the minimum or maximum temperatures were found to be more significant than the variables derived from the precipitation for predicting moderate-to-very severe agricultural droughts. The 1975-1997 data were used for model development while the 1998-2002 data were used for model testing. About 83% accuracy was achieved in predicting the non-drought category while 71% accuracy was achieved in predicting the drought category. It was concluded that pattern recognition techniques could be applied for predicting drought qualitatively, which would aid current methods of drought prediction. Key words: Discriminant analysis, crop yield, prairies, wheat.