Statistical Drought Forecasting in South Central Oklahoma

The Southern Great Plains have been hard hit in recent years by reoccurring drought despite the long history of drought in the region. Many communities have been ill-prepared to deal with the consequences of long-lasting drought, which has led to a need to rethink how water is used in order to prevent existing aquifers from running dry. As part of a larger project looking at water sustainability in the Arbuckle-Simpson Aquifer region in south central Oklahoma, this research focuses on trying to forecast the PDSI, as a proxy for drought, in the climate division encompassing the majority of land above the aquifer, OK Climate Division 8. The goal was to forecast a year forward in time the PDSI score in this climate division using only prior monthly PDSI and climate oscillation information as inputs. Techniques from multiple fields were used toward this end: from exploratory data analysis, smoothing was applied to PDSI and the climate oscillations; from the analysis of nonlinear chaotic time series, lags for the inputs were chosen to minimize the mutual information across lags for each of the variables; from machine learning, the data were randomly divided into training and testing sets. Then regression was used on the training set with stepwise forward-backward model selection based on the Bayesian Information Criteria. Finally to evaluate performance, the automatically selected forecast model was tested on data that were not used during training.The training and testing procedure was repeated 500 times to thoroughly assess the model fit and generalizability. Results indicated very strong fit on cross-validation: the median R-squared across the 500 testing data sets for the forecast model was 74.8%. Hence, this forecast modeling procedure had excellent fit, explaining three fourths of the variability in smoothed PDSI using only prior PDSI and climate indicators from more than one year earlier. This includes accurately forecasting rapid regime changes from pluvial to drought and vice versa. This level of forecasting precision would potentially allow water managers and agricultural interests to plan ahead for drought events on a timescale of up to a year, exceeding current forecasting capabilities.