Modelling mean annual rainfall for Zimbabwe
Rainfall has a substantial influence on agriculture, food security, infrastructure development, water quality and the economy. Zimbabwe, like most other Southern African countries, has distinctive meteorological features which are characterized by a high variability of temporal and spatial rainfall distributions, flash floods and prolonged drought periods. Because people struggle to adapt to these diverse rainfall patterns, a better understanding of rainfall characteristics, its distribution and potential predictors will help mitigate the effects of these adverse weather conditions. The aim of this thesis is to develop an early warning tool that can help predict a drought and/or flash flood in Zimbabwe, and to estimate the amount of rainfall during the year. In this thesis, mean annual rainfall figures from 1901 to 2015 obtained from 40 rainfall stations scattered throughout Zimbabwe were used. The thesis consists of three sections. In the first section, appropriate statistical models are applied to a set of annual rainfall figures that have been divided by 12 to produce a mean annual rainfall figure for the year with a view towards finding potential predictors for rainfall in Zimbabwe. Monthly-based indicator variables associated with the Southern Oscillation Index (SOI) and the standardised Darwin sea level pressure readings (SDSLP) were considered as predictor variables with the SOI and SDSLP readings for August (two months before the onset of the rainfall season) producing the most important predictor variables for future rainfall in Zimbabwe. In the second part of the thesis, several characteristics associated with the mean annual rainfall for Zimbabwe are studied using an appropriately fitted theoretical probability distribution. More specifically, the annual rainfall figures from 1901 to 2009 were used to fit a gamma, lognormal and log-logistic distribution to the annual rainfall data. The relative performance of the fitted distributions were then assessed using the following goodness-of-fit tests, namely; the relative root mean square error (RRMSE), relative mean absolute error (RMAE) and the probability plot correlation coefficient (PPCC). All the fitted distributions, however, were not able to adequately predict periods of extreme rainfall. Extreme value distributions such as generalised extreme value and generalised Pareto distributions were then fitted to the mean annual rainfall data. The possibility that periods of extreme rainfall may be time-dependent and be influenced by weather/climate change drivers was then considered. This study shows that, although rainfall extremes for Zimbabwe are not time-dependent, they are highly influenced SDSLP anomalies for April. In the third and last part of this thesis, we categorized rainfall data using a drought threshold value of 570 mm. We compared the relative performance of the logistic regression model in estimating drought probabilities for Zimbabwe with that of a generalised extreme value regression model for binary data. The department of meteorological services in Zimbabwe uses 75% of normal annual rainfall (usually a 30-year time series data) to declare a drought year. Results show that the GEVD regression model with SDSLP anomaly for April is the best performing model and can be used to predict drought probabilities for Zimbabwe.