Spatiotemporal rainfall forecasting models for agricultural management

The main aim of the current PhD thesis is to develop forecast systems for Australia over medium time scales such as weekly, monthly, seasonal and annual for Agricultural planning. Common data driven algorithms in hydrology and climate studies including statistical methods, Artificial Intelligent (AI), machine learning and data mining techniques are sought to improve the rainfall prediction using historical data from land and oceans. First, spatiotemporal monthly rainfall forecasting is developed for south-eastern and eastern Australia using climatic and non-climatic variables. To improve model performance, climate regionalization and regionalization of the climate drivers are considered as initial steps for Neural Network model. The outcome of this study indicates that climate regionalization can improve performance of space-time prediction model for monthly rainfall in eastern and south-eastern Australia. The second part of the study investigates the stability and reliability of the lagged relationship between climate drivers and leading modes of seasonal rainfall in south-eastern Australia. Strength and polarity of correlation between climatic indices and leading mode of seasonal rainfall vary in different seasons and over time. This suggests using suitable lagged climatic indices rather than fixed climatic indices for each season leads to better rainfall predictions. Finally, annual rainfall, using Gene Expression Programming (GEP) method, significant predictors that were identified are Geographic Information System (GIS) variables, long-term mean and median annual rainfall, seasonal rainfall, previous annual rainfall and lagged climatic indices. The results indicate that the best predictors for modelling Australian annual rainfall in space-time are climatology (median and mean of rainfall) in comparison with GIS variables