Rain-fed agriculture is extremely important in sub-Saharan Africa, thus the ability to forecast and monitor regional crop yields throughout the growing season would be of enormous benefit to decision makers. Of equal importance to be able to assign a measure of uncertainty to the forecast, especially considering that many predictions are made in the context of a complex climate and sparse meteorological and agricultural observations. This work investigates these issues in the context of an operational updating regional crop yield forecast, concentrating in particular on a case study forecasting Ethiopian maize. Part 1 of the work presented a detailed discussion of Ethiopia' s climate and agricultural systems. As real-time ground based weather observations are sparse in Africa, Part 2 contains an investigation into remotely sensed satellite rainfall estimates. A daily TAMSAT calibration and the geostatistical process of sequential simulation were used to create a spatially correlated ensemble of Meteosat-derived rainfall estimates. The ensemble mean was evaluated as a daily deterministic rainfall product and was found to be as good as or better than other products applied in the same region. A validation of the full ensemble showed that they realistically estimated Ethiopian rainfall fields that agreed both with observed spatial correlations and input pixel level statistics. Part 3 of the work includes a discussion on regional crop simulation modelling and presents a new parameterisation of the GLAM crop simulation model for tropical maize. GLAMMAIZE was then driven using individual members of the satellite ensemble; this was shown to exhibit the correct sensitivities to climate inputs and performed reasonably against yield observations. Finally, Part 4 presented a new method of creating stochastic spatially and temporally correlated rainfall fields. This 'regional weather generator' was tested using a case study on Ethiopian April rainfall and a detailed discussion was included about future development plans.

This thesis addresses the problem of estimating rainfall rates from satellite imagery. The potential for using cloud motion and texture to estimate rain rates has been examined. The main types of textural information, i.e. statistical, structural, frequency and spatio-temporal, have been used to derive features from the satellite measurements and then used to determine a relationship to the radar-observed rain rates. These features were ranked by two scoring functions that were devised to assess their relationship to rain rates. The first scoring function selected a feature set for classifying samples into three rain rate classes. The selected features successfully classify rain rates of a mid-latitude cyclone seen on Meteosat7 with 84.8-99.3 % accuracy with a significant Hanssen-Kuipers discriminant score when a probabilistic neural network was used. A similar accuracy was found when a support vector machine (SVM) was used. Another scoring function was used for the selection of the features for estimating rain rates of each class. A Gaussian-kernel SVM that has been trained by these features produced visually agreeable rain estimates, which were much better than those produced by other methods that used only spectral information. Using the same types features at different time also achieved the similar accuracy. The method was robust and continuous rain estimates were obtained. Unlike other techniques in which additional information has always been required, the results showed that textural information alone can be used for rain estimation. This is preferable when only satellite measurements are available. Frequent updating of the observed rain rates can be done to improve the accuracy of the estimation. The potential for using cloud motion to predict rain rates was also examined. It was found that a combination of the maximum cross correlation and optical flow techniques provided the best estimate of the velocity of clouds. A cloud’s displacement derived by the maximum cross correlation technique was used for the approximation of the future location of its corresponding rain and the final velocity derived by the optical flow technique predicts how the rain rates would change. The rain rates predicted by this novel method provided good correlation to the observed rain rates at an hour later.