Abstract
Making use of the open data from the Central Weather Bureau in Taiwan, this thesis develops a statistical model for multi-step rainfall probability forecasts. The data considered include the rainfall gauge data at the monitoring sites, the satellite cloud image and the radar reflectivity images around the Taiwan area, which are naturally informative to the rainfall tendency. The data are further integrated according to various spatial and temporal resolutions and summarized into different statistic measures. Via a boosting technique, most effective spatial-temporal summaries with predictive abilities (possibly with nonlinear effect) are explored in a logistic regression framework. Accordingly, the spatial forecasting map for rainfall probability can be generated. The proposed methodology is implemented to the hourly data collected from May 19 to May 31 in 2015. The empirical result shows that the proposed prediction model with integrated spatial and temporal variables provides reasonable good multi-step rainfall probability forecasts for 3 hours in advance (with AUC greater than 0.7). In particular, the complexity of selected model is reduced to 20% in total variables and saves about 88% of computational time after introducing the boosting scheme in the modeling procedure, while the reduced model remains a similar forecasting ability evaluated by AUC.