Abstract
Operational, real-time rainfall estimation on a daily timescale is potentially of great benefit for hydrological forecasting in African river basins. Sparseness of ground-based observations often means that only methodologies based predominantly on satellite data are feasible. An approach is presented here in which Cold Cloud Duration (CCD) imagery derived from Meteosat thermal infrared imagery is used in conjunction with numerical weather model analysis data as the input to an artificial neural network. Novel features of this approach are the use of principal component analysis to reduce the data requirements for the weather model analyses and the use of a pruning technique to identify redundant input data. The methodology has been tested using 4 yr of daily rain gauge data from Zambia in central Africa. Calibration and validation were carried out using pixel area rainfall estimates derived from daily rain gauge data. When compared with a standard CCD approach using the same dataset, the neural network shows a small but consistent improvement over the standard method. The improvement is greatest for higher rainfalls, which is important for hydological applications.
Corresponding author address: Dr. D. I. F. Grimes, Department of Meteorology, University of Reading, Earley Gate, Reading, Berkshire, RG6 6BB, United Kingdom. Email: d.i.f.grimes@reading.ac.uk