dc.description.abstract | This study examined the level of skill of daily precipitation forecasts of Numerical Weather
Prediction (NWP) models over Kenya. These models are the United Kingdom
Meteorological office (UKMO) model, United States' National Centers for Environmental
Prediction (NCEP) global forecasting system and India's National Center for Medium Range
Weather forecast (NCMRWF) global spectral model. The models' grid box rainfall was
averaged and assigned to an observing station within the box. Verifications were then
carried out against 24-hr observations. The rainfall data were from 22 synoptic stations for a
period of one year, 2004/2005.
Verification scores were derived from two way contingency tables of events and non events.
Thresholds (i.e. 0.1, 0.2, 0.5, 1, 2, 5, 10, 15, 20, 25 mm/day) were considered to define the
transition between a rain versus a non-rain event.
At each verification point, each event/non rain .. event was scored as falling under one of the
four categories of false alarms, misses, hits or correct non-rain forecasts.
The main objective of the study was to perform a comparative verification of skills of three
NWP models currently in use at Kenya Meteorological Department at predicting
precipitation. In order to achieve the objective, categorical statistics based on contingency
tables were derived. These comprised of bias score, probability of detection, false alarm
ratio, and Hansen-Kuipers score. Root mean square errors for checking the accuracy of the
different models were also calculated.
Statistics are presented of the bias score, probability of detection, false alarm ratios and the
Hansen-Kuipers score also known as the 'true skill statistic'. The results showed significant
monthly variations of root mean square values of rainfall. For all the models the root mean
square error was found to be largest during the rainy season of March-May. However the
error was found to be relatively low for the drier months. Bias results indicated that the bias
to be well above one at low thresholds and below one at higher thresholds. This means that
the models overestimate the frequency of rainfall for light rainfall while underestimating the
same for heavy rainfall. The probabilities of detection values were found to be very high
especially for the UKMO model for rainfall up to about athreshold of Smm. The false alarm
ratios scores also varied geographically with higher values in the dry southeastern,
Northeastern and Northwestern parts of the country. Model skill as measured by Hansen-
Kuipers score indicated that UKMO and NCMR WF models had greater skill in forecasting
rainfall than the NCEP model over most parts of the country. The Hansen-Kuipers(HK) score
peaks between the thresholds of2mm -Smm. Thereafter the HK score decreases rapidly. The
results of this study may provide a basic source of information to operational weather
forecasters on the skill of precipitation forecasts from the various models used as input to
their daily operations. The scores may also form a benchmark against which to measure skill
improvement ofNWP models in future.
On the overall, no single model was better th~n the others in all aspects of accuracy and skill.
The UKMO tended to have higher values of bias than the other models. It also overestimated
amounts of rainfall. However it was able to detect presence of rainfall better than the other
models and its Hansen-Kuipers score was higher. On the other hand NCMRWF model
tended to have lower values of bias. Its false alarm ratio scores were also lower. Its Hansen -
Kuipers values were also higher than the other models especially over western Kenya. On the
whole NCEP model seemed to perform poorly in most of the skill scores than the other two
models. However it was giving fewer alarms than the UKMO model. | en |