Content area
Abstract
Rain is an elegant phenomenon of nature that has aroused human interest and curiosity through the ages. Its complexity has challenged our knowledge and technology and continues to be an important issue of the meteorologists. This thesis is a continuum of essays dealing with our current ability in the forecast of rain by studying the errors, by offering methods for estimating them, by producing improved short-term forecasts, and by measuring forecast skills with an information-theoretic score.
In the first part of our study, errors were obtained in terms of innovations by comparing model forecasts against observations from radars. Bias of each forecast was quantified by comparing the overall distribution of average errors with a normal distribution. Variance of error was obtained from ensemble-based forecasts and innovation-based method, and the results were compared. It was observed that the innovation-based estimation using a local sampling can produce a similar non-uniform variance field of errors as that of the ensembles.
Errors of precipitation forecast are correlated through space and time. These correlations are important for practical purposes such as data assimilation. A non-parametric correlogram was used along with the Wiener-Khinchin theorem to estimate the auto-correlation function of errors (ACF). It was observed that these ACFs form shapes that can be approximated by ellipses and exponential profiles in lag distance units. Thus it became possible to reduce the representation of anisotropic error correlations to three scalar parameters by defining the major axis, the minor axis and the angle of rotation of these ellipses. Furthermore, local error correlation distances were examined and obtained from the variance field, by both ensemble-based and innovation-based estimations. It was observed and discussed that local error correlations of precipitation are better represented by less smooth functions such as low-order autoregressive functions rather than Gaussian.
In the second part, the practical knowledge about error covariance of rain was put to test by designing a predictive error covariance estimation and modeling within a data assimilation scheme. The assimilation method was an adoption of variational schemes modified for the problem of blending two independent sources of forecast in an optimum interpolation. The ability to implement non-homogeneous error correlations within grid-point space demanded an error modeling module that was equipped with an implicit diffusion scheme. The Assimilation Method for Blending Extrapolated Radar (AMBER) was tested on 60 real nowcasting cases and the results indicated average improvements over both forecast skills within 8 hours of lead-time, although this improvement was not always above statistical significance threshold.
Finally, the problem of precipitation forecast verification was addressed by introducing a measure of joint information between binary forecasts and observation with a distinction between favorable and unfavorable mutual information. The information-theoretical score (ITS) was compared with other traditional scores in a hypothetical test scenario and by means of an analytical sensitivity measure. Results indicate that ITS can be used as an alternative skill measure.





