Premium
Using a neural network to benchmark a diagnostic parametrization: the Met Office's visibility scheme
Author(s) -
Claxton B. M.
Publication year - 2008
Publication title -
quarterly journal of the royal meteorological society
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.744
H-Index - 143
eISSN - 1477-870X
pISSN - 0035-9009
DOI - 10.1002/qj.309
Subject(s) - parametrization (atmospheric modeling) , visibility , benchmark (surveying) , artificial neural network , computer science , algorithm , artificial intelligence , physics , meteorology , optics , geology , geodesy , radiative transfer
Within the Met Office's Unified Model the visibility is diagnosed from a set of the model's prognostic variables using a parametrization. This parametrization has been optimized to minimize the error in the model visibility, relative to observed visibility. The performance of the parametrization is dependant on two aspects; (1) the quality of the input meteorological variables and (2) the structure of the parametrization itself. This paper describes a technique for obtaining a quantitative assessment of how much improvement is possible in the structure of the visibility parametrization. This is achieved by constructing an alternative visibility diagnostic scheme using a neural network. This statistical model provides a benchmark against which the performance of the current visibility parametrization can be judged, irrespective of the input error. It was found that the neural network achieved significant improvements over the current diagnostic parametrization: a 22% improvement in the geometric mean of the fractional error, and a 15% improvement in the geometric variance of the fractional error. ©Crown Copyright 2008. Reproduced with the permission of Her Majesty's Stationery Office. Published by John Wiley & Sons, Ltd.