MESAN Mesoscale analysis of precipitation

The Mesoscale Analysis System (MESAN) has been running operationally since April, 1997, providing science and consumers of weather information with sp atially continuous fields of nine analysed meteorological parameters every hour. Data input to MESAN consists of surface observations from different o bservation systems, numerical weather prediction model fields, weather rada r and satellite imageries, and climate information. Each data source is qua lity controlled before being subjected to an optimal interpolation (OI) sch eme, together with data from the other sources. This paper presents MESAN's accumulated precipitation product. The methods used for interpolation of the multisource data are presented and discussed, as are the methods used to quality control each data source. Results from August-October 1995, using multisource data including gauge observations fr om the countries in the Baltic Sea Experiment (BALTEX) Region, exemplify th e product. OI, used with a variable first guess error, has been compared with conventi onal inverse distance interpolation of precipitation in two catchments in m ountainous terrain. Verification was conducted through modelled runoff, usi ng areally integrated accumulated precipitation, compared with hydrograph o bservations. Significant improvements using OI were found in one of the cat chments. The relative contribution (or importance) of each data source to the analys is has been evaluated using cross validation. Results show that gauge netwo rks are the single most important sources and that radar imagery makes a si gnificant contribution in areas lacking networks of dense gauges, such as t he Baltic Sea. Analysis quality improves with the use of a greater number o f input data sources. MESAN is an appropriate tool for creating an overall best estimate precipit ation analysis and should be useful in applications where such information is required. In validating precipitation produced by numerical weather pred iction models. analyses generated without the use of such model fields is r ecommended.