Delineation of precipitation areas from MODIS visible and infrared imagery with artificial neural networks

Abstract

An important phase in a nowcasting system is the diagnosis of the forecast variables. This work focuses on the diagnosis of precipitation. The Nimrod automatic nowcasting system at the Met Office (UK) has long used Meteosat visible and infrared data to supplement the data it receives from the UK weather radar network to produce rainfall analyses. With the advent of Meteosat Second Generation (MSG) attention has focused on how best to use the larger range of spectral information from MSG to improve the rainfall analyses. Earlier work at the Met Office had suggested artificial neural networks (ANNs) to be a useful tool for such applications. Pending the availability of data from MSG, ANNs were used to process data from appropriate visible and infrared channels on the MODIS instrument. Sixty daytime winter cases were collected, and Nimrod radar rainfall analyses provided ‘ground truth’ for both training and testing the ANNs. The optimal combination of MODIS channels was investigated and it was found that almost all the skill in rain/no rain discrimination was provided by the radiance values from six selected spectral channels. A notable result was that the 1.64 μm channel had no value as a discriminator when used alone, but produced a large increase in skill when used in conjunction with a visible channel. The ANN with MODIS data was found to outperform the corresponding Nimrod look-up table technique applied to Meteosat data. Application of the technique to SEVIRI data is proposed, as is extension to other seasons.