Spectral radiative-transfer modeling with minimized computation time by use of a neural-network technique

A new approach based on a neural-network technique for reduction in the computation time of radiative-transfer models is presented. This approach gives high spectral resolution without significant loss of accuracy. A rigorous radiative-transfer model is used to calculate radiation values at a few selected wavelengths, and a neural-network algorithm replenishes them to a complete spectrum with radiation values at a high spectral resolution. This method is used for the UV and visible spectral ranges. The results document the ability of a neural network to learn this specific task. More than 20,000 UV-index values for all kinds of atmosphere are calculated by both the rigorous radiative-transfer model alone and the model in combination with the neural-network algorithm. The agreement between both approaches is generally of the order of ?1%; the computation time is reduced by a factor of more than 20. The new algorithm can be used for all kinds of high-quality radiative-transfer model to speed up computation time.

Method to determine snow albedo values in the ultraviolet for radiative transfer modeling

For many cases modeled and measured UV global irradiances agree to within +/-5% for cloudless conditions, provided that all relevant parameters for describing the atmosphere and the surface are well known. However, for conditions with snow-covered surfaces this agreement is usually not achievable, because on the one hand the regional albedo, which has to be used in a model, is only rarely available and on the other hand UV irradiance alters with different snow cover of the surface by as much as 50%. Therefore a method is given to determine the regional albedo values for conditions with snow cover by use of a parameterization on the basis of snow depth and snow age, routinely monitored by the weather services. An algorithm is evolved by multiple linear regression between the snow data and snow-albedo values in the UV, which are determined from a best fit of modeled and measured UV irradiances for an alpine site in Europe. The resulting regional albedo values in the case of snow are in the 0.18-0.5 range. Since the constants of the regression depend on the area conditions, they have to be adapted if the method is applied for other sites. Using the algorithm for actual cases with different snow conditions improves the accuracy of modeled UV irradiances considerably. Compared with the use of an average, constant snow albedo, the use of actual albedo values, provided by the algorithm, halves the average deviations between measured and modeled UV global irradiances.

Comparison of models used for UV index calculations

Eighteen radiative transfer models in use for calculation of UV index are compared with respect to their results for more that 100 cloud-free atmospheres, which describe present, possible future and extreme conditions. The comparison includes six multiple-scattering spectral models, eight fast spectral models and four empirical models. Averages of the results of the six participating multiple-scattering spectral models are taken as a basis for assessment. The agreement among the multiple-scattering models is within +/- 0.5 UV index values for more than 80% of chosen atmospheric parameters. The fast spectral models have very different agreement, between +/- 1 and up to 12 UV index values. The results of the empirical models agree reasonably well with the reference models but only for the atmospheres for which they have been developed. The data to describe the atmospheric conditions, which are used for the comparison, together with the individual results of all participating models and model descriptions are available on the Internet: http://www.meteo.physik.uni-muenchen.de/++ +strahlung/cost/.