Towards modelling ultraviolet index in global scale. Artificial neural networks approach

A unique method for the reconstruction of global maps of erythemal Ultraviolet index (UV index) is proposed. The feed forward, multilayered supervised artificial neural network dedicated to this task was constructed and trained using purely experimental meteorological parameters (covariates: date, latitude, UV index for previous day, two days earlier and one year earlier and response variable: erythemal local noon UV irradiance expressed as UV index) for all dates from a 3-year range period (2001–2003) collected by Total Ozone Mapping Spectrometer, TOMS. The data from the 3-year period provide 1095 grids of 288 × 180 , i.e. 56,764,800 training vectors. The 4-year period of 2001–2004 was used for the forecast validation. Global UV index maps for any location and date can be predicted with accuracy comparable to the detection error (5%, 0.7 unit of UV index). The prediction is better than that obtained for the artificial neural network using ozone levels, aerosols and reflectivity as inputs or that obtained for the network using ozone levels as inputs (6%). The accuracy of prediction is much higher for medium and extremely high UV index in comparison with that for low UV index. The novelty of our approach relies also on using only archival data of UV index and not taking any additional meteorological or environmental data as predictors.