Using Self-Organizing Maps in Creation of an Ocean Forecasting System

Tuesday, 16 December 2014: 3:10 PM
Ivica Vilibic1, Nedjeljka Zagar2, Simone Cosoli3, Vlado Dadic1, Damir Ivankovic1, Blaz Jesenko2, Hrvoje Kalinic1, Hrvoje Mihanovic1, Jadranka Sepic1 and Martina Tudor4, (1)Institute of Oceanography and Fisheries, Split, Croatia, (2)University of Ljubljana, Ljubljana, Slovenia, (3)National Institute of Oceanography and Experimental Geophysics - OGS, Trieste, Italy, (4)Croatian Meteorological and Hydrological Service, Zagreb, Croatia
We present the first results of the NEURAL project (, which is dedicated to creation of an efficient and reliable ocean surface current forecasting system. This system is based on high-frequency (HF) radar measurements, numerical weather prediction (NWP) models and neural network algorithms (Self-Organizing Maps, SOM). Joint mapping of mesoscale ground winds and HF radars in a coastal area points to a high correlation between two sets, indicating that wind forecast may be used as a basis for forecasting ocean surface currents. NEURAL project consists of three modules: (i) the technological module which covers installation of new HF radars in the coastal area of the middle Adriatic, and implementation of data management procedures; (ii) the research module which deals with an assessment of different combinations of input variables (radial vs. Cartesian vectors, original vs. detided vs. filtered series, WRF-ARW vs. Aladin meteorological model), all in order to get the best hindcasted surface currents; and finally (iii) the operational module in which NWP operational products will be used for short-term forecasting of ocean surface currents. Both historical and newly observed HF radar data, as well as reanalysis and operational NWP model runs will be used within the (ii) and (iii) modules of the project. Finally, the observed, hindcasted and forecasted ocean current will be compared to the operational ROMS model outputs to compare skill reliability of the forecasting system based on neural network approach to the skill and reliability of numerical ocean models. We expect the forecasting system based on neural network approach to be more reliable than the one based on numerical ocean model as it is more exclusively based on measurements. Disadvantages of such a system are that it can be applied only in areas where long series surface currents measurements exist and where the recognized patterns can be properly ascribed to a forcing field.