Interpolating monthly precipitation by self-organizing map (SOM) and multilayer perceptron (MLP)

There are needs to find better and more efficient methods to interpolate precipitation data in space and time. Interpolation of precipitation is explored using a self-organizing map (SOM) in a region with large complexity of precipitation mechanisms (northern Iran). The technique is used both for regionalization and for interpolating monthly precipitation for stations with missing data for 1-, 2-, 5- and 10-year periods using a jack-knife procedure to obtain objective results. The SOM is able both to find regions with similar precipitation mechanisms and to interpolate with accuracy. The results show that precipitation interpolation can be improved considerably by taking into account the regionalization properties in the SOM modelling. The SOM results are compared with those from a well-defined multilayer perceptron (MLP). The findings suggest that, without regionalization, MLP modelling is generally better than SOM. However, when regionalization is included, SOM performs better than MLP.