Tools for enhancing the application of self-organizing maps in water resources research and engineering

Environmental measurements generate great volumes of high-dimensional data (often noisy and with missing values) from which meaningful messages may be extracted through appropriate organisation and summarisation. The self-organizing map (SOM) is an artificial neural network popular for recognizing patterns, relationships and clusters in such data. Through the finetuning of configuration and training parameters, the SOM can be tailored to best suit a specific data set. Presently, most hydrologic applications continue to rely on heuristics, software defaults or arbitrarily chosen parameters, forfeiting some of the potential benefits of the method, and often resulting in an output that does not best represent the actual features of the data. This paper guides researchers through the knowledgeable creation and interpretation of an appropriately customised SOM relevant to their particular high-dimensional, nonlinear data set. An understanding of the effects that parameter selection and training options have on the extracted information is developed, practical guidance is given for appropriately modifying parameter sets, and comparisons are made with closely-related methods. Data pre-processing, parameter selection, initialisation, training, visualisation and interpretation of the output map for pattern extraction and clustering are discussed. Through individually customised SOMs, more meaningful information can be gained from data-driven exploratory analysis of the inter-component relationships involved in water resource systems.