Hydrogeochemical analysis and groundwater pollution source identification based on self-organizing map at a contaminated site

Groundwater contamination at the site has become a very serious problem. A clear understanding of the hydrogeochemical characteristics of groundwater is indispensable for pollution remediation. It requires taking a number of samples and continuous monitoring. However, it is challenging to interpret hydrogeochemical datasets with diverse compositions and wide range of concentration by linear method. In this work, combination of self-organizing map (SOM) and K-means clustering was applied to investigate the hydrogeochemical characteristics at a contaminated site. The results showed that shallow groundwater hydrogeochemical characteristics were performed by 42 neurons and were classified into 5 clusters. The NO3- in cluster 1 widely distributed in the site. The application of fertilizers led to high NO3- concentration in groundwater. Cluster 2 was dominated by Ca2+, Mg2+, Cr(VI) and NO2- and cluster 3 was characterized by TDS, Na+, Cl-, HCO3- and SO42-. Pollutants were mainly from the migration of components at the chromium slag heap under the effect of convection and dispersion. Cluster 4 was dominated by pH, As and CO32-. Furthermore, the pH with the minimum of 8.3 and the presence of CO32- in groundwater provided a favorable opportunity for arsenic enrichment. Pollutants in cluster 4 originated from rainfall leaching on the chromium slag. Moreover, the migration of components from cluster 4 to cluster 2 was also observed by SOM and numerical simulation. Cluster 5 was mainly dominated by Mn and Fe. Reduced environment and anthropogenic activities caused Fe and Mn to exceed standards. The deep groundwater characteristics were performed using 20 neurons and were identified into 4 clusters. Its contamination was due to the leakage of shallow groundwater. Finally, the Gibbs diagram and the saturation index method performed the chemistry control mechanisms of different clusters. This study demonstrated that SOM could be used to interpret nonlinear and complex contamination datasets.