Using a Machine Learning Method (t-SNE) to Delineate Spatial Zones of Groundwater Geochemistry in Regional Aquifers

Abstract:

It is always a challenge to delineate spatial zones of groundwater geochemistry at the regional scale because of the complex spatial and temporal variation of hydrogeology and groundwater biogeochemistry. Zone delineation requires not only a large amount of long-term groundwater geochemistry data, but also statistical analyses of the data and professional judgement to relate the analyses results to site geology, hydrogeology, and biogeochemical processes. To facilitate the professional judgement, we used the t-distributed stochastic neighbor embedding (t-SNE) method, a state-of-the-art nonlinear dimensionality-reduction technique to visualize similarities and dissimilarities of cluster data. We demonstrated that t-SNE was suitable for delineating zones of groundwater geochemistry using three published geochemical datasets: (1) the Oslo transect dataset with 360 samples and 25 geochemical variables (Templ et al., 2008), (2) the Taiyuan karst water dataset with 37 samples and 31 geochemical variables (Ma et al., 2011), and (3) the Jianghan Plain groundwater dataset with 1,184 samples and 21 geochemical variables (Yang et al., 2020). We developed a protocol of integrating hierarchical cluster analysis (HCA) and t-SNE for creating more accurate hydrogeological zones to advance our understanding of spatial patterns of groundwater geochemistry in regional aquifers. We showed that t-SNE satisfactorily explored large and high-dimensional geochemical datasets and provided a new visualization-based zoning technique to gain advanced understanding of the hydrogeological and biogeochemical processes that control spatial patterns of groundwater geochemistry.