Upscaling of point-scale groundwater recharge measurements using machine learning: A case study in New Zealand and Colombia

Estimating groundwater (GW) recharge rates is essential for water resources decision-making, in particular for dynamic regional-scale allocation. Typically, recharge has been estimated either based on models that require observed historic climatic and soil data for calibration or through measurements at a lysimeter monitoring site. Lysimeters are known as the most direct method of measuring drainage, yet utilization for decision making in regional water management is limited as merely point-scale measures of recharge are provided. In the past, machine learning techniques such as artificial neural networks (ANNs) have been found robust for modelling nonlinear hydrologic processes in relation to groundwater management. For this study, an ANN was selected in order to evaluate whether decision making in groundwater allocation can be improved by upscaling lysimeter-measured recharge. Model uncertainty for the ANN scheme was estimated employing a “Dropout” Monte Carlo (MC) technique. The ANN was trained and assessed in terms of its predictive performance to match lysimeter-measured recharge. The ANN was trained on daily time scale, employing recharge data recorded at three lysimeter stations in the Canterbury plains of New Zealand i.e. Dorie, Dunsandel and Methven sites. The best model in terms of accuracy and parsimony, provided R² values ranging from 0.65 up to 0.86 and a mean absolute error ranging from 0.41 to 0.99 when tested at the three lysimeter locations, with a model uncertainty of 6%. The model was implemented in a geographic information system (GIS) environment, in order to predict the spatial variability of land surface recharge, but also to calculate GW allocation for three of the groundwater allocation zones of the Canterbury Region (i.e. Rakaia-Selwyn, Ashburton and Chertsey). GW available for allocation was estimated to be approximately 650 * 106 m³ year⁻¹ or the Rakaia-Selwyn allocation zone; whereas allocation limits of 284.41 * 106 m³ year⁻¹ and 332.45 * 106 m³ year⁻¹ were estimated for Ashburton and Chertsey respectively. The suitability of applying an ANN to estimate LSR in a comparably data scarce region in Colombia was also tested. The results support how the inclusion of lysimeter data into the analysis, improves our confidence regarding the estimation of groundwater recharge. The methodology developed in this study couples a supervised machine learning technique i.e. ANNs with a visualisation tool in a GIS to predict land surface recharge employing rainfall, potential evapotranspiration and dominant soil texture data as inputs. The tool developed here can be utilised to provide support to water managers in order to identify sustainable dynamic regional groundwater allocation strategies