Using machine learning to determine acceptable levels of groundwater consumption in Iran

Abstract

Groundwater footprint index (GFI) is an essential indicator to assess the sustainability of groundwater aquifers. Prediction of future GFI can significantly help managers and decision-makers of groundwater supply to better plan for future resilient consumption of surface and groundwater. In this context, artificial intelligence and machine learning models can aid to predict GFI in view of lacking or uncertain data. We used this technique to predict GFI for 178 Iranian aquifers. To our knowledge, this is the first time that GFI was predicted using machine learning models. Four models, i.e., adaptive neuro-fuzzy inference system, least-squares support vector regression, random forest, and gene expression programming, were used to predict GFI. Systematic combinations of eight variables, including precipitation, recharge, return water, infiltration from the river to the aquifer, groundwater exploitation, aquifer area, evaporation, and river drainage from the aquifer were used in the form of nine input scenarios for GFI prediction. The results showed that inclusion of all input variables gave the best results for predicting the GFI. Predicted GFIs were generally between 0.5 and 8 with an average of 1.9. A value above 1 indicates that groundwater consumption is not resilient that can adversely affect available groundwater resources in the future. Over-use of groundwater can lead to land subsidence. Especially, aquifers located in Qom, Qazvin, Varamin, and Hamedan provinces of Iran may be affected due to large over-use. Among the four models, least-squares support vector regression resulted in the highest prediction performance. Due to the poor performance of adaptive neuro-fuzzy inference system, the novel Harris hawks optimization algorithm was used to improve the performance of adaptive neuro-fuzzy inference system. The Harris hawks optimization - adaptive neuro-fuzzy inference system hybrid model improved the GFI prediction performance. Machine learning methods improve prediction of GFI for aquifers and thus, can be used to better manage groundwater in areas with less reliable data.