Groundwater balance estimators using Machine Learning

Groundwater use for irrigation, stock and domestic purposes from shallow unconfined aquifers is rarely metered in most parts of the world despite significant increase in the rate of use in the recent decades. Most aquifers systems are poorly characterized and monitored rendering assessment of groundwater balance and informed management decision making difficult.

Advances in automated data collection through remote sensing and other technologies in the recent years, makes available pertinent data sets that can indirectly inform groundwater balance from which groundwater recharge and discharge can be estimated. Assimilating such data sets using traditional means would require simulating the physics across multiple domains including climate, surface water, unsaturated and saturated zones and may be untenable in most decision-making contexts.

Recent advances in Artificial Intelligence and Machine Learning techniques (AI/ML) have created the opportunity to estimate groundwater balance components probabilistically by considering the correlation and causal relationships with other climatic, hydrological and geospatial variables for which data sets are readily available, for example estimates of actual evapotranspiration from remote sensing data. Such machine learning-based models need not necessarily be underpinned by the explicit solutions of governing equations pertaining to the physical processes involved across multiple domains. These ML-based models can be used either independently or in combination with physically based models for retrospective or predictive assessments of groundwater balance components and quantification of recharge and discharge components including historical pumping rates from an aquifer.

This study develops ML-based groundwater balance estimators using machine learning based on suitable supervised learning algorithm for a selected unconfined aquifer system that spans across 16 districts in northwest Bangladesh region. The study uses daily precipitation data, evapotranspiration estimates using Moderate Resolution Imaging Spectroradiometer (MODIS) data, interpolated river stages, and weekly observed water levels from monitoring bores to train, test and validate a Deep Neural Network model implemented using PYTORCH.

Simulated groundwater levels obtained using the trained and tested ML models are used to estimate long-term groundwater storage changes in region and are compared to estimates from a numerical groundwater MODFLOW model developed and history-matched using Flopy and PEST++ frameworks. Both the ML and MODFLOW models are implemented for a rectangular grid with 1500 m × 1500 m cells.  The workflow scripted using PYTORCH and Flopy libraries enabled the ready comparison of ML and numerical models’ outputs and evaluate the applicability of ML models for groundwater balance simulation.