Machine learning models for predicting spatiotemporal dynamics of groundwater recharge

Abstract

Although groundwater is one of the most important sectoral exposures to climate change [1], a vital component of maintaining the world's food supply, the primary source of fresh water, an essential component of preserving the ecological balance of the planet, and a necessary component of the earth's crust that prevents the earth from burning, it is hardly ever fully utilized. Despite its significance and limited availability, groundwater management is rarely done in most countries of the world [2]. Continuous monitoring and precise projections of spatiotemporal groundwater recharge change can aid in sustainable development and effective groundwater resource management. Open public remote sensing datasets were used to develop machine learning prediction models (Random Forest, XGBoost, Keras models, etc.) and time series forecasting models (LSTM, CNN, etc.) for predicting and forecasting groundwater sheet recharge, respectively. The publicly available datasets are merged, processed, and organized into three parts for training, testing, and validation: 2002–2009, 2010–2015, and 2015–2020. A comparison of spatiotemporal prediction models' estimates of groundwater recharge in Morocco revealed AdaBoost and RF were the more accurate methods for temporal and spatial prediction, with RMSE values of 10.9712 mm/month and 5.0089 mm/month, respectively. In terms of time series forecasting, the LSTM model performed better, with an RMSE of 20.05 mm/month. Our models' performance on validation datasets demonstrates the utility and scalability of our combined remote sensing and artificial intelligence-based technology, opening up a new pathway for large-scale groundwater management.  Our established workflow enables the study to be extended to any other site.

Received: 20 July 2024

Accepted: 03 October 2024

Published: 03 November 2024