Data-Driven Analysis of Pond-Based Recharge in Daratoo and Tarin Pilot Sites in Erbil City Using Internet of Things and Machine Learning

Abstract

Groundwater depletion is a pressing issue in semi-arid regions. In Erbil’s environs, over-extraction and reduced precipitation have driven local water-table declines. This study evaluates the effectiveness of pond-based managed aquifer recharge (MAR) at two pilot sites, Tarin and Daratoo, by integrating high-frequency Internet of Things (IoT) monitoring with machine learning (ML). Diver-HUB sensors recorded water pressure, temperature, electrical conductivity, and salinity every 15 minutes in recharge and monitoring wells. Pressure readings were barometrically corrected and used for time-series prediction of groundwater depth. Four modelling approaches (linear regression, neural networks, eXtreme Gradient Boosting (XGBoost), and gradient boosting (GBoost) were compared, and a random forest feature-importance analysis identified water pressure as the dominant predictor. During the wet season, observed water-table rises were 1.29 m at Tarin and 2.4 m at Daratoo, indicating successful short-term recharge. Linear regression produced the best predictive performance on test data (near-unity R² and very low RMSE), but these results are tempered by site specificity, the limited number of pilot sites, and the short observation window. The contribution of this work is a data-driven, deployable framework that combines IoT and ML to support real-time monitoring, inform MAR design and operation, and guide adaptive groundwater management. Broader validation across more sites and longer periods is recommended to establish transferability and long-term sustainability.