نوع مقاله : مقاله پژوهشی
نویسندگان
1 دانشجوی کارشناسی ارشد مهندسی منابع آب، گروه مهندسی آبیاری و آبادانی دانشکدگان کشاورزی و منابع طبیعی، دانشگاه تهران، کرج، ایران.
2 دانشجوی کارشناسی ارشد مهندسی منابع آب، گروه مهندسی آبیاری و آبادانی ، دانشکدگان کشاورزی و منابع طبیعی، دانشگاه تهران
3 دانشیار گروه مهندسی آبیاری و آبادانی دانشکدگان کشاورزی و منابع طبیعی، دانشگاه تهران، کرج، ایران.
چکیده
کلیدواژهها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
Accurate modeling of groundwater quality requires a proper understanding of the spatial dependency structure of hydrochemical parameters and the application of advanced predictive approaches. In this study, six machine learning algorithms including K Nearest Neighbors (KNN), Multi Layer Perceptron (MLP), Random Forest (RF), Gaussian Process Regression (GPR), Gradient Boosting Regression (GBR) and Extreme Gradient Boosting (XGBoost) were employed to predict electrical conductivity (EC) and calcium (Ca²⁺) concentrations in Guilan Province, Iran. To incorporate spatial dependency into the predictive framework, the distances to neighboring wells were introduced as input features, and model performance was evaluated under different neighborhood scenarios (4 to 8 neighboring wells) using annual groundwater quality data from 2002 to 2018.
The results indicate that ensemble learning models outperform the other algorithms; however, the optimal spatial configuration varies between parameters. For electrical conductivity (EC), the Random Forest model achieved the best performance when eight neighboring wells were considered, with an R‑squared value of 0.888, a root mean square error of 117.165, and a mean absolute error of 84.173. This suggests a broader spatial dependency for this parameter within the aquifer system. In contrast, for calcium ion (Ca²⁺), the XGBoost model yielded the optimal performance with four neighboring wells, achieving an R‑squared value of 0.811, a root mean square error of 1.154, and a mean absolute error of 0.820, indicating that local hydrogeochemical processes exert a stronger control on the spatial distribution of this ion.
The comparison of neighborhood scenarios further demonstrates that increasing the number of neighboring wells does not necessarily improve prediction accuracy for all parameters, highlighting the importance of parameter specific spatial dependency analysis. Overall, the findings confirm that integrating machine learning with spatial feature engineering provides an effective framework for intelligent groundwater quality modeling and can support the optimization of monitoring networks and water resources management strategies.
کلیدواژهها [English]
)