Scholarlink

<p>Located in northwestern Morocco, the Rmel aquifer represents a vital resource for both drinking water supply and irrigated agriculture. It is under increasing stress, with an estimated annual deficit of 3.3&#xa0;million m³. This study introduces an innovative methodological framework that integrates a comparative analysis between machine learning (ML) algorithms namely Random Forest, XGBoost, and AdaBoost within a two-level modeling architecture, and advanced deep learning (DL) models, including Long Short-Term Memory (LSTM), Generative Adversarial Networks (GAN), Deep Learning Neural Networks (DLNN), and Recurrent Neural Networks (RNN). The primary objective is to delineate high groundwater potential zones by incorporating 17 conditioning factors and modeling their complex, nonlinear interdependencies. The hybrid model combining Random Forest and AdaBoost within XGBoost delivers outstanding predictive performance, achieving an AUC of 0.981, outperforming all other individual ML algorithms and DL models. The resulting spatial distribution indicates that approximately 10% of the study area, mainly in the northern part, exhibits very high potential, while about 25%, predominantly in the south, shows very low suitability. The most influential factors include piezometric level (information gain: 0.19), precipitation (0.18), and aquifer depth (0.16). This integrative approach significantly outperforms conventional methods, providing robust decision-support tools to enhance artificial recharge, regulate groundwater extraction, and optimize well placement thereby contributing to the sustainable management of groundwater resources in similar contexts.</p> Graphical Abstract <p>The Rmel aquifer, Located in Northern Morocco, Spans Approximately 240&#xa0;km² and Serves as a Vital Water Resource for 178,338 inhabitants, Supporting Agriculture and the Local Economy Amid Increasing Sustainability challenges. This Study Presents an Advanced Groundwater Potential Modeling Approach by Integrating a two-level Machine Learning Framework (Random Forest, XGBoost, AdaBoost) with Deep Learning Techniques (RNN, LSTM, GAN, DLNN). the Analysis Is Based on a Rigorously Validated Dataset Comprising 200 Observation Points and Multiple Data sources, Including Precipitation and Temperature records, Soil Type (pedological map), Satellite Imagery (Landsat-8), and a Digital Elevation Model (USGS). the Key Influencing Factors Were Categorized into Five Groups: Topographic, Land use, geological, hydrological, and Hydrogeological parameters, with the Most Significant Being Piezometric Level (IG = 0.19), Precipitation (IG = 0.18), and Groundwater Depth (IG = 0.16). Groundwater Potential mapping, Performed Through Kriging Interpolation and Spatial standardization, Highlights a Pronounced North-south Contrast: High Potential in the North (Larache) Due To Favorable Recharge Conditions and Low Potential in the south, where Steep Slopes and Low Permeability Hinder Water retention. the Study results, Particularly the Model Performance Evaluation Via AUC-ROC, Confirm the Superiority of the Ensemble RF-AdaBoost-XGBoost Approach (AUC = 0.981), Providing a More Stable and Accurate Classification Compared To Individual Machine Learning and Deep Learning Models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparison between Two-Level Machine Learning and Deep Learning for Groundwater Potential Mapping in the Rmel Aquifer (Northwestern Morocco)

  • 期刊论文

  • 发表日期：

    2025年10月7日

查看全文
  • Morad Chahid,
  • Jamal Eddine Stitou El Messari,
  • Ismail Hilal,
  • Zouhir Dichane,
  • Rabin Chakrabortty,
  • Sanju Purohit,
  • Mourad Aqnouy

摘要

Located in northwestern Morocco, the Rmel aquifer represents a vital resource for both drinking water supply and irrigated agriculture. It is under increasing stress, with an estimated annual deficit of 3.3 million m³. This study introduces an innovative methodological framework that integrates a comparative analysis between machine learning (ML) algorithms namely Random Forest, XGBoost, and AdaBoost within a two-level modeling architecture, and advanced deep learning (DL) models, including Long Short-Term Memory (LSTM), Generative Adversarial Networks (GAN), Deep Learning Neural Networks (DLNN), and Recurrent Neural Networks (RNN). The primary objective is to delineate high groundwater potential zones by incorporating 17 conditioning factors and modeling their complex, nonlinear interdependencies. The hybrid model combining Random Forest and AdaBoost within XGBoost delivers outstanding predictive performance, achieving an AUC of 0.981, outperforming all other individual ML algorithms and DL models. The resulting spatial distribution indicates that approximately 10% of the study area, mainly in the northern part, exhibits very high potential, while about 25%, predominantly in the south, shows very low suitability. The most influential factors include piezometric level (information gain: 0.19), precipitation (0.18), and aquifer depth (0.16). This integrative approach significantly outperforms conventional methods, providing robust decision-support tools to enhance artificial recharge, regulate groundwater extraction, and optimize well placement thereby contributing to the sustainable management of groundwater resources in similar contexts.

Graphical Abstract

The Rmel aquifer, Located in Northern Morocco, Spans Approximately 240 km² and Serves as a Vital Water Resource for 178,338 inhabitants, Supporting Agriculture and the Local Economy Amid Increasing Sustainability challenges. This Study Presents an Advanced Groundwater Potential Modeling Approach by Integrating a two-level Machine Learning Framework (Random Forest, XGBoost, AdaBoost) with Deep Learning Techniques (RNN, LSTM, GAN, DLNN). the Analysis Is Based on a Rigorously Validated Dataset Comprising 200 Observation Points and Multiple Data sources, Including Precipitation and Temperature records, Soil Type (pedological map), Satellite Imagery (Landsat-8), and a Digital Elevation Model (USGS). the Key Influencing Factors Were Categorized into Five Groups: Topographic, Land use, geological, hydrological, and Hydrogeological parameters, with the Most Significant Being Piezometric Level (IG = 0.19), Precipitation (IG = 0.18), and Groundwater Depth (IG = 0.16). Groundwater Potential mapping, Performed Through Kriging Interpolation and Spatial standardization, Highlights a Pronounced North-south Contrast: High Potential in the North (Larache) Due To Favorable Recharge Conditions and Low Potential in the south, where Steep Slopes and Low Permeability Hinder Water retention. the Study results, Particularly the Model Performance Evaluation Via AUC-ROC, Confirm the Superiority of the Ensemble RF-AdaBoost-XGBoost Approach (AUC = 0.981), Providing a More Stable and Accurate Classification Compared To Individual Machine Learning and Deep Learning Models.