Scholarlink

<p>Coastal aquifers are vital freshwater resources in arid regions but are increasingly threatened by seawater intrusion (SWI), which compromises groundwater quality and long-term sustainability. While traditional models have provided valuable insights into SWI dynamics, they often struggle to capture the non-linear, spatiotemporal variability of coastal systems, particularly under data scarcity and incomplete environmental characterization. This study presents the first comparative application of four deep learning architectures, Feed-Forward Neural Network (FFNN), Long Short-Term Memory (LSTM), Transformer, and a hybrid LSTM-FFNN model, for the prediction of SWI using Total Dissolved Solids (TDS) as a salinity proxy. The models were trained on over 14,000 daily records from 16 coastal wells in the Emirate of Fujairah, UAE. Data pre-processing included K-Nearest Neighbors (KNN) imputation, outlier removal, and feature normalization to enhance temporal continuity and statistical robustness. While the Transformer initially outperformed other models under mean imputation, the integration of KNN and an attention layer within LSTM significantly improved accuracy, allowing LSTM and LSTM-FFNN to surpass the Transformer across most evaluation metrics. The LSTM model achieved the highest overall performance (MAE = 401&#xa0;mg/l, R² = 0.983). The spatial and seasonal analyses further emphasized the superior stability of LSTM-based models, particularly in high-stress summer months and inland zones This study highlights the strength of attention-augmented recurrent models in forecasting complex groundwater salinity dynamics. The proposed deep learning framework offers a scalable and adaptable solution for early warning and salinity risk management in vulnerable coastal aquifers, especially under evolving climatic and anthropogenic pressures. This graphical abstract offers a concise visual summary of the study’s methodology and key findings. The research presents a novel deep learning framework to predict seawater intrusion (SWI) in Fujairah, United Arab Emirates, a hyper-arid coastal region highly susceptible to SWI. The approach models total dissolved solids (TDS) as a proxy for SWI progression, utilizing over 14,000 daily observations from 16 coastal wells. Influential features include aquifer hydraulic properties, precipitation, and proximity to the shoreline. The workflow is structured into three main stages: data collection, pre-processing, and implementation of three deep learning models—Feed-Forward Neural Network (FFNN), Long Short-Term Memory (LSTM), and Transformer along the hybrid LSTM-FFNN model. Among these, the LSTM model outperformed others due to its ability to capture long-range temporal dependencies inherent in time-series data upon equipping it with attention layer and KNN imputation. The LSTM model demonstrated consistent spatial accuracy and temporal robustness during seasonal salinity peaks. These results underscore the model’s potential to support real-time, adaptive groundwater management in hyper-arid coastal aquifers vulnerable to SWI.</p> Graphical Abstract <p>This graphical abstract offers a concise visual summary of the study’s methodology and key findings. The research presents a novel deep learning framework to predict seawater intrusion (SWI) in Fujairah, United Arab Emirates, a hyper-arid coastal region highly susceptible to SWI. The approach models total dissolved solids (TDS) as a proxy for SWI progression, utilizing over 14,000 daily observations from 16 coastal wells. Influential features include aquifer hydraulic properties, precipitation, and proximity to the shoreline. The workflow is structured into three main stages: data collection, pre-processing, and implementation of three deep learning models—Feed-Forward Neural Network (FFNN), Long Short-Term Memory (LSTM), and Transformer along the hybrid LSTM-FFNN model. Among these, the LSTM model outperformed others due to its ability to capture long-range temporal dependencies inherent in time-series data upon equipping it with attention layer and KNN imputation. The LSTM model demonstrated consistent spatial accuracy and temporal robustness during seasonal salinity peaks. These results underscore the model’s potential to support real-time, adaptive groundwater management in hyper-arid coastal aquifers vulnerable to SWI.</p> <p></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

Dynamic Prediction of Seawater Intrusion in Hyper-Arid Coastal Aquifer Using Advanced Deep Learning

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Assaad Kassem,
  • Ahmed Sefelnasr,
  • Nadeem Iqbal Kajla,
  • Faisal Baig,
  • Mohsen Sherif

摘要

Coastal aquifers are vital freshwater resources in arid regions but are increasingly threatened by seawater intrusion (SWI), which compromises groundwater quality and long-term sustainability. While traditional models have provided valuable insights into SWI dynamics, they often struggle to capture the non-linear, spatiotemporal variability of coastal systems, particularly under data scarcity and incomplete environmental characterization. This study presents the first comparative application of four deep learning architectures, Feed-Forward Neural Network (FFNN), Long Short-Term Memory (LSTM), Transformer, and a hybrid LSTM-FFNN model, for the prediction of SWI using Total Dissolved Solids (TDS) as a salinity proxy. The models were trained on over 14,000 daily records from 16 coastal wells in the Emirate of Fujairah, UAE. Data pre-processing included K-Nearest Neighbors (KNN) imputation, outlier removal, and feature normalization to enhance temporal continuity and statistical robustness. While the Transformer initially outperformed other models under mean imputation, the integration of KNN and an attention layer within LSTM significantly improved accuracy, allowing LSTM and LSTM-FFNN to surpass the Transformer across most evaluation metrics. The LSTM model achieved the highest overall performance (MAE = 401 mg/l, R² = 0.983). The spatial and seasonal analyses further emphasized the superior stability of LSTM-based models, particularly in high-stress summer months and inland zones This study highlights the strength of attention-augmented recurrent models in forecasting complex groundwater salinity dynamics. The proposed deep learning framework offers a scalable and adaptable solution for early warning and salinity risk management in vulnerable coastal aquifers, especially under evolving climatic and anthropogenic pressures. This graphical abstract offers a concise visual summary of the study’s methodology and key findings. The research presents a novel deep learning framework to predict seawater intrusion (SWI) in Fujairah, United Arab Emirates, a hyper-arid coastal region highly susceptible to SWI. The approach models total dissolved solids (TDS) as a proxy for SWI progression, utilizing over 14,000 daily observations from 16 coastal wells. Influential features include aquifer hydraulic properties, precipitation, and proximity to the shoreline. The workflow is structured into three main stages: data collection, pre-processing, and implementation of three deep learning models—Feed-Forward Neural Network (FFNN), Long Short-Term Memory (LSTM), and Transformer along the hybrid LSTM-FFNN model. Among these, the LSTM model outperformed others due to its ability to capture long-range temporal dependencies inherent in time-series data upon equipping it with attention layer and KNN imputation. The LSTM model demonstrated consistent spatial accuracy and temporal robustness during seasonal salinity peaks. These results underscore the model’s potential to support real-time, adaptive groundwater management in hyper-arid coastal aquifers vulnerable to SWI.

Graphical Abstract

This graphical abstract offers a concise visual summary of the study’s methodology and key findings. The research presents a novel deep learning framework to predict seawater intrusion (SWI) in Fujairah, United Arab Emirates, a hyper-arid coastal region highly susceptible to SWI. The approach models total dissolved solids (TDS) as a proxy for SWI progression, utilizing over 14,000 daily observations from 16 coastal wells. Influential features include aquifer hydraulic properties, precipitation, and proximity to the shoreline. The workflow is structured into three main stages: data collection, pre-processing, and implementation of three deep learning models—Feed-Forward Neural Network (FFNN), Long Short-Term Memory (LSTM), and Transformer along the hybrid LSTM-FFNN model. Among these, the LSTM model outperformed others due to its ability to capture long-range temporal dependencies inherent in time-series data upon equipping it with attention layer and KNN imputation. The LSTM model demonstrated consistent spatial accuracy and temporal robustness during seasonal salinity peaks. These results underscore the model’s potential to support real-time, adaptive groundwater management in hyper-arid coastal aquifers vulnerable to SWI.