Predicting sediment contamination in Tunisia's coastal lagoons using an OP-LSTM deep learning model: A case study...

Coastal areas are increasingly threatened by marine sediment contamination resulting from industrial discharge, agricultural runoff, and urban expansion, posing serious risks to marine ecosystems and human health. This study aims to predict sediment contamination risks in the Bizerte Lagoon, Tunisia, by applying an Optimized Long Short-Term Memory (OP-LSTM) deep learning model, supported by comprehensive geochemical and mineralogical analyses. The methodology involved characterizing sediment samples using X-ray diffraction (XRD) to identify mineral species and quantify the clay fraction, while atomic absorption spectroscopy (AAS) was used to determine major and trace element concentrations, with major elements expressed as oxides. The OP-LSTM model was trained and validated using these datasets alongside contamination scoring systems to enhance predictive accuracy. Mineralogical findings revealed eleven dominant minerals: quartz, feldspar, calcite, dolomite, gypsum, hematite, goethite, aragonite, muscovite, kaolinite (75.47 %), illite (27.87 %), and smectites (4.05 %), indicating combined terrigenous and anthropogenic inputs. Major element oxides varied spatially, with concentrations ranging from 0.04 to 1.20 % for K₂O, 0.13 to 1.31 % for Na₂O, 3.86 to 19.47 % for CaO, and 0.18 to 0.48 % for MgO. Trace metals were also detected at concerning levels, with Pb up to 3.974 ppm, Cr up to 1.800 ppm, and Cd up to 817 ppm-exceeding Tunisian environmental standards. The OP-LSTM model outperformed the standard LSTM, exhibiting lower RMSE (0.142-0.164), MSE (0.020-0.027), and training loss (from 0.413 to 0.200), alongside higher R² scores (up to 0.413), confirming its robustness and suitability for effective sediment contamination risk prediction and management in coastal lagoon environments.