Real-Time Machine Learning and Wireless Sensor Network for Coastal Water Quality Monitoring in the Gulf of Aqaba

This study presents an advanced real-time monitoring system integrating wireless sensor networks with machine learning to assess water quality in the Gulf of Aqaba. Our hybrid machine learning framework combines Random Forest (500 trees, node size = 5) for feature selection with Artificial Neural Network ensembles (3-layer MLP with Monte Carlo dropout) for probabilistic forecasting. The system continuously monitors six critical parameters, demonstrating strong predictive performance through rigorous validation: dissolved oxygen (R² = 0.92, RMSE = 0.45 mg/L, 95 % CI (Confidence interval): 0.41 - 0.49), nitrite (R² = 0.85, RMSE = 0.08 mg/L, CI: 0.07 - 0.09), and turbidity (R² = 0.89, RMSE = 2.3 FTU, CI: 2.1 - 2.5). Comprehensive uncertainty analysis revealed prediction intervals of ±0.38 mg/L for DO and ±0.10 mg/L for nitrite, with spatial variability lowest in open waters (CV = 8.2 %) and highest near coastal zones (CV = 15 %). Residual autocorrelation analysis confirmed model reliability (Moran’s I < 0.12, p > 0.05) across the study area. Spatial-temporal analysis identified nitrite as a sensitive pollution indicator, with concentrations reaching 0.12 mg/L near urban outflows compared to background levels (< 0.05 mg/L). The system achieved 92 % accuracy in the early detection of environmental risks, including coral bleaching precursors (temperature anomalies > 1 °C) and pollution events (nitrite spikes > 0.1 mg/L). Compared to conventional monitoring, the platform demonstrated 20.4 % greater predictive accuracy (ΔR² = +0.17, p < 0.01) while reducing operational costs by 30.2 %, primarily through automated data collection and reduced manual sampling. The integration of high-frequency sensing, adaptive machine learning, and cloud-based analytics establishes a replicable framework for coastal ecosystem management, particularly in anthropogenically stressed environments like the Red Sea.