Innovating saltwater intrusion monitoring: Deep Learning-driven sensor fusion for soil salinity and moisture measurement

Saltwater intrusion (SWI) presents a growing threat to coastal groundwater resources, with serious implications for agriculture, infrastructure, and environmental sustainability. This study introduces a sensor fusion methodology that integrates three distinct measurement principles (i.e., dielectric permittivity, electrical conductivity, and acoustic impedance) into a heteroscedastic deep neural network (DNN) framework to accurately predict soil pore salinity and volumetric water content (VWC). The model was trained on a laboratory dataset spanning four representative soil types, including SC (clayey sand), CL-ML (low-plasticity silty clay), CL (lean clay), and SP (poorly graded sand). The model achieved high predictive performance in the lab (MAPE = 2.68% for VWC, 5.326% for salinity) and showed encouraging performance in a preliminary evaluation on unseen field samples from two coastal sites in Maine (MAPE = 9.185% for VWC, 9.011% for salinity). The proposed framework demonstrates generalization across the tested soil types, substantially reducing the burden of site-specific calibration. Additionally, the study employs heteroscedastic regression incorporating Monte Carlo (MC) dropout to quantify both epistemic (model-based) and aleatoric (data-based) uncertainties associated with the predictions. Predictions were most reliable for moderate-to-high moisture conditions (VWC > 0.17 m³/m³) and low-to-moderate salinity levels (<20 g/L), which are typical field monitoring ranges. Uncertainty quantification enables users to identify predictions requiring additional verification. This research underscores the importance of multimodal sensing and advanced machine-learning techniques in environmental monitoring. The proposed solution offers a more accurate, versatile, and reliable method for SWI detection, benefiting applications in agriculture, environmental monitoring, and geotechnical engineering.