Observation-Domain-Aware Multisensor Fusion for Intelligent Coastal Water-Quality Monitoring

Abstract

Reliable monitoring of coastal water quality requires the integration of sensors that differ in physical observation domain, spatiotemporal support, and uncertainty structure. We propose an observation-domain-aware multisensor fusion architecture in which CTD/DO/H2S/Eh profiling, laboratory measurements, Sentinel-2 MSI, Sentinel-3 OLCI/SLSTR, Landsat, and hydrometeorological context are combined only after preserving sensor identity, QA/QC status, and the independent physical observation unit. The experimental field dataset comprises four seasonal campaigns, 32 station–campaign groups, 336 depth records, 111 water samples, 1110 laboratory determinations, 16 satellite scenes, and 64 MSI/OLCI–in situ matchups. MSI/OLCI products showed strong agreement with surface chlorophyll-a, turbidity, and TSM measurements (ρ = 0.906–0.991). For hypoxia, nested leave-one-campaign-out (LOCO) validation yielded ROC-AUC/Brier scores of 0.860/0.157 for full fusion, 0.705/0.228 for optical-only sensing, and 0.899/0.045 for the geometry/context baseline, demonstrating non-monotonic fusion gain. A row-wise split that violated group independence produced an optimistically inflated ROC-AUC of 1.000. These results show that an intelligent sensing system should enforce endpoint-specific sensor compatibility, preserve censoring and uncertainty information, and be evaluated using physically independent observation units. Because the same sectors were revisited, the hypoxia metrics represent cross-campaign transfer within one transect, not geometry-independent within-sector change detection.