Spatiotemporal mapping of inland water quality variables: A comparative analysis of Gaussian process regression and spectral indices using Sentinel-2 multispectral imagery

Monitoring ecological water quality in optically complex inland waters remains challenging due to nonlinear interactions among multiple optically active constituents (OACs) and the limited spectral information content of multispectral satellite sensors. This study evaluates the potential of Sentinel-2 (S2) Level-2 A (L2A) multispectral imagery for monitoring key ecological water quality indicators in the hypertrophic Albufera of València lagoon through a systematic comparison between spectral index (SI)-based retrievals and Gaussian process regression (GPR). Using a common in situ dataset collected between 2016 and 2023, retrieval models were developed for chlorophyll– a (Chl– a ), phycocyanin (PC), and Secchi disk depth (SDD), and implemented within a cloud-based Google Earth Engine (GEE) workflow to reconstruct multi-annual trajectories between 2017 and 2026. Cross-validation results demonstrated high retrieval performance for both approaches. The GPR framework achieved modest improvements over SI-based models specifically for Chl– a ( R 2 = 0.64 vs. 0.61, respectively) and SDD ( R 2 = 0.64 vs. 0.57, respectively). Conversely, the targeted SI configuration exhibited superior predictive resilience for PC, outperforming GPR with an R 2 of 0.59 compared to 0.57. The added value of GPR was most evident for capturing complementary multi-band relationships in Chl– a and SDD retrieval. In contrast, the fine-tuned empirical spectral formulation proved more robust against full-spectrum background noise for the highly complex PC signal, where full-spectrum nonparametric flexibility led to slight over-parameterisation. Multi-annual reconstructions revealed persistent eutrophic-to-hypertrophic conditions characterised by elevated phytoplankton biomass, reduced water transparency, and pronounced seasonal and interannual variability. Overall, the study highlights the complementary roles of tailored semi-empirical spectral formulations and uncertainty-aware machine learning methods for scalable long-term monitoring of ecological water quality in optically complex inland waters.