A Reliability Framework for Sentinel‑2 Water‑Quality Retrieval in Optically Complex, Data‑Scarce Tropical Inland Water: Evidence from Oguta Lake (Nigeria)

Abstract

A reliability-oriented Sentinel-2 workflow was implemented for Oguta Lake, Nigeria, to retrieve Coloured Dissolved Organic Matter (CDOM), turbidity, chlorophyll a (Chl-a), Secchi disk depth (SDD), and total suspended solids (TSS). Sentinel-2 MSI Level-2A imagery spanning 2020 to 2025 was processed in Google Earth Engine using cloud and shadow masking, NDWI water masking, and radiometric normalisation. Field and laboratory measurements were collected at six sites on 6 February 2025; one site (GT6) was excluded after NDWI water‑domain screening, so model calibration and LOOCV used N = 5 match-ups. Given the extreme calibration scarcity, high capacity regressors were avoided and final models were restricted to low capacity, positivity by design log linear forms, with fixed pipeline leave one out cross validation used for internal stability assessment (MAE and MdAPE). Ground-truth representativeness across optical regimes was evaluated by sampling a K-Means optical stratification (k = 8) at the sampling sites, and mapping-domain coherence was examined using Pearson and Spearman plausibility diagnostics across a stratified auxiliary time series. Results were parameter dependent. CDOM and turbidity supported cautious semi‑quantitative/relative interpretation, while SDD was best treated as relative‑only; TSS and Chl‑a remained exploratory under N = 5.