Satellite Remote Sensing of Inland Water Quality: Sensor Evaluation, Atmospheric Correction, and Empirical Modeling

Abstract

Inland waters represent a critical yet understudied environment in satellite Remote Sensing (RS), occupying a unique position between terrestrial and open oceans systems that makes them particularly challenging to monitor from space. Despite their ecological and socioeconomic importance, only 60% of known inland waters meet good ambient water quality standards, underscoring the need for cost-effective, large-scale monitoring solutions. Satellite RS offers such a solution, but its effectiveness depends on three interconnected components: suitable sensors, accurate atmospheric correction, and robust retrieval models. This thesis advances each of these components through four studies.

Chapter 1 provides a comprehensive review of satellite RS for inland Water Quality Parameter (WQP) retrieval, synthesizing developments in sensor technologies, atmospheric correction algorithms, and Machine Learning (ML) models. The review identifies three key research gaps that motivate the subsequent chapters: the potential of hyperspectral data for inland WQP retrieval, the impact of Adjacency Effect (AE) correction on WQP retrieval, and the influence of spatial autocorrelation on ML model evaluation.

Chapter 2 addresses hyperspectral sensor capability by evaluating Wyvern Dragonette-001 against Sentinel-2 MSI for retrieving chlorophyll-a, colored dissolved organic matter, dissolved organic carbon, suspended solids, and turbidity in the St. Lawrence River, Canada. Wyvern demonstrated competitive performance for turbidity retrieval but underperformed relative to Sentinel-2 for chlorophyll-a, colored dissolved organic matter, presumably due to the absence of blue spectral bands. These results highlight that hyperspectral capability alone does not guarantee superior WQP retrieval performance.

Chapter 3 addresses atmospheric correction by investigating the impact of AE correction on turbidity retrieval from Landsat 8/9 and Sentinel-2 imagery in the St. Lawrence River. Applying T-Mart AE correction prior to atmospheric correction with ACOLITE's Dark Spectrum Fitting technique improved retrieval accuracy for both sensors, with Root Mean Square Error reductions of 48.8% and 28.4% for Landsat 8/9 and Sentinel-2 respectively, compared to atmospheric correction alone. The most pronounced improvements were observed within 300 meters of the shoreline, where AE from surrounding land are strongest, demonstrating that AE correction is a meaningful and necessary preprocessing step in riverine environments.

Chapter 4 addresses WQP retrieval modeling through a comprehensive comparison of 24 statistical and ML models for salinity mapping in the Karun River, Iran, using Landsat 8 OLI imagery. Decision tree-based ensemble models, particularly Gradient Boosted Decision Trees, achieved the highest accuracy while also exhibiting non-significant residual spatial autocorrelation. A key finding is that strong statistical performance metrics do not necessarily indicate spatially uncorrelated residuals, and vice versa, demonstrating that residual spatial autocorrelation analysis is an essential complement to conventional performance evaluation when satellite imagery is used to map WQP.

Together, these contributions demonstrate that advances in inland WQP retrieval depend on progress across three distinct but interconnected components: sensor design, atmospheric correction algorithms, and retrieval methodology. Moreover, the findings provide actionable guidance for researchers and practitioners developing satellite-based monitoring systems for inland waters.