From Standard to Bayesian: Revisiting Ocean Color Model Evaluation

Abstract

Ocean color models, facilitated by satellite-based remote sensing technologies, have transformed our ability to monitor chlorophyll-a concentrations at the ocean’s surface. These models decode changes in the spectral properties of reflected sunlight to map phytoplankton biomass across extensive spatial and temporal scales. Such capabilities are crucial for understanding marine primary productivity, carbon cycling, and ecosystem reactions to environmental change. Additionally, improvements in atmospheric correction and inwater algorithms over the past decades have reduced the uncertainties associated with chlorophyll-a retrieval, rendering satellite-derived ocean color data more suitable for fisheries management, climate research, and coastal water quality monitoring. However, the most commonly used ocean color products (e.g., Sentinel datasets) do not provide uncertainties, which poses challenges in their application for analyzing events like algal blooms and for risk assessment. Moreover, these models are typically evaluated using deterministic error metrics that may not always capture the subtleties in model performance. This paper presents a probabilistic alternative to traditional ocean color models, including the maximum band and color index models, utilizing Bayesian regression and underpinned by an exhaustive database of validated bio-optical match-ups. We further propose a method for comparing ocean color models using the Bayesian information criterion, which evaluates models more effectively by emphasizing those with minimal deviation and mean difference relative to reference measurements. Our findings indicate that, depending on the application, the commonly used OC4 model may not necessarily be the most effective option available.