From Standard to Bayesian: Revisiting Ocean Color Model Evaluation

We revisit the evaluation framework used to assess ocean color algorithms by moving from standard deterministic comparisons toward a Bayesian perspective. Rather than ranking models by point-wise error metrics alone, we propagate observational and parameter uncertainty through the retrieval problem and compare the resulting posterior predictive distributions. The approach yields uncertainty-aware skill scores, calibrated credible intervals on chlorophyll-a retrievals, and a transparent way to weigh model performance across regimes (oligotrophic, mesotrophic, eutrophic) and ecological provinces. We argue that adopting a Bayesian evaluation protocol changes which algorithms appear “best” in operational contexts and offers a more honest baseline for the next generation of data-driven ocean color models.

AGU Earth and Space Science