Training-Domain Gaps in Data-Driven Ocean Color Algorithms

Bio-optical algorithms serve a critical role, enabling satellite-based monitoring of the global ocean and its ecosystems. These largely empirical algorithms invert satellite measurements of water leaving reflectance (“ocean color”) to biogeochemical parameters such as chlorophyll a, but often fail to generalize beyond their training data, leading to increased uncertainty when extrapolated. We present a simple geometric approach for assessing bio-optical model domain using a convex hull. Using NASA Aqua-MODIS observations, and selected comparisons with PACE-OCI, we find large swaths of the ocean are not within the domain of existing bio-optical models, with size and location of gaps varying across seasons, and becoming more pronounced as input dimensionality increases. Understanding these limitations is critical in a changing ocean. Our work provides an interpretable metric that points to specific regions and times that should be flagged in the short-term and sampled in the long-term, to better understand the ocean's ecological and biogeochemical diversity.