Classification of optical water types in aquatic radiometry

Some 35 years ago Morel and Prieur (1977) introduced the concept of Case-1 and Case-2 waters to ocean colour research in order separate waters that are dominated by chlorophyll-a and those that are not. This separation aided the development of algorithms for each of the respective water types, as different underlying assumptions/generalisations could be made about the optical properties of the observed water bodies, but also created something of an artificial division in the aquatic optics community. At the global scale, most waters exist as part of a continuum of optical conditions, partially due to a continuum of physical and biological forcing factors and partially due to the fact that we are observing a fluid environment that can physically mix or blend.

It remains the case that no single perfect algorithm works optimally across all optical conditions or water types, and we should not expect to find such an algorithm in the near future. Instead, a promising development in recent years has been the move towards a broader optical water type classification and algorithm blending. This approach is founded on the premise that multiple optimal algorithms exist, but, for each, we can define the most suitable optical environments. The strength of the optical water type classification approach has seen its use grow in limnological and oceanographic remote sensing research (Moore et al. 2001, 2014, Jackson et al. 2017, Spyrakos et al. 2018). The utility of optical water classes has also grown beyond algorithm blending (Moore et al. 2001) to include product uncertainty estimation (Jackson et al. 2017), data quality flagging (Wei et al. 2016, Jiang et al 2023), water quality monitoring (Uudeberg et al. 2020) and environmental phenology studies (Trochta et al. 2015).

Unfortunately, although the limnology and ocean optics communities may agree that optical classification is useful, a harmonised approach to the creation and use of the classes has not yet emerged from the research community. Despite recent efforts to move to a unified fuzzy logic scheme (Jai et al. 2021), a diversity of distance metrics, data transformations and cluster optimisation schemes are applied at local scales (Bi et al. 2019, Botha et al. 2020, da Silva et al. 2020, Uudeberg et al. 2020). Though all these approaches provide interesting and useful results, the fragmented nature of the research makes the comparison of water types difficult, impeding collaboration and optimisation of methods. Also, as with most machine learning techniques, unsupervised clustering is susceptible to the problems of insufficient or biased training data, the ‘central tendency’ (Malik, 2020), and overtraining.

It is timely to convene and reconcile these growing issues under a common framework to unify and standardize definitions, interpretations, and uses to establish guidelines for a growing body of developers and users, as well as to close methodological gaps. This need is of particular urgency in light of new missions with hyperspectral capabilities. Where classification strategies most likely will have to be revised and expanded.