Evaluating remote sensing model specification methods for estimatingwater quality in optically diverse lakes throughout the growing season

Spectral images from remote sensing platforms are extensively used to estimate chlorophyll-a (chl-a) concentrations for water quality studies. Empirical models used for estimation are often based on physical principles related to light absorption and emission properties of chl-a and generally relying on spectral bands in the green, blue, and near-infrared bands. Because the physical characteristics, constituents, and algae populations vary widely from lake to lake, it can be difficult to estimate coefficients for these models. Many studies select a model form that is a function of these bands, determine model coefficients by correlating remotely-measured surface reflectance data and coincidentally measured in-situ chl-a concentrations, and then apply the model to estimate chl-a concentrations for the entire water body. Recent work has demonstrated an alternative approach using simple statistical learning methods (Multiple Linear Stepwise Regression (MLSR)) which uses historical, non-coincident field data to develop sub-seasonal remote sensing chl-a models. We extend this previous work by comparing this method against models from literature, and explore model performance for a region of lakes in Central Utah with varying optical complexity, including two relatively clear intermountain reservoirs (Deer Creek and Jordanelle) and a highly turbid, shallow lake (Utah Lake). This study evaluates the suitability of these different methods for model parameterization for this area and whether a sub-seasonal approach improves performance of standard model forms from literature. We found that while some of the common spectral bands used in literature are selected by the data-driven MLSR method for the lakes in the study region, there are also other spectral bands and band interactions that are often more significant for these lakes. Comparison of model fit shows an improvement in model fit using the data-driven parameterization method over the more traditional physics-based modeling approaches from literature. This suggests that the sub-seasonal approach and exploitation of information contained in other bands helps account for lake-specific optical characteristics, such as suspended solids and other constituents contributing to water color, as well as unique (and season-specific) algae populations, which contribute to the spectral signature of the lake surface, rather than only relying on a generalized optical signature of chl-a. Consideration of these other bands is important for development of models for long-term and entire growing season applications in optically diverse water bodies.