Improving extreme value prediction for water clarity using weighted regression models

Previous work on predicting water quality indicators has mainly consisted of using both semi-analytical algorithms (SAAs) and empirical approaches, but recently new datadriven machine learning approaches such as neural-networkbased regression models are increasingly being explored for their utility and potential adoption. Although these types of data-driven models may achieve higher accuracy compared to previous methods, they can also be prone to biasing their outputs towards the mean value of the target distribution if model inputs are noisy. This paper investigates using a recently published weighted regression approach to alleviate "mean-centric" bias on these types of water clarity estimators in the Chesapeake Bay. Experiments comparing standard and weighted data-driven regression approaches for Chesapeake Bay Secchi disk depth prediction are performed and results are discussed.