Interpretable Machine Learning Identifies Eutrophication Factors of a Landscape Lake Recharged by Reclaimed Water

Abstract

The water quality of lakes recharged by reclaimed water is affected by both the fluctuation of reclaimed water quality and the biochemical processes in the lakes, so the main controlling factors of algae blooms are always difficult to identify. Taking a landscape lake recharged by reclaimed water as an example, the spatiotemporal distribution characteristics and correlation of water quality indexes were analyzed. It was found that the trend of lake water quality was highly consistent with but different from reclaimed water quality, and the lake alternated between hyper-eutrophication and eutrophication throughout the year by seasonally driven, but chlorophyll a (Chl-a) had no significant correlation with other indicators, except for a moderate negative correlation with NO-3-N and a weak negative correlation with NO-2-N. Furthermore, taking nutrient difference indexes (NDIs) between reclaimed water and lake water into account, Chl-a concentration was predicted by random forest, and the relationship between Chl-a and each variable was analyzed by a partial dependence plot. Results showed that ∆NO-3-N was an important and positive NDI explanatory variable for Chl-a prediction, which proved that NO-3-N input from reclaimed water was the dominant factor causing eutrophication, and the negative correlation between NO-3-N and Chl-a in lake water was the result of algae blooms rather than the cause. Our study provides new insight into the identification of eutrophication factors for lakes recharged by reclaimed water.

Keywords: Interpretable machine learning, Reclaimed water, landscape lake, Eutrophication, Chl-a prediction, random forest