Climate change is likely to have strong effects on ecosystems, including water quality in lakes and streams. However, few studies have evaluated the effects of climate across lakes over broad spatial extents, or contrasted the relative importance of climate predictors for several water quality response variables. While extensive spatial and temporal data required for broad scale comparisons are difficult to synthesize, macroscale studies are important because the effects of change and important climate drivers are likely to differ across regions. To address this gap, we used a database that integrates both lake water quality and geospatial data, and developed a new multi-task learning algorithm to examine how well climate data can predict water quality in ~11,000 lakes over a 17 US state extent. Our analysis included 48 climate predictors that describe temperature, precipitation, and climate at monthly, seasonal and annual time steps. We identified which climate variables were the best predictors of four water quality response variables: Secchi depth, total nitrogen, total phosphorus, chlorophyll a. The multi-task learning algorithm allowed models for different responses to be trained simultaneously and accommodated spatial autocorrelation and correlations among response variables, enabling predictions for all lakes, even those with missing data.
Results/Conclusions
Multi-task learning models showed relationships between water quality and climate, with R2 values above 0.4 for all response variables and over 0.7 for Secchi and total nitrogen. For all four water quality response variables, precipitation from the previous winter was most important for predicting summer water quality. Early summer temperatures were also important for predicting chlorophyll a and nutrients, but not Secchi depth. In contrast, annual precipitation and temperature from the previous year were important predictors of Secchi depth but not the other water quality metrics. Although we found evidence of spatial structure in the association between climate predictors and lake water quality, it was difficult to directly link these associations with lake characteristics because they were spatially correlated in a similar manner as the climate predictors. Our results suggest that winter precipitation is a dominant driver of summer lake water quality but that the magnitude or direction of those relationships can differ among response variables, reflecting the different mechanisms by which temperature and precipitation influence the chemical, physical and biological aspects of lakes.