Modeling lake trophic state: a random forest approach

Abstract

Productivity of lentic ecosystems is well studied, and it is widely accepted that as nutrient inputs increase, productivity increases and lakes transition from lower trophic state (e.g., oligotrophic) to higher trophic states (e.g., eutrophic). These broad trophic state classifications are good predictors of ecosystem condition, services (e.g., recreation and esthetics), and disservices (e.g., harmful algal blooms). While the relationship between nutrients and trophic state provides reliable predictions, it requires in situ water quality data to parameterize the model. This limits the application of these models to lakes with existing and, more importantly, available water quality data. To address this, we take advantage of the availability of a large national lakes water quality database (i.e., the National Lakes Assessment), land‐use/land‐cover data, lake morphometry data, and other universally available data, and we apply data‐mining approaches to predict trophic state. Using these data and random forests, we first model chlorophyll a and then classify the resultant predictions into trophic states. The full model estimates chlorophyll a with both in situ and universally available data. The mean‐squared error and adjusted R2 of this model was 0.09 and 0.8, respectively. The second model uses universally available GIS data only. The mean‐squared error was 0.22, and the adjusted R2 was 0.48. The Kappa coefficients of the trophic state classifications derived from the chlorophyll a predictions were 0.57 for the full model and 0.29 for the “GIS‐only” model. Random forests extend the usefulness of the class predictions by providing prediction probabilities for each lake. This allows us to make trophic state predictions and also indicate the level of uncertainty around those predictions. For the full model, these predicted class probabilities ranged from 0.42 to 1. For the GIS‐only model, they ranged from 0.33 to 0.96. It is our conclusion that in situ data are required for better predictions, yet GIS and universally available data provide trophic state predictions, with estimated uncertainty, that still have the potential for a broad array of applications. The source code and data for this manuscript are available from https://github.com/USEPA/LakeTrophicModelling.

Mapping the temporary and perennial character of whole river networks

“…Moreover, recent studies have shown that RF models predict spatial patterns in river characteristics better than other more conventional methods [e.g., Booker and Snelder , ]. The RF technique has been previously applied in water resource studies to predict spatial patterns of different ecosystem components such as river bed surface grain size [ Snelder et al ., ], biotic indices [ Álvarez‐Cabria et al ., ], and lake trophic state [ Hollister et al ., ], among others.…”

A Modeling Comparison of Groundwater Potential Mapping in a Mountain Bedrock Aquifer: QUEST, GARP, and RF Models

“…Furthermore, RF provides two importance ranking indices, "mean decrease Gini" and "mean decrease accuracy" (also termed percent increase in mean square error (MSE)) [58]. The higher the MSE percentage, the greater the importance of the independent variable considered [59]. These indices can be applied to reveal the sensitivity of the model to input variables and to help decrease uncertainty.…”

“…The variable importance assessment in this study was conducted using the RF model. As one of its main benefits, RF can estimate the importance of predictor variables applied in the modeling process using the index percentage increase in mean square error (PIMSE) [59,71]. Table 6 shows the importance of predictor variables in predicting groundwater potential.…”

Improving Predictive Models of In‐Stream Phosphorus Concentration Based on Nationally‐Available Spatial Data Coverages

“…While readily available spatial data coverages empower users across all levels of research and governance, they must be used with some caution in specific applications. In studies of lakes throughout the U.S., for example, lake‐specific variables are known to produce significantly improved predictive models of water quality and trophic state than models based on nationally‐available spatial covariates alone (Read et al ., ; Hollister et al ., ). Thus, while national databases enable prediction of response variables in lakes lacking in situ data, these predictions can be greatly improved with additional data not currently available nationally.…”