Predicting biological conditions for small headwater streams in the Chesapeake Bay watershed

Abstract

A primary goal for Chesapeake Bay watershed restoration is to improve stream health and function in 10% of stream miles by 2025. Predictive spatial modeling of stream conditions, when accurate, is one method to fill gaps in monitoring coverage and estimate baseline conditions for restoration goals. Predictive modeling can also monitor progress as additional data become available. We developed a random forests model to predict biological condition of small streams (<200 km 2 in drainage) in the Chesapeake Bay watershed. Biological condition was measured with the Chesapeake Bay Basin-wide Index of Biotic Integrity (Chessie BIBI), a stream macroinvertebrate index. Our goal was to predict biological condition in all unsurveyed small streams present in a 1:24,000 scale catchment layer as a 2004-2008 baseline. We reclassified the 5-category Chessie BIBI ratings into two categories, poor and fair/good, to align with management goals of the Chesapeake Bay Program. The model included 12 geospatial predictor variables including measures on spatial location, bioregion, land cover, soil, precipitation, and number of dams in local catchments. We trained the model with a random 75% subset of Chessie BIBI data (n 5 1449), and used the remaining 25% of Chessie BIBI data (n 5 484) as test data. The model performed well, correctly predicting 72% of samples in training data and 73% of samples in test data, but model accuracy varied among bioregions. We performed uncertainty analyses by adding bands of either ±0.05 or ±0.10 BIBI units to the cutoff between poor and fair/good. These uncertainty analyses resulted in 14.5% (±0.05 band) and 24.8% (±0.10 band) of samples in test data being classified as in uncertain condition. For 95,877 small stream reaches in the Chesapeake Bay watershed, the model predicted 64% in fair/good condition, the ±0.05 uncertainty analyses predicted 57% in fair/good condition, and the ±0.10 uncertainty analysis predicted 50% in fair/good condition. These reported values have different implications for the number of improved stream miles required to meet the goal of improving 10%. Incorporating uncertainty provides an assessment of model strength as well as confidence in predictions. We, therefore, suggest increased reporting of uncertainty in studies that spatially predict stream conditions.

Disentangling the potential effects of land‐use and climate change on stream conditions

“…Model accuracy is an important component of any prediction study. Our accuracy diagnostics are similar to prediction‐based ordinal stream condition studies performed previously within the study region (e.g., Maloney, Weller, Russell, & Hothorn, ) but slightly less than those from binomial stream condition studies both inside and outside the region (Hill et al, ; Maloney, Smith, et al, ). Overall, this suggests our model is suitable for use in both prediction of current conditions of unsampled areas and under future land‐use and climate conditions.…”

“…Our prediction of 66.1% of small stream reaches in the Chesapeake Bay watershed currently in Fair or Good condition is slightly higher than the 64.0% predicted in Maloney, Smith, et al (). Close agreement was surprising given the latter used a finer base layer (1:24,000), different covariates, and Chessie BIBI data from 2004 to 2008.…”

“…We used the family‐level, bioregion‐scale Chessie BIBI (Maloney, Smith, Buchanan, Nagel, & Young, ; Smith et al, ). We chose this version because it (a) optimized the trade‐off between taxonomic resolution and sample size, (b) adjusted for natural effects of geology, topography, and biogeography (i.e., bioregion), and (c) selected by stakeholders as their preferred stream health indicator (Buchanan, Maloney, Smith, Nagel, & Young, ).…”

“…We used data from StreamCat (Hill, Weber, Leibowitz, Olsen, & Thornbrugh, ), which contains 517 metrics representing both natural and anthropogenic landscape information summarized using NHDPlusV2 at local, cumulative upstream and riparian scales. We selected 15 uncorrelated ( r < 0.70) StreamCat metrics that were identified in literature as important surrogates of instream drivers of stream condition as they relate to habitat for benthic macroinvertebrates, including: Upstream cumulative watershed area because it is strongly related to many stream variables, such as discharge, energy process, and biological communities (Vannote, Minshall, Cummins, Sedell, & Cushing, ); Elevation since it was important in previous modeling efforts of stream condition in the study area and because it correlates with slope and instream temperature (Maloney, Smith, et al, ); Seven soil predictors (mean season water depth (cm), mean organic matter content (% by weight), mean permeability (cm/hr), mean depth (cm) to bedrock, mean percent clay content, mean percent sand content, and mean soil erodibility (Kf) factor) because of the importance of soils and resultant drivers (e.g., sediment) on stream macroinvertebrates (Waters, ); Three measures of geochemical content in surface or near surface geology—mean percent of lithological calcium oxide (CaO) because of its high correlation with many stream chemistry variables, mean lithological hydraulic conductivity (micrometers per second) because of its influence on rock/waters interaction, and mean lithological uniaxial compressive strength as a measure of susceptibility to weathering (megaPascals; Olson & Hawkins, ) all of which affect local habitat for macroinvertebrates; Summaries of mean runoff (mm) and baseflow index because of the importance of hydrology to streams (Poff et al, ); and Mean composite topographic index (topographic wetness index), which relates upslope area to local slope and is used to quantify topographic control on hydrological processes and estimate water accumulation—that is, valley bottoms have a high index whereas ridge and crests have a low index (Beven & Kirkby, ; Sörensen, Zinko, & Seibert, ). We hypothesized streams near or surrounded by ridges and crests would have less anthropogenic stress due to less accessibility or suitability, thus improving stream condition. …”

“…Random forests are well suited for predicting stream condition scores because they can use continuous and categorical predictors, often perform exceptionally when compared with other methods (Fernández‐Delgado, Cernadas, Barro, & Amorim, ), are relatively insensitive to correlated covariates, and account for nonlinear relationships and complex interactions among predictors (Cutler et al, ). In total, 36 predictors were used (Table S1), including cumulative upstream watershed values for 15 natural predictors from StreamCat (Hill et al, ), 12 LULC categories representing baseline 2005 landscape conditions, seasonal average air temperature and total precipitation from PRISM, and dominant bioregion upstream of each stream reach given its importance in previous modeling efforts (Maloney, Smith, et al, ). The model was developed using the randomForest R package (Liaw & Wierner, ) with 1,000 trees and 19 variables randomly sampled at each split, which was optimal during tuning.…”