Model Metadata
Overview of specific conductance, total nitrogen, and total phosphorous predictive models and reference criteria
For the water quality indicators lacking state standards, we utilized predictive models to establish predicted natural conditions (PNC) in the absence of anthropogenic impacts. PNC use empirical models based on geospatial predictors to understand the spatial variability among reference sites for a given indicator. Such models account for natural environmental gradients and are used to make predictions of chemical, physical, or biological values expected at a site in the absence of anthropogenic impairment (i.e., site potential). Condition is then determined based on the deviation of the observed indicator value from the site-specific predicted value. If this deviation is greater than specified percentiles of model error (e.g., 95th), the value is assigned a condition of ‘not meeting’ a given water quality benchmark. Predictive modeling approaches are advantageous because they result in site-specific predictions, take into account natural environmental gradients, and have known levels of accuracy and precision.
Specific conductance (SC) benchmarks were established using a revised method of Olson and Hawkins (2012, 2013). The SC model uses 10 StreamCat-derived predictor variables (Hill et al. 2016; Table 1) to explain 74% of the spatial variability in baseflow specific conductance concentrations (root-mean-square error 81.1 µS/cm) among 1912 reference sites throughout the contiguous western U.S. Benchmarks were then established by taking the site-specific predicted natural conditions from the model and adding the 75% and 95% of model error, 24.4 µS/cm and 106.7 µS/cm, respectively, to the prediction.
Total nitrogen (TN) and total phosphorus (TP) benchmarks were established using the methods in Olson and Hawkins (2013). The TN model uses 6 StreamCat-derived variables (Table 1) and the day of year (DOY) on which the sample was collected to explain 39% of the spatial variability in baseflow TN values (root-mean-square error 120.3 µg/L). The TP model uses 9 StreamCat-derived variables (Table 1) to explain 38% of the spatial variability in baseflow TP values (root-mean-square error 19.4 µg/L) among reference sites throughout the contiguous western U.S. Benchmarks for TN and TP were established similar to specific conductance. The 75% and 95% of model error are 25.1 µg/L and 217.7 µg/L, respectively, for TN. The 75% and 95% of model error are 6.0 µg/L and 37.5 µg/L, respectively, for TP.
The reference site networks (Figure 1) were derived from Olson and Hawkins (2012, 2013) and Olson and Cormier (2019). Observations of SC and concentrations of TN and TP were taken directly from Olson and Hawkins (2012, 2013) or downloaded from the National Water Quality Monitoring Council Water Quality Portal (WQP; https://www.waterqualitydata.us). Sampled sites were identified as being in reference quality by the original collection agency and confirmed by Olson and Hawkins (2012, 2013) and Olson and Cormier (2019) following a two-step process. First, field-based physical habitat and water quality data for the sampled sites were used to screen data for anomalous water quality values. Second, google earth and USGS quad maps were used to screen sites for any evidence of human impacts (e.g., ranches, mines, agriculture, clearcuts).
The models outlined in this document are revised versions of previously used SC, TN, and TP predictive models. Models were revised to include a greater number of reference sites, be easier to compute, and to alleviate potential shifting baseline issues. Each of the revised models contain a greater number of reference sites than the previously used models (Table 1). In addition, the revised models use easily obtainable StreamCat predictor variables as input, reducing the need for complex GIS computations. The revised models should also avoid serious shifting baseline issues. A shifting baseline can occur when expected (reference) conditions are adjusted for the effects of a human-caused factor rather than anchored at a standard period in the past. Climate is a good example. Some water chemistry models use non-stationary climate variables to predict stream chemistry under current climatic conditions. These models can provide more precise estimates, which may be useful in some situations (for example to parse effects of land use from climate effects), but routinely incorporating current climate conditions into models to predict reference conditions can create a potentially serious shifting baseline issues. The revised models use stationary climate predictors from StreamCat (e.g., PRECIP8110WS, TMAX8110WS; see Table 1 for description) and do not include predictors related to land cover which can change over time.
Performance of the revised models is similar to that of the former models (Table 2). The revised SC and TP models perform very similarly to the former SC and TP models, whereas the revised TN model performs noticeably better than the former TN model.
Table 1
List of predictor variables included in each water quality indicator model.
| Indicator | Predictor Name | Description |
|---|---|---|
| SC | PRECIP_MINUS_EVTWS | Precipitation minus evapotranspiration is the net flux of water from the atmosphere to the earth’s surface |
| SC | TMAX8110WS | PRISM climate data - 30-year normal maximum temperature (°C): Annual period: 1981-2010 within the watershed |
| SC | TMIN8110WS | PRISM climate data - 30-year normal minimum temperature (°C): Annual period: 1981-2010 within the watershed |
| SC | AL2O3WS | Mean % of lithological aluminum oxide (Al2O3) content in surface or near surface geology within watershed |
| SC | CAOWS | Mean % of lithological calcium oxide (CaO) content in surface or near surface geology within watershed |
| SC | SWS | Mean % of lithological sulfur (S) content in surface or near surface geology within watershed |
| SC | RUNOFFWS | Mean runoff (mm) within watershed |
| SC | BFIWS | Baseflow is the component of streamflow that can be attributed to ground-water discharge into streams. The baseflow index (BFI) is the ratio of baseflow to total flow, expressed as a percentage, within watershed. |
| SC | ELEVWS | Mean watershed elevation (m) |
| SC | NWS | Mean % of lithological nitrogen (N) content in surface or near surface geology within watershed |
| TN | PRECIP_MINUS_EVTWS | Precipitation minus evapotranspiration is the net flux of water from the atmosphere to the earth’s surface |
| TN | PRECIP8110WS | PRISM climate data - 30-year normal mean precipitation (mm): Annual period: 1981-2010 within the watershed |
| TN | RUNOFFWS | Mean runoff (mm) within watershed |
| TN | ELEVWS | Mean watershed elevation (m) |
| TN | BFIWS | Baseflow is the component of streamflow that can be attributed to ground-water discharge into streams. The baseflow index (BFI) is the ratio of baseflow to total flow, expressed as a percentage, within watershed. |
| TN | PERMWS | Mean permeability (cm/hour) of soils (STATSGO) within watershed |
| TN | DOY | Day of year on which sample was collected |
| TP | PRECIP_MINUS_EVTWS | Precipitation minus evapotranspiration is the net flux of water from the atmosphere to the earth’s surface |
| TP | PRECIP8110WS | PRISM climate data - 30-year normal mean precipitation (mm): Annual period: 1981-2010 within the watershed |
| TP | TMAX8110WS | PRISM climate data - 30-year normal maximum temperature (°C): Annual period: 1981-2010 within the watershed |
| TP | RUNOFFWS | Mean runoff (mm) within watershed |
| TP | CLAYWS | Mean % clay content of soils (STATSGO) within watershed |
| TP | P2O5WS | Mean % of lithological phosphorous oxide (P2O5) content in surface or near surface geology within watershed |
| TP | WETINDEXWS | Mean Composite Topographic Index (CTI) [Wetness Index] within watershed |
| TP | CAOWS | Mean % of lithological calcium oxide (CaO) content in surface or near surface geology within watershed |
| TP | SANDWS | Mean % sand content of soils (STATSGO) within watershed |
Table 2
Metrics of model performance for the former and revised water quality models. Metrics are based on the relationship between observed and predicted values of reference sites.
| Indicator | n | R2 | RMSE | No. Predictors |
|---|---|---|---|---|
| Former Models | - | - | - | - |
| Specific Conductivity | 1390 | 0.78 | 67.3 | 19 |
| Total Nitrogen | 665 | 0.32 | 113.9 | 9 |
| Total Phosphorus | 752 | 0.40 | 20.5 | 15 |
| Revised Models | - | - | - | - |
| Specific Conductivity | 1912 | 0.74 | 81.1 | 10 |
| Total Nitrogen | 699 | 0.39 | 120.3 | 7 |
| Total Phosphorus | 966 | 0.38 | 19.4 | 9 |
References
- Hill, R.A., Weber, M.H., Leibowitz, S.G., Olsen, A.R., and D.J. Thornbrugh. 2016. The Stream- Catchment (StreamCat) dataset: a database of watershed metrics for the conterminous United States. Journal of the American Water Resources Association 52:120-128.
- Olson, J.R. and S.M. Cormier. 2019. Modeling spatial and temporal variation in natural background specific conductivity. Environmental Science and Technology 53:4316-4325.
- Olson, J.R. and C.P. Hawkins. 2012. Predicting natural base-flow stream water chemistry in the western United States. Water Resources Research 48, W01504, doi:10.1029/2011WR011088.
- Olson, J.R. and C.P. Hawkins. 2013. Developing site-specific nutrient criteria from empirical models. Freshwater Science 32:719-740.