Chapter 11: Probability of Streamflow Permanence (PROSPER) Model

Dataset Overview

The PRObability of Streamflow PERmanence (PROSPER) model is a geospatial model that predicts streamflow permanence for unregulated, minimally impaired streams throughout the Pacific Northwest of the United States at 30-m resolution, corresponding to the stream network represented by the medium-resolution National Hydrography Dataset (U.S. Environmental Protection Agency and U.S. Geological Survey, 2012). The PROSPER model was constructed using stream observations (wet/dry) and a suite of climate and physiographic variables to predict streamflow permanence. For each 30-m pixel of a stream, the model provides streamflow permanence probability at annual time steps from 2004 through 2016 (figure 11.1; Jaeger et al., 2018; Sando and Hockman-Wert 2019). Stream-network pixels were then categorized as wet (remaining wet or with flowing water throughout the year) or dry (lacking water at some point in the year) for each year using the streamflow permanence probabilities.

Data Access:

Figure 11.1


Conservation Applications

Potential conservation applications of this dataset could include the following:

  • Outputs from the PROSPER model could be combined with local knowledge of stream networks in an area to assess vulnerability of streams to summer drying under a range of climate conditions, i.e., in wet years and dry years. Because streamflow permanence (whether or not flow is maintained throughout the year) can be an important influence on riverine ecosystems, including habitat for some species of conservation concern, PROSPER-derived estimates of streamflow permanence probabilities could be used to guide assessments of riverine habitat, community dynamics, and species vulnerability to changing environmental conditions such as droughts and climate change. Additionally, flow-permanence regime (perennial, ephemeral, and intermittent) is a primary consideration in decisions related to the application of chemical herbicides and pesticides. PROSPER results can be used to better inform decisions about the location and quantity of herbicides and pesticides applied on the landscape.

Applicable scales for detailed spatial assessments:

  • For conservation applications requiring detailed assessment of spatial variation of the dataset within a geographic boundary (such as a protected area), the following geographic scales may be most appropriate (see appendix 3 for more information).
  • At the scale of: a state park or state wildlife area, a local watershed (12-digit hydrologic unit code [HUC-12]), a Bureau of Land Management (BLM) district, a river watershed (8-digit hydrologic unit code [HUC-8]), an individual county, a national forest, a level-3 ecoregion (e.g. the North Cascades), a single state (Washington, Oregon, or Idaho), a region (multiple states in the Pacific Northwest).

Applicable scales for assessing general patterns:

  • Due to spatial resolution, the dataset may not show detailed spatial variation at the following geographic scales, however, the dataset may be useful to assess general patterns or for comparison to other locations (see appendix 3 for more information).
  • At the scale of a small (< 1 km2) nature preserve.

Use of the dataset in conservation applications may be limited by the following considerations:

  • Actual streamflow permanence for a section of a river network may not be accurately described by the PROSPER model outputs as described in section 8 above. Thus, PROSPER outputs should be combined with local knowledge, other streamflow permanence datasets (e.g., National Hydrography Dataset Plus), and field observations for conservation planning purposes. Additionally, the PROSPER model outputs do not represent hydroperiod—the length of time a stream remains flowing—which can be critically important to riverine ecosystems and to individual species of conservation concern. Other important characteristics that shape riverine habitat include water temperature, water quality (e.g., dissolved oxygen, nutrients, suspended sediment), timing and magnitude of streamflow characteristics such as flooding, stream bed and bank characteristics, and characteristics of the riparian environment.

Past or current conservation applications:

  • The dataset has not yet been used in any on-the-ground conservation applications to the knowledge of the authors of this chapter.

Dataset citation:

Jaeger, K. L., R. Sando, R. R. Mcshane, J. B. Dunham, D. P. Hockman-wert, K. E. Kaiser, K. Hafen, J. C. Risley, and K. W. Blasch. 2018. Probability of Streamflow Permanence Model (PROSPER): A spatially continuous model of annual streamflow permanence throughout the Pacific Nortwest. Journal of Hydrology X 2:100005.

Sando, R., and D. P. Hockman-Wert. 2019. Probability of Streamflow Permanence (PROSPER) Model Output Layers (ver. 2.0, February 2019): U.S. Geological Survey data release.

Dataset documentation link: (open access) (open access)

Data access:

The dataset can be downloaded from:

The dataset is available for interactive viewing under the Exploration Tools in StreamStats:

Metadata access:

Formal metadata is available from:

Dataset corresponding authors:
Roy Sando
U.S. Geological Survey
Kristin Jaeger
U.S. Geological Survey

Kyle Blasch
U.S. Geological Survey

Jason Dunham
U.S. Geological Survey

Data type category (as defined in the Introduction to this guidebook):  Hydrology, stream and riparian

Species or ecosystems represented:  This dataset represents stream and river ecosystems. No individual species are represented.

Units of mapped values:

Streamflow permanence probabilities
Streamflow permanence classes

Range of mapped values:

0 to 1 (for streamflow permanence probabilities)
-5 to 5 (for streamflow permanence classes)

Spatial data type: a raster dataset (grid)

Data file format(s): GeoTiff (.tif), Esri Service Definition file

Spatial resolution: 30 m

Geographic coordinate system: North American Datum of 1983

Projected coordinate system: Albers Conic Equal Area

Spatial extent: Regional

Dataset truncation: The dataset is truncated at the border between the United States and Canada

Time period represented: Current or recent (2004 to 2016)

Methods overview:

The PROSPER model was constructed using over 3,800 stream observations, roughly half wet (with either flowing water or pools observed after July 1) and half dry, with no observable surface water. To predict streamflow permanence, the PROSPER model used a suite of climate and physical predictor variables, representing land cover, topography, soils, permeability, temperature, precipitation, snow-water equivalent, and evapotranspiration. For climate variables that change through time, monthly or annual values were used. For each 30-m pixel in the model, all variables were summarized for the entire upstream catchment. A random forest classification model was used to produce a streamflow permanence probability for every stream pixel in the study area for each year (2004-2016), along with confidence intervals. Finally, each pixel was categorized into a streamflow permanence class with an associated confidence level (wet or dry) using a locally optimized threshold developed to correct for regional bias. For more information, please consult the dataset citation listed in section 2 of this chapter.

This dataset relied on the following general types of models: Hydrologic models

This dataset employed the following specific models:  Random forest classification (Breiman 2001), Empirical Bayesian Kriging (Krivoruchko and Gribov 2014)

Major input data sources for this dataset included: Current land use, including protected areas, historical climate observations or models, digital elevation models (DEMs) or topography, soil characteristics, streamflow data or other hydrologic data.

The mapped values of the dataset may be interpreted as follows:

For each pixel in each year, high streamflow permanence probability (SPP) values (0.5 to 1) and streamflow permanence class (SPC) values (1 to 5) indicate high likelihood of streamflow permanence, i.e., streamflow continuing throughout the year. Low SPP values (0 to 0.5) and SPC values (-5 to -1) indicate high likelihood of that location of the stream becoming dry at some point that year. Agreement between SPP and SPC values indicate more reliable predictions. For example, a pixel with a predicted SPP value of 0.7 and SPC value of 5 is more reliable than a pixel with a predicted SPP value of 0.54 and SPC of -3.

Representations of key concepts in climate-change ecology:

Outputs from the PROSPER model relate to climate-change vulnerability in that they represent predicted streamflow permanence across a range of years with varying climate conditions. From the perspective of conserving aquatic habitat, streams that remain wet during relatively dry years—and for snow-dominated basins, in years with reduced snowpack or early snowmelt—may suggest greater potential resistance (reduced vulnerability) to future climate changes. Additionally, PROSPER outputs relate to climate sensitivity by estimating interannual variability in streamflow permanence that can be directly attributed to the fluctuations in climatic variables included in the PROSPER model.

Streamflow permanence or intermittency can have important implications for aquatic ecosystems. For example, ephemeral streams (those with intermittent flow) may have unique ecological characteristics and groups of species, determined largely by seasonal patterns of streamflow availability (Datry et al., 2017). Thus, categorizing streams as perennial or ephemeral can be an important step in species vulnerability assessments. Streamflow permanence can be ecologically important not only to aquatic communities, but also to terrestrial animals using streams as sources of water or food.

This dataset involves the following assumptions, simplifications, and caveats:

PROSPER is a regional-scale model applied to streams spanning a broad range of climate, geology, and topography. For some streams, local-scale controls related to soil and geology might be very important in determining streamflow permanence, but these local-scale factors are not represented in the PROSPER model or its outputs. The most important predictors in the PROSPER model were precipitation, forest cover, and temperature. As a result, streams with consistent year-round flow in arid regions would have lower predicted streamflow permanence probabilities than streams in wetter climates, suggesting underestimation of streamflow permanence for some perennial streams (e.g., spring-fed streams) in arid landscapes. Similarly, in very wet regions such as the coast range of Oregon, there were relatively few observations of dry streams to use as inputs to train the PROSPER model. Thus, the model may not adequately predict dry streams in this region. Figure A1B in Jaeger et al. (2018) shows standard error of prediction values related to classifying streams as wet or dry. Dataset users may want to use the wet and dry classifications with caution in areas that have high error rates.

In addition, the PROSPER model does not account for streamflow regulation through dams or diversions. Stream segments downstream from reservoirs likely have greater streamflow permanence than is predicted by PROSPER because their hydrology has been altered by humans. Conversely, streams with substantial water withdrawals would likely have lower streamflow permanence that is predicted by PROSPER.

Quantification of uncertainty:

Out-of-bag error rates from an internal cross-validation process in the random forest classification algorithm were used to assess the accuracy of the PROSPER model at the locations of the calibration data.

To translate predicted streamflow permanence probability (a continuous variable ranging from 0 to 1) to a binary streamflow permanence classification of wet versus dry, a threshold probability value was needed below which pixels would be categorized as dry, and above which they would be categorized as wet. This threshold varied throughout the study area and was accompanied by a standard error of prediction value, presented in figure A1 of Jaeger et al. (2018). Regions with relatively low standard error of prediction values are associated with greater confidence for the wet/dry stream classifications, whereas regions with high standard error of prediction values indicate less confidence in the wet/dry classification. Several watersheds were flagged with particularly high standard error rates suggesting users should exercise caution in interpreting wet/dry classifications in these regions.

Field verification:

Streamflow permanence probabilities from the PROSPER model were compared to flow classifications (flowing versus not flowing) from U.S. Geological Survey (USGS) gage locations across a range of climate conditions for time periods outside the modeling time period of PROSPER (2004-2016). In five of six climate classes (all except the wettest climate class), the streamflow permanence probabilities predicted by PROSPER were statistically significantly different between the flowing and non-flowing USGS stream gages, helping to validate the relative streamflow permanence probabilities within those climate classes. In the wettest climate class, flowing and non-flowing USGS gages had predicted streamflow permanence probabilities that were not statistically different, probably due to a relatively limited number of dry observations.

Prior to dataset publication, peer review was conducted by external review (at least two anonymous reviewers, each from a different institution).

Breiman, L. 2001. Random forests. Machine learning 45:5–32.

Datry, T., N. Bonada, and A.Boulton. 2017. Chapter 1 – General introduction. Pages 1-20 in T. Datry, N. Bonada, and A. Boulton, editors. Intermittent Rivers and Ephemeral Streams: Ecology and Management. Academic Press, London

Jaeger, K. L., R. Sando, R. R. Mcshane, J. B. Dunham, D. P. Hockman-wert, K. E. Kaiser, K. Hafen, J. C. Risley, and K. W. Blasch. 2018. Probability of Streamflow Permanence Model (PROSPER): A spatially continuous model of annual streamflow permanence throughout the Pacific Nortwest. Journal of Hydrology X 2:100005.

Krivoruchko, K., and A. Gribov. 2014. Pragmatic Bayesian kriging for non-stationary and moderately non-Gaussian data. Pages 61–64 in E. Pardo-Igúzquiza, C. Guardiola-Albert, J. Heredia, L. Moreno-Merino, J. Durán, and J. Vargas-Guzmán, editors. Mathematics of Planet Earth. Springer, Berlin.

Sando, R., and D. P. Hockman-Wert. 2019. Probability of Streamflow Permanence (PROSPER) Model Output Layers (ver. 2.0, February 2019): U.S. Geological Survey data release.

U.S. Environmental Protection Agency and U.S. Geological Survey. 2012. National Hydrography Dataset Plus – NHDPlus version 2.10. U.S. Environmental Protection Agency, Washington, DC.