Mapping the upper extent of fish in streams through modeling
Land-Water Connections
In western North America, the regulations governing the management of forests and streams highlight the desire to protect and utilize the ecosystem goods and services provided by both land and water resources. For example, forest harvest practices are regulated to protect fish, streams, and clean water which are valued socially and economically, as is the timber. Balancing the protection and production of these sometimes competing goods and services has contributed to a rich evolution of research, regulation, and management.
Photo of a small waterfall on Muletail Creek in the Nestucca River basin in the Oregon Coast Range that forms a barrier to fish passage, marking the upper extent of fish in a stream. The “last fish” would be found in the pool below the waterfall.
The uppermost fish in Muletail Creek in the Nestucca River basin in the Oregon Coast Range is found in the pool below the small falls.
Illustration of ecosystem services at the upper limit of fish. Forest-freshwater ecosystems jointly produce benefits from nature, also known as ecosystem services, such as carbon storage, greenhouse gases, and climate regulation, aquatic and terrestrial biodiversity and habitat, mitigation of soil erosion and floods, timber and nontimber products, and management of water quality and quantity. Ecosystem services are jointly produced, thus when forests are harvested, the trajectory of ecosystem services associated with forest-freshwater ecosystems may change. Riparian buffer regulations for forest harvests depend on the upper limit of fish. Forest management practices near the upper extent of fish affect the levels of co-produced ecosystem services associated with the riparian buffer.
Identifying the upper distribution of fish within a stream network is key to contemporary forest and fisheries management. Stream reaches with fish receive more protection than reaches without fish. These protections include harvest restrictions on lands adjacent to fish-bearing reaches and wider riparian buffers.Regulations around timber harvest and other forest management practices are designed to protect fish and their habitats while benefitting other species and protecting water quality.
A shared map that offers a visual context for where fish are, where they are not, and where their distributions end would help facilitate management of multiple resources and inform decisionmakers.
Header: Why focus on the upper extent of fish?
The upper distribution boundary for fish in forested streams receives special attention because fish-bearing portions of streams are managed differently and are more protected than fishless portions.
Two sliding images compare an illustration of fish distributions in a watershed where the image on the left has fish in the streams and consequently it also has additional protections, including wider riparian buffers. The image on the right does not have any fish in the streams shown in the watershed and accordingly there are narrower riparian buffers along the streams.
Examples of riparian forest buffers, western Oregon
Examples of riparian forest buffers, western Oregon.
Examples of riparian forest buffers, western Oregon.
Examples of riparian forest buffers, western Oregon.
Examples of riparian forest buffers, western Oregon.
Examples of riparian forest buffers, western Oregon.
Mapping the upper distribution limit of fish is complicated by land ownership, land use, and the limitations of survey data. Distribution maps often products of a mosaic of data gathered via different methods at different scales. In North America, fish distribution maps are maintained by multiple entities, including private companies, states/provinces, and federal land managers. The maps are populated with different kinds of information depending on their mission and objectives. A consistent, shared method that captures the complex ecological processes that contribute to predicting the upper extent of trout in streams would be useful for landowners, managers, and policymakers.
Map of landownership patterns in a segment of western Oregon. Mountainous areas are primarily owned and managed by the USDA Forest Service and USDI Bureau of Land Management, or are private industrial forests, with some areas of state forest lands scattered here and there. In some areas, a checkerboard pattern of federal and private provides challenges to management, as different rules for riparian buffer protections apply to public versus private lands. Lowland areas are most often in private non-industrial forest land or agricultural ownership, where rules are much less restrictive.
Land ownership patterns, western Oregon.
Illustration of riparian buffer protections on streams in western Oregon, USA. Buffer widths vary by landownership and management. The widest buffers (up to 160 m) apply to federal forest lands (USDA Forest Service and USDI Bureau of Land Management) under the Northwest Forest Plan, with fish-bearing streams given more protection. State forest plans specify somewhat narrower buffers, consistent among several categories of characteristics. Private industrial timber lands are regulated under the Oregon Forest Practices Administrative Rules. Buffer widths vary based on whether or not the stream is fish-bearing, contains salmon, steelhead, or bull trout, or provides domestic water, and on whether the streamflow is perennial or seasonal. Non-forest agricultural lands are covered by Agricultural Water Quality Management Plans, in which all protections are voluntary, none prescribed.
Example map of predictions by Fransen et al. (2006) from Panther Creek, Oregon showing a distribution of where fish are predicted and where they aren’t.
Example of predictions by Fransen et al. (2006) from Panther Creek, Oregon showing a distribution of where fish are predicted and where they aren’t.
Most fish distribution maps come from models based on occurrence information or habitat features. They may also include information from mechanistic, process-based, and correlative models. Distribution maps can also be based on direct observations, often from electrofishing, trapping, or snorkeling. However, data collected through direct observation can be labor intensive, rely on taxonomic expertise, and are influenced by both seasonal flow and the life cycle of the fish. Sampling every stream reach across a region is almost impossible. Slopes of 20 percent are recommended as the cutoff for the uppermost extent of fish across various western states and provinces of the United States and Canada. Another model that predicts the upper extent of trout is the optimal Fransen model (Fransen et al. 2006), which is a logistic regression model that was developed on stream layers derived from the National Hydrography Dataset (NHD) for 10-m reaches on private lands in western Washington.
Photograph of coastal cutthroat trout being netted in Mack Creek, McKenzie River the H.J. Andrews Experimental Forest.
Coastal cutthroat trout being netted in Mack Creek, McKenzie River at the H.J. Andrews Experimental Forest.
Oregon Department of Fish and Wildlife fish species distribution LEFT
Oregon Department of Forestry fish presence/absence RIGHT
Two sliding images compare fish species distribution data from the Oregon Department of Fish and Wildlife (left) and the Oregon Department of Forestry (right). The important takeaway is the upper extent of fish distribution in different sources does not always match up, so it matters which set is used for analysis or decision making.
National Hydrography Data (previous) LEFT
National Hydrography Data (new) RIGHT
Two sliding images compare the stream network from the National Hydrography Dataset, with previous data on the left and newer data on the right. Newer data based on better LiDAR imagery provides a more detailed and precise mapping, especially of small tributaries, that can greatly enhance the ability to model barriers that might block occupancy by fish.
Here, we focus on coastal cutthroat trout, a species iconic to the Oregon Coast Range. They are found in small streams at their upper extent as well as downstream into rivers, through the watershed, and into coastal habitats. In this work, we only focus on their distribution edge at the most upstream limit.
Map showing the geographic extent of Coastal Cutthroat Trout range, west of the Cascades and the British Columbia and Alaska coastal ranges, from coastal northern California to the Kenai Peninsula of Alaska.
Coastal cutthroat trout range.
Coastal cutthroat trout from the Willamette River basin, Oregon, in fish viewer tank.
Photo example from the field of actual upper extent of fish on streams in the Pacific Northwest: uppermost fish occurs below small waterfall and cascades.
Photo example from the field of actual upper extent of fish on streams in the Pacific Northwest: soft edge where fish drop out owing to lower streamflows.
Photo example from the field of actual upper extent of fish on streams in the Pacific Northwest: cascade blocks fish passage.
Photo example from the field of actual upper extent of fish on streams in the Pacific Northwest: dense cobbled substrate leaves too little waterflow for fish to swim past obstacles.
Photo example from the field of actual upper extent of fish on streams in the Pacific Northwest: bedrock slide blocks fish passage.
Examples of the upper extent of fish from across streams in the Pacific Northwest.
UPRLIMET
Acronym explained: Upstream Regional LiDAR Model for Extent of Trout
UPRLIMET is a spatially explicit and standardized model to predict the upper extent of fish across land ownerships in western Oregon. Using trout occurrence information and a stopping rule, we implemented UPRLIMET as a logistic regression model that uses geophysical aspects of the landscape, including stream size, slope, and elevation. We found this combination resulted in the lowest error.
Generalized development workflow for UPRLIMET, a single logistic regression model fit to trout occurrence observation data with a stopping rule.
It is impractical and inefficient to collect observations across the hundreds of thousands of kilometers of streams in western Oregon. Instead, we can use a few fish sightings and information associated with fish presence that is both less expensive to acquire and available at a broad spatial extent to calibrate a prediction model that allows us to infer where fish are likely to be. This is a complex and extensive process with numerous pitfalls, so we assembled a custom 4-step model development, validation, and prediction framework using R, Python, and ArcGIS software.
01 / 09
How does UPRLIMET work?
Photo of small waterfall below a log that marks the upper extent of fish from North Fork Ecola Creek, Coast Range. The uppermost fish would occur in the pool below the falls.
We electrofished in 103 streams across federal, state, and private lands in western Oregon to identify both the upper-most fish occurrence and habitat characteristics at or above the fish occurrence.
Map of field data showing sites sampled for this study in the Cascades and Coast Range of western Oregon. Right: photo of crew member with electrofishing gear surveying a stream.
It is impractical and inefficient to collect observations across the hundreds of thousands of km of streams in western Oregon. Instead, we can use a limited number of fish presence observations and information associated with fish presence that is both less expensive to acquire and available at a broad spatial extentnto calibrate a prediction model that allows us to infer where fish are likely to be present. This is complex and extensive process with numerous pitfalls, so we assembled a custom 4-step model development, validation, and prediction framework using R, Python, and ArcGIS software.
Illustration of a detailed workflow for the UPRLIMET model. Observation data, National Hydrography Dataset flowlines, LiDAR digital elevation model data, HUC 12 boundaries, and other covariate data feed the model. Step 1 compiles spatial data for model training. Step 2 is model development. In Step 3, a stopping rule (SR) is developed. In Step 4, model selection results in UPRLIMET, the best fit model, and model predictions. Details of the process can be found in Penaluna et al. 2022 Scientific Reports, Supplementary Information.
We used high-detail, fine-scale LIDAR-derived flowlines split into small reaches ranging from 5 to 7m in length. Length varies depending on whether the reach crosses orthogonally across the associated 5m x 5m spatial data raster grid cell at the location of the reach, or diagonally.
We compiled a database of big data composed of the 5 to 7m flowline hydrography attributed with 67 variables representing geologic, hydrotopographic, landcover, and climatic conditions for each of the 381 12-digit hydrologic unit code (HUC12) subwatersheds.
Field observations were propagated up and down the flowlines to generate a large set of model training data (~100,000 reaches) that captures a large range of conditions for each of the 67 candidate predictor variables.
Binomial classification models were fit to several combinations of the 67 candidate prediction variables and the propagated observation data with logistic regression. The most accurate model that we tested contained four variables related to hydrotopographic conditions on the landscape; it was selected as stage 1 of UPRLIMET. The four variables were upper stream length, upstream drainage area, slope, and elevation. The graphs to the right depict how the probability of trout presence changes across the distribution of values for a given variable, when all other variables are held constant.
The first stage of the UPRLIMET model was applied to flowlines divided into 5 to7m reaches having attribute values representing the four variables: upper stream length, upstream drainage area, slope, and elevation. The results are predicted probabilities (ranging from 0 to 100 percent) that any given reach belongs to the “trout” class.
We then applied the second stage of the UPRLIMET model, which is a stopping rule (SR1 from Penaluna et al. 2022) and serves to infer a discrete upper limit of trout locations from the predicted trout presence probabilities in stage 1. Once a discrete upper limit is identified, the upper limit of fish distribution can be estimated on a given stream network by labeling stream reaches as “trout” downstream of the upper limit location and “no trout” upstream of the location.
Coffee Creek, South Umpqua River fish distribution map produced by the UPRLIMET model. Predicted fish presence is most likely in higher-order mainstem reaches of each subwatershed, and unlikely in small headwater streams.
Ecola Creek (in the Oregon Coast Range) fish distribution map produced by the UPRLIMET model. Predicted fish presence is most likely in higher-order mainstem reaches of each subwatershed, and unlikely in small headwater streams.
Panther Creek, North Umpqua River fish distribution map produced by the UPRLIMET model. Predicted fish presence is most likely in higher-order mainstem reaches of each subwatershed, and unlikely in small headwater streams.
West Fork Smith River fish distribution map produced by the UPRLIMET model. basin. Predicted fish presence is most likely in higher-order mainstem reaches of each subwatershed, and unlikely in small headwater streams.
UPRLIMET predictions of fish distributions in four HUC12 sub-watersheds, including Coffee Creek (South Umpqua River), Ecola Creek (Coast Range), and Panther Creek (North Umpqua River), and West Fork Smith River (Umpqua River).
Key Findings
Our analysis shows that UPRLIMET distinguishes the upper distribution of fish better than all other models considered in this study, including random forest, logistic regression, and the optimal Fransen models, as evidenced by the lowest error estimates. Although some of the model combinations exhibited relatively small errors, those errors accumulated across tens of thousands of predicted upper limit points could potentially alter management decisions and outcomes along tens to thousands of kilometers of stream.
Chart of comparison among selected models ranked by mean absolute error (MAE; in meters) of linear distance between the observed upper limit and the predicted upper limit. For the top five models, the model description specifies the development algorithm [e.g., random forest (RF) or logistic regression (LR)], the stopping rule (SR) and its number (1, 2, or 3), and the type of training data used, with occurrence = O or habitat = H. In addition to showing the MAE for the top five models, two additional models are included, the Fransen et al. (2006) model, and a 20% slope cut off, where the lowest point on the network with a slope greater than or equal to 20% becomes the upper limit point. The model with the smallest MAE is called UPRLIMET.
Although there is no single, general explanation for distribution limits for coastal cutthroat trout, the intersection of stream size (upper stream length and upstream drainage area), slope, and elevation together locate the upper limit of fish.
Stream size (a; b) corresponds to major ecosystem changes along a stream continuum including energy sources, ecosystem metabolism, habitat characteristics, and biodiversity, as well as the upper distribution limit of fish as shown here. Stream size accounts for the top two variables in the model suggesting that it is the major driver of the upper distribution limit of fish with the probability of trout increasing with increasing upstream stream length and upstream drainage area.
Slope (c), the next variable of importance influencing the upper extent of fish, exerts control on physical habitats in streams, including channel morphology, hydraulics, sediment transport, substrate, and habitat.
Elevation (d) or vertical topographic position may indirectly integrate broad influences of other landscape-scale or climate factors or also indirectly capture stream size, influencing the likelihood of fish presence. The multiple factors associated with elevation correspond to the relationship found for stream size that smaller streams are less likely to have fish.
Partial-dependence profile plots of the four variables in UPRLIMET in relationship to the probability of trout presence, including (a) total upstream channel length in kilometers, (b) drainage area, log transformed, in square kilometers, (c) downstream channel slope over 1000 m (as percent), and (d) elevation, median-normalized, in meters. Plots are arrayed in decreasing order of model importance.
UPRLIMET predicted more fish on private lands than on state, U.S. Bureau of Land Management (BLM), or U.S. Forest Service (USFS) lands, highlighting the importance of using spatially consistent maps across a region and working across land ownerships.
Bar chart of the distribution of the length of streams (in kilometers) with trout, with no trout, and with no predictions, along with the number of upper limit datapoints (thousands) predicted by UPRLIMET across landownership categories of private non-industrial, private industrial, BLM (Bureau of Land Management), USFS (USDA Forest Service), state, other, and other federal lands. Stream length was estimated from the HUC12 scale. Note that streams without predictions occur when there is less than 1000 m of stream length over which to evaluate slope, or for channel-initiation reaches where upstream drainage area cannot be calculated. The greatest stream length predicted as having trout occurred on private non-industrial lands, followed by private industrial forest lands. Private industrial lands also had the greatest total stream length of the categories.
To provide a management-relevant understanding of UPRLIMET performance, we compared upper-limit predictions to four other sources of fish-distribution data, including (1) the Interagency Coastal Cutthroat Trout dataset; (2) the Oregon Department of Forestry (ODF) fish layer; (3) predictions from the optimal Fransen model; and (4) predictions of the upper limit based on the downstream-most presence of a 20 percent or greater slope over a 20-m run of stream. Differences among databases for the upper distribution limits of fish come from both the upper limit points and depiction of fish-bearing reaches, underscoring the importance of having a shared map with common coverage of the fish extent across landscapes and ownerships.
Ridge plots of frequency distributions of distances to the upper trout limit for UPRLIMET compared to trout occurrence and habitat distributions from Oregon Department of Forestry (ODF), trout occurrence distributions from the Interagency Coastal Cutthroat Trout (ICCT) database, Fransen et al. (2006) optimal model13, and 20% slope cutoff for western Oregon. X-axis is distance to UPRLIMET in kilometers, and y-axis is relative frequency. Positive numbers represent overestimation relative to UPRLIMET and negative numbers are an underestimation. Note that the previously used estimates for upper trout limit by Fransen et al. (2006) optimal, a 20% slope, and the ICCT database are biased toward underestimation and ODF overestimates.
Implications
UPRLIMET maps both the probability of trout and the upper limit of trout across multiple ownerships, including private, state, and federal lands. The resulting cross-boundary distribution map may be a useful tool for policymakers and forest managers.
Map of UPRLIMET predictions for the West Fork Smith River (Umpqua River basin) overlaid with landownership layer, an example of how predictions from this model can apply across ownership boundaries in a way that older models that relied on more restricted datasets have difficulty in doing. The checkerboard pattern of private and federal landholdings creates sharply different riparian protections in adjacent reaches of the same stream.
UPRLIMET crosses ownerships.
This work provides a transferable prediction modeling framework for systematically and comprehensively estimating the upper distribution limit of fish, which could be applied to watersheds and fish species around the globe.
Photo panel of five examples of the upper extent of fish in small streams on private lands in western Oregon. In each case, the uppermost fish would be found in the pool or reach below the barrier, whether that barrier is a cascade, a small waterfall, a very high-gradient reach, or an area of insufficient depth or other obstructions.
Photo panel of five examples of the upper extent of fish in small streams on private lands in western Oregon. In each case, the uppermost fish would be found in the pool or reach below the barrier, whether that barrier is a cascade, a small waterfall, a very high-gradient reach, or an area of insufficient depth or other obstructions.
Photo panel of five examples of the upper extent of fish in small streams on private lands in western Oregon. In each case, the uppermost fish would be found in the pool or reach below the barrier, whether that barrier is a cascade, a small waterfall, a very high-gradient reach, or an area of insufficient depth or other obstructions.
Photo panel of five examples of the upper extent of fish in small streams on private lands in western Oregon. In each case, the uppermost fish would be found in the pool or reach below the barrier, whether that barrier is a cascade, a small waterfall, a very high-gradient reach, or an area of insufficient depth or other obstructions.
Photo panel of five examples of the upper extent of fish in small streams on private lands in western Oregon. In each case, the uppermost fish would be found in the pool or reach below the barrier, whether that barrier is a cascade, a small waterfall, a very high-gradient reach, or an area of insufficient depth or other obstructions.
Examples of the upper extent of fish from across streams in the Pacific Northwest.
Look for the upper limit of fish in your favorite western Oregon watershed!
Authors
Brooke E. Penaluna 1 , Jonathan D. Burnett 1 , Kelly Christiansen 1 , Ivan Arismendi 2 , Sherri L. Johnson 1 , Kitty Griswold 3 , Brett Holycross 4 , and Sonja H. Kolstoe 5
1 U.S. Department of Agriculture, Forest Service, Pacific Northwest Research Station, 3200 SW Jefferson Way, Corvallis, OR 97331
2 Oregon State University, Department of Fisheries, Wildlife, and Conservation Sciences, 104 Nash Hall, Corvallis, OR 97331
3 Idaho State University, Department of Biological Sciences, 921 S. 8th Ave Mail, Stop 8007 | Pocatello, ID 83209-8007
4 Pacific States Marine Fisheries Commission, 205 SE Spokane St., Portland, OR 97202
5 U.S. Department of Agriculture, Forest Service, Pacific Northwest Research Station, 1220 SW 3rd Ave, Suite 1410, Portland, OR 97204
Further Reading
Fransen, B.R.; Duke, S.D.; McWethy, L.G.; Walter, J.K.; Bilby, R.E. 2006. A logistic regression model for predicting the upstream extent of fish occurrence based on geographical information systems data. North American Journal of Fisheries Management. 26: 960–975. https://doi.org/10.1577/M04-187.1.
Penaluna, B.E.; Burnett, J.D.; Christiansen, K.; et al. 2022. UPRLIMET: UPstream Regional LiDAR Model for Extent of Trout in stream networks. Scientific Reports. 12:20266. https://doi.org/10.1038/s41598-022-23754-0.
The uppermost fish in Muletail Creek in the Nestucca River basin in the Oregon Coast Range is found in the pool below the small falls.
Example of predictions by Fransen et al. (2006) from Panther Creek, Oregon showing a distribution of where fish are predicted and where they aren’t.
Coastal cutthroat trout being netted in Mack Creek, McKenzie River at the H.J. Andrews Experimental Forest.
