Remote sensing framework details riverscape connectivity fragmentation and fish passability in a forested landscape

Abstract

Fragmentation of stream networks from anthropogenic structures such as road culverts can affect the health of a catchment by negatively affecting the ecosystem’s biota, their movements, abundance, and species richness. We present a framework using publicly available LiDAR and orthophotography to locate and identify road crossings, i.e. the most prolific of barriers in forested landscapes, and evaluate fragmentation and passability at the landscape scale. Coupling the LiDAR stream network and private road network in the 3,223 km² study area, we identified 1,052 stream crossings of which, 32% were culverts and 12% of the total stream network was potentially inaccessible due to these culverts. We correctly identified the type of stream-road crossings at >90% of any stream order and at 100% at Orders >2. The 10 culverts restricting the most stream kilometers, restricted >34% of the potential stream habitats for four species of fish, a result that provides the resource management with a first assessment for effective improvement of connectivity across this landscape. With this framework, managers equipped with appropriate imagery can create a stream crossing database with minimal funding, create an inventory of instream barriers, and prioritize removals at a landscape-scale, thus providing an effective assessment and decision-making tool for their habitat restoration efforts.

Keywords:

• Culvert
• LiDAR
• remote sensing
• stream connectivity

1. Introduction

Anthropogenic, in-stream structures such as dams, weirs, and road-crossing culverts are widespread barriers that can cause significant fragmentation of aquatic ecosystems (Trombulak and Frissell 2000; Anderson et al. 2012). Effects of fragmentation are an important issue for managers and planners because in-stream barriers can impact ecosystem structure and function, i.e. the stream ecosystem’s goods and services (e.g. Mahlum et al. 2014; Torterotot et al. 2014; Erkinaro et al. 2017). The most common anthropogenic barrier causing fragmentation of river networks in forested landscapes are road crossings (Khan and Colbo 2008; Park et al. 2008). Many road crossings use culverts to facilitate the passage of water through road embankments (Clay 1995; Erkinaro et al. 2017). Culverts are hydraulic structures that carry water under roadways or other embankments (Doehring et al. 2011; Hall et al. 2011). Pipe-culverts are the most common in-stream structure used at stream crossings due to their low costs (Khan and Colbo 2008; Park et al. 2008). They can be variable in size, e.g. as small as 0.3 m in diameter, and are typically installed for the singular purpose of transporting water through stream crossings by roads and thus rarely consider fish passage (Blakely et al. 2006; Makrakis et al. 2012). While there are guidelines intended to reduce environmental impacts on fish (e.g. Chilibeck 1992; New Brunswick Department of Environment (NBDE)) 2012), culverts alter stream morphology at the installation location and often increasingly so as a function of time, thus affect fish passability (Poplar-Jeffers et al. 2009; Price et al. 2010).

Culverts are not natural pathways and can alter fish behaviour and movement patterns (Fahrig 2003; Gibson et al. 2005). Potential passability is further compromised when culverts are poorly installed or not maintained. More broadly, problem culverts can disrupt aquatic ecosystem connectivity or the ‘exchange pathway of matter, energy, and organisms’ (Ward and Stanford 1995). The fragmentation of natural waterways can create habitat patches (e.g. Frissel et al. 1986; Cote et al. 2009) that can threaten habitat availability, biodiversity, and abundance (e.g. Khan and Colbo 2008; Nislow et al. 2011). Problem culverts can be repaired or removed to re-establish connectivity, but often there is competition for limited funds and culvert repairs are expensive; thus, culvert remediation becomes a low management priority (Gibson et al. 2005; Poplar-Jeffers et al. 2009). There is a growing desire to identify and establish a priority for remediating non-functional culverts to improve ecological connectivity across catchments (Januchowski-Hartley et al. 2014; Erkinaro et al. 2017).

Previous studies have demonstrated that the slope of the culvert and the elevation drop (i.e. a hanging culvert) at the culvert influence fish passability at stream crossings (e.g. Peake 2008; Bourne et al. 2011). With adequate funding and time, managers and planners could assess and evaluate the passability of each stream crossing (Bowen et al. 2006); however, this is not a feasible for landscape-scale planning, e.g. forest management, where thousands of existing stream crossings need to be evaluated (Kemp and O'Hanley 2010; Bourne et al. 2011). Fortunately, potential assessment parameters are now measurable with remotely sensed data. Januchowski-Hartley et al. (2014) demonstrated that statistical models can be used at the landscape-scale to predict passability of culverts using boosted regression trees if fine-scale data exists. Diebel (2014) demonstrated that culvert slope can be extracted from Light Detection and Ranging (LiDAR) digital elevation models (DEM). Doehring et al. (2011) illustrated that slope (also referred to as gradient) was a primary factor affecting passability through culverts for juvenile fish, i.e. slope has a positive relationship with water velocity in the culvert (Bouska and Paukert 2010).

Morphological and physiological characteristics of the fish can limit passability for each species based on swim speed, water velocity, and the length of time the fish can maintain that speed (Olsen and Tullis 2013; Khodier and Tullis 2018). Additionally, culverts with higher slopes can create “hanging” culverts at their downstream outlet that are a barrier if the fish cannot jump from the stream into the culvert (Taylor 2000; Burford et al. 2009). There have been several studies outlining frameworks to measure fish passability for culverts (e.g. Rayamajhi et al. 2012; Khodier and Tullis 2018), including mathematical models (Kraft et al. 2019), LiDAR technology (Diebel 2014), and simulation software (Bourne et al. 2011). These emerging technologies and interpretations demonstrate the promise of measuring passability of culverts and therefore stream connectivity remotely. The challenge remains moving from a fine-scale to landscape-scale, while sustaining cost effectiveness (Kemp and O'Hanley 2010).

Problem culverts can fragment stream connectivity, especially in lower order streams (Keller et al. 2011). Dependent on climatic, geologic, and topographic settings, lower order streams can provide critical flow and thermal refugia during extreme temperatures, such as summer highs and winter lows (French et al. 2017; Wilbur et al. 2020). As the climate warms in eastern Canada, access to such flow and thermally resilient streams will become critical to the survival of stenotherms, such as Atlantic Salmon (Salmo salar) and Brook Trout (Salvelinus fontinalis). Targeted connectivity remediation is required to ensure prioritization is given to connect the most resilient and therefore critical stream networks. The first step towards effect, and efficient, remediation is the identification of problem areas, at the landscape-scale. The goal of this research was to develop a framework and practical GIS tool to assess stream network connectivity and fish passability across large landscapes using remote-sensing data, specifically LiDAR generated digital elevation models and orthophotography. We also sought to examine the accuracy of publicly available road network data compared to a high-resolution private road network. We selected the most resolute stream connectivity model and then used the framework to assess passability at the landscape-scale for four common stream fishes based on criteria in existing literature and produced a count of potential barriers for each species, and the amount of cumulative stream kilometers that these barriers restricted. Finally, we produce an assessment of cumulative effects of barriers removed to demonstrate how the framework and tool could be applied by managers and planners.

2. Methods

2.1. Study area

The Restigouche River has a total drainage area of ∼13,000 km², 51% of which is located in New Brunswick and 49% in Quebec, Canada (Figure 1). The study area encompassed ∼ 3,200 km² within New Brunswick. The catchment lies in the Atlantic Maritime Ecozone and is underlain by calcareous bedrock and blanketed with glacially reworked surficial deposits (Rampton et al. 1984; Fyffe and Richard 2007). The catchment’s geomorphology is characterized by steep valleys and is comprised of large cobble substrate near the valley walls, with siltstone substrate on the valley floor (Amiro 1983). Forest cover makes up 96% of land use within the Restigouche River Watershed (Restigouche River Watershed Management Council (RRWMC)) 2015).

Figure 1. LiDAR range and sample area of the Eastern New Brunswick portion of the Restigouche catchment, New Brunswick, Canada.

2.2. Workflow

The framework developed for this study follows six steps. In steps 1 and 2, a synthetic stream network is derived using a1 m resolution LiDAR derived digital elevation model (DEM) and intersecting the stream network with a road network data set, respectively. In steps 3 and 4, culverts are visually identified and classified using high resolution (1 m) 3-band (Red-Green-Blue) aerial image, and elevation is extracted from the DEM upstream and downstream of the culvert, respectively. In steps 5 and 6, slope and length are calculated at the site, and the site is tested for species specific passability. An illustration of the workflow is provided in Figure 2.

Figure 2. The 6-step framework for (1) generating data, where FDR is the flow direction raster; (2) determine stream crossing locations; (3) classifying the crossing; (4) extracting elevation data from LiDAR DEM; (5) calculating slope and length of culverts; (6) export slopes to determine fish passability.

2.3. Hydrographic network delineation

LiDAR DEM data were obtained from the Province of New Brunswick’s online data (Service New Brunswick-, SNB 2017) for 3,223 km² of the catchment acquired in the summer of 2016. The LiDAR mission produced a data set with a point cloud density of 6 points per m² with the root mean square error (RMSE) = 0.06 m and the vertical error was 0.12 m at 12.4 cm at 95% accuracy (Service New Brunswick (SNB) 2016–2018).

The 1 m LiDAR DEM was utilized to simulate surface water flow through the landscape (Lindsay and Dhun 2015). A known limitation with surface flow modeling is closed depressions, that is, cells within the DEM that do not have an outlet, which can affect the delineation of flow paths (Hayashi et al. 2003). Several methods have been developed to remedy these limitations (Tarboton 2015; Jackson 2013), of which we use a selective breaching method to reduce the number of false depressions in the DEM (Wall et al. 2015), i.e. this allows flow to pass through embankments by lowering the cells to the stream elevation (Wall et al. 2015). This method involves locating all stream-road crossings and lowering the elevation of the cells of the embankment surrounding the crossing to allow the simulated flow to pass through the road embankment, thus preventing a false, closed depression (Wall et al. 2015). The result was hydrologically corrected DEM that has all true and artificially derived flow paths with no closed depressions from man-made topography (Wall et al. 2015).

To extract the stream network, a flow accumulation threshold was to be set in the DEM to determine the required number of accumulated hillslope pixels to initiate a channel, i.e. how much of the (upstream) drainage area is needed before a channelized stream is predicted to initiate (O'Callaghan and Mark 1984). In this study, a flow initiation threshold of 40 ha was chosen. The threshold was determined by visually comparing different thresholds to the National Hydrographic Network and by examining boundaries in relation to permanent streams (Service New Brunswick (SNB) 2016–2018). At 40 ha, we assume the synthetic stream network can account for ephemeral streams. Our selective breaching model applied the “Hydrocutter” flow path model (Wall et al. 2015). The Hydrocutter model is a toolbox that can remove false detections of closed depressions that are created by artificial dams within the DEM. The toolbox uses the intersections of stream and road crossings to locate where the water should flow and lower the elevation cells to allow the surface flow across the embankment.

We used the methods above to create our own unique stream network that measured 4,122 km. The stream network provided by Service New Brunswick (SNB) follows a similar protocol as the method stated above (SNB 2016–2018). In addition, SNB has a publicly available road network. The SNB stream network and road network are 3867 km and 3673 km, respectively. Data from J.D. Irving, Limited was a combination of logging and private roads totaling 6,035 km. The data type and sources are summarized in Table 1.

Table 1. Overview of variables used for stream crossing framework, data type, and source.

Variable	Data type	Source
LiDAR DEM	Raster	Service New Brunswick
Provincial road network	Feature	Service New Brunswick
J.D. Irving, Limited road network	Feature	J.D. Irving, Limited
Provincial stream network	Feature	Service New Brunswick
Modeled stream network	Raster Derived (Flow accumulation threshold 40 ha)	DEM Hydrocutter (Wall et al. 2015)

2.4. Culvert identification

Roads can generally be identified with orthophotography, and where not possible, the road was identified on the DEM by a steep elevation gain, followed by no elevation change, then returning to a steep elevation decrease, which would be the embankments of the road (Figure 3). All stream crossings were identified by the result from intersecting the stream and road network, then examined and verified visually on the DEM. Culverts were identified by visually observing the stream channel becoming more restricted and passing the road on the DEM (Figure 3). If the culvert is large enough, it was apparent on the orthophotography, and some culvert ends were visible on the DEM (Figure 3). Embankments were absent for bridges and there was no channel restriction. Various combination of these layers can help the user to identify where the road crossed a stream, i.e. the layers are interchanged to represent the crossing and emphasize topographical features. For example, if a crossing was thought to be a culvert, the 1-m resolution DEM with the hillshade was overlaid. This would make the channel and the embankment of the road more apparent. As per Diebel (2014), culverts or other crossings were identified directly from the 1 m orthophotography layer.

Figure 3. Digitized culvert using LiDAR crossing model overlaid with a hillshade DEM (a), and orthophoto reference showing digitized culvert (b).

For each stream crossing identified as a culvert, the culvert was digitized manually. This was accomplished by locating the lowest elevation of the up-and downstream ends of the culvert on the LiDAR DEM within 5 pixels (5 meters) of the road embankment. Once points were created at the lowest point on each side of the crossing (up and downstream of the road embankment) the culvert ends can be digitized, and thus, giving the distance (m) or culvert length and elevations above mean sea level (MSL) or culvert slope automatically. A positive slope was assumed to represent the water flowing from the upstream, down through the culvert (Wilson and Gallant 2000). To establish a digital pathway between the upstream and downstream points, we used ‘model builder’ (ESRI, 2018) which finds the nearest points with the same unique identifier and produces a table pairing these two points. The model then creates points from the table and connects them to generate a predicted (digitized) culvert. Length and elevations of the paired points were identified and used in a raster calculator to determine the slope of each culvert.

Known waterfalls were added to the data set and we considered these to be impassable (see Discussion for an assessment of GIS tools for waterfall detection). We attempted to identify natural barriers, specifically detecting waterfalls in the orthophotography and DEM, based on knickzones. A knickzone is a large change in elevation along a longitudinal gradient (Zahra et al. 2017). Zahra et al. (2017) created a tool to extract knickzones from DEM’s. The tool was applied on a 10 m DEM and a curvature threshold value = 2 based on the terrain shape with moderate relief.

2.5. Field validation

Culvert location predictions were validated against field-verified measurements using n = 242 randomly selected, stream crossings that were proportionally distributed amongst Strahler stream orders 1 to 6 ensuring a minimum 10% of sites within each stream order. Each crossing was classified as a culvert (regardless of culvert type or design), a ford (stream flow traveling across a forest road), or a bridge. A false detection was a model prediction of a crossing that was either not present or where there was no evidence of waterflow or a road-crossing structure in the field. Sites where a culvert was present but no waterflow, was observed were classified as a “culvert, no channel” (i.e. potential ephemeral stream)

At each site, GPS coordinates were taken at the center of the road at the stream crossing using a Garmin GPSMAP 78 s (accuracy of 3.6 m). The slope was measured with a Zip Level which measured the elevation change between the upstream and downstream ends of the culvert with an accuracy of 0.01 mm (Technidea 2012). For each culvert examined in the field, the length and slope of the culvert predicted by our model was compared to the actual observed length and slope were compared using a paired t-test on the data (α = 0.05).

2.6. Species passability

We selected four common stream fishes to examine passability based on available information in the literature: Burbot (Lota lota); Lake Chub (Couesius plumbeus); Atlantic Salmon (Salmo salar), and Brook Trout (Salvelinus fontinalis). Passable conditions for each of the fish species were defined by slope thresholds from previous studies, see MacPherson et al. (2012) – Burbot and Lake Chub; Bourne et al. (2011) – Atlantic Salmon; Burford et al. (2009) – Brook Trout. As an additional test of passability we utilized elevation drops, or the threshold for entrance into a culvert (see Diebel 2014) and considered a barrier existed when drops exceeded 0.3 m. To assess the potential expansion of fish habitat with the removal of a potential barrier, we assumed that making an “impassable” culvert passable, i.e. removing it as a “barrier”, resulted in potential habitat expansion (measured in stream length – km). This assumption does not consider the species-specific suitability of habitats upstream and assesses only access to the stream corridor per se, but it provides a first step towards detailed assessments that managers can develop later. Beginning at the first “barrier” in a stream network, we calculated the stream distance (km) added between the barrier and the next barrier or multiple barriers when the stream network split into lower order tributaries. Next, we assessed the stream slope between barriers by examining slope at 100 m stream segments using the “Editor” and “Add Surface Information Tool” in ArcGIS. The maximum slope between “barriers” was compared against species’ gradient tolerances, and either classified as “barrier” or “added habitat”.

2.7. Prioritization

Prioritizing barrier removal begins with a management objective, e.g. increase available potential habitat for fish species. To prioritize amongst potential barriers to remove we ranked each barrier based on the total km that would be potentially available if the barrier were overcome (e.g. removed). We report the expected improvement in cumulative passability for each species if the 10 culverts that restricted the greatest cumulative stream distance were made passable.

3. Results

3.1. Provincial data and LiDAR crossing model comparison

The Hydrocutter model produced a stream network of 4,122 km which was a 6% increase in comparison to the provincial data set of 3,867 km. The inclusion of private roads (forestry industry data) increased linear road kilometers by 64% compared to the provincial road network. The Provincial hydrographic network and Provincial road network identified 431 potential stream crossings: 181 (42%) first order, 127 (29%) second order, 82 (19%) third order, 35 (8%) fourth order, 2 (1%) fifth order, and 4 (1%) sixth order (Table 2). The LiDAR crossing model initially created 1,714 potential stream crossings prior to removal of overlapping or duplicate points and the final, corrected model predicted 1,633 potential stream crossings: 1,182 (72%) first order, 285 (17%) second order, 122 (7%) third order, 38 (2%) fourth order, 2 (1%) fifth order, and 4 (1%) sixth order (Table 2). The total number of potential crossings increased by 278% using the private road network and LiDAR derived stream network in comparison to using the publicly available Provincial data. The crossing density (crossings/km) using Provincial data was 0.11 crossings/km while our model data resulted in 0.39 crossings/km. Given the large differential in potential crossing density, we continued our analysis using the crossing model generated with the private road data.

Table 2. Number of stream crossings by Strahler stream order based on provincial data (Service New Brunswick (SNB) 2016–2018) and the LiDAR modeled data.

Stream order	Provincial data n = 431	LiDAR crossing model n = 1633	Detection difference (%)
1	181	1,182	553
2	127	285	124
3	82	122	49
4	35	38	9
5	2	2	0
6	4	4	0

3.2. Crossing classification

We surveyed n = 242 sites in-situ: culverts = 78 (32%); drainage ditches = 45 (19%); bridges = 58 (24%); sites with no channel upstream = 25 (24%); fords = 8 (3%); and false (desktop) detections = 28 (12%; Table 3). The accuracy of the model predictions ranged from 90–100% depending on the stream order (Table 3). A false detection of a stream crossing occurred where a stream and road did not cross, which was the result of incorrect flowlines or geometry of the roads in these two networks, respectively. These false detections occurred most frequently in first order streams (18%, n = 22 – Figure 4).

Figure 4. Stream-road crossings and false detections based on the results of the LiDAR crossings model in the Restigouche catchment, New Brunswick, Canada.

Table 3. Site survey in the field (n = 242) by Strahler stream order and model accuracy (%).

Stream order	Culvert (model accuracy)	Drainage ditch (model accuracy)	Bridge (model accuracy)	No channel upstream (model accuracy)	Ford (model accuracy)	False detection (model accuracy)
Order 1	30 (93%)	37 (100%)	4 (100%)	20 (90%)	7 (100%)	22 (95%)
Order 2	30 (93%)	3 (100%)	0	4 (100%)	1 (100%)	2 (100%)
Order 3	18 (100%)	3 (100%)	15 (100%)	1 (100%)	0	3 (100%)
Order 4	0	2 (100%)	33 (100%)	0	0	1 (100%)
Order 5	0	0	2 (100%)	0	0	0
Order 6	0	0	4 (100%)	0	0	0

The knickpoint extraction tool generated many false detections, (i.e. a waterfall was predicted, but no crossing feature was present when inspected visually). If the waterfall was identified visually, then the elevation drop can be used to determine if it’s a barrier to fish passage.

3.3. Field validation of culvert characteristics

A total of 78 predicted culverts were compared with field data. The predicted culvert lengths ranged from 5 m to 47 m and slopes ranged from 0.03% to 4% (Table 4). Field and LiDAR predicted culvert lengths and slopes were similar (Table 4). First order culvert lengths had a p-value of 0.79, second order was 0.81, and third order was 0.92, highlighting no statistically significant differences (Table 4). Similarly, there was no statistically significant difference for culvert slopes, first order culvert slope p-values were 0.56, 0.93 for second order, and 0.8 for third order, (Table 4).

Table 4. Mean of culvert length and slope measured using the LiDAR-based model compared to field observations.

	Length				Slope
Stream Order	Mean LiDAR	Mean Observed	T-Stat	p	Mean LiDAR	Mean Observed	T-Stat	p
1	12.4 m	12.4 m	−0.26	0.79	0.93%	0.85%	−0.59	0.56
2	13.5 m	13.0 m	−0.24	0.81	1%	1%	0.08	0.93
3	22.6 m	21.4 m	−0.10	0.92	1.20%	1.20%	−0.26	0.80

3.4. Landscape-scale crossing analysis

The landscape analysis covered an area of 3,223 km² and 4,100 km of streams. The suite of predicted crossings was filtered to remove false detections, and resulted in n = 1,052 crossings, 339 (32%) of which were culverts. Culvert slope varied from 0.15–16.38%. Forty-one percent of culverts had a slope <2%, 24% had slopes 2–4%, 22% had slopes 4–8%, 13% had slopes >8%. Elevation drop represented 172 barriers (for all species – Table 5). Most barriers (due to slope and drop, collectively) were in first order streams (44%). There were few stream segments with multiple barriers (5%).

Table 5. Thresholds of passability, determined by slope, for four species, followed by the number of barriers impeding each species as a function of slope and elevation drop expressed in absolute terms and as a percentage of total culverts in the study area, the kilometers of stream potentially restricted by the predicted slope barriers, and finally, the estimated habitat (stream kilometers) predicted to be inaccessible within the study area.

Species	Slope threshold (%)	Number of barriers due to culvert slope (% of total)	Number of barriers due to culvert elevation drop  (% of total)	Culvert slope – restricted stream in km	Culvert elevation drop – restricted stream in km
Burbot^1	2%	199 (12.2%)	172 (10.5%)	502.8	431
Lake Chub^1	2%	199 (12.2%)	172 (10.5%)	502.8	431
Atlantic Salmon^2	4%	120 (6.9%)	172 (10.5%)	285.3	431
Brook Trout^3	4.5%	105 (5.8%)	172 (10.5%)	237.4	431

¹ MacPherson et al. 2012;

² Bourne et al. 2011; ³Burford et al. 2009.

Using our barrier analysis, we examined the benefit of removing the top 10 most influential culverts, where benefit is measured by additional stream km opened/reconnected. Removing these 10 barriers an estimated 123 km (3%) for Lake Chub, 85 km (2.1%) for adult Atlantic Salmon, and 83 km (2%) for Brook Trout (Table 6).

Table 6. The length of restricted streams, classified by species and quantified in km, and the potential habitat reconnected, expressed as a percentage of the entire study stream network – 4,100 km, assuming the removal of the 10 most influential culverts that are predicted to be barriers.

Species	Stream restricted (km)	Habitat reconnected (% of total stream network)
Burbot	123	3.0
Lake Chub	123	3.0
Atlantic Salmon	85	2.1
Brook Trout	83	2.0

4. Discussion

Stream connectivity is a primary control on the movement of biota and their population and community-scale structure (e.g. Hitt and Angermeier 2008; Huttunen et al. 2017), and is equally important for sediment transport and therefore geomorphic and hydrologic processes that effects both riverscapes and landscapes (e.g. Wondzell and Gooseff 2013; Winter 1999). The goal of this research was to develop a framework and practical GIS tool to assess stream connectivity and then examine fish passability across large landscapes using remote-sensing data, specifically LiDAR generated, digital elevation models and orthophotography. We found that a publicly available road network dataset largely under predicted the extent of stream fragmentation in the study area. By utilizing a higher resolution road network dataset, the numbers of potential barriers increased by 278%. This highlights a critical need to include more resolute road network data to accurately identify/estimate potential barriers and therefore, stream fragmentation. Next, our GIS classification framework proved successful for identification of multiple barrier types: we correctly identified the type of stream-road crossings at >90% of any stream order and at 100% at orders >2. Our desktop assessment of culvert length and slope were also highly accurate. Finally, our framework demonstrated it value for assessing stream network connectivity and predicting the potential habitat that may become available if barriers were overcome at the landscape scale, i.e. a > 3,000 km² region with a > 1,000 km stream network. As one example, 75% of the culverts in our study area exceeded the local, regulatory guidelines for non-baffled, closed-bottom culvert slope (New Brunswick Department of Environment (NBDE)) 2012) which suggests policy and practice are not in sync.

The effectiveness of our framework is underpinned by the resolution of the LiDAR and red-green-blue imagery available to researchers/natural resource managers. Whilst global coverage of LiDAR is yet lacking, high resolution satellite imagery is globally available (e.g. O'Sullivan et al. 2019 – Worldview 3 − 0.3 m resolution). Further, a partnership between the European Space Agency and Airbus has produced a worldwide high resolution 10 m digital elevation model. Future work could test the efficacy of our framework using these types of globally available datasets. The potential success arising from these tests would serve to make these methods globally applicable.

Stream fragmentation is a symptom of all managed landscapes. Although, both managers, scientists, and conservation groups are increasingly taking action to remediate connectivity issues (see for example Correa Ayram et al. 2016), remediation can be costly, and hence the necessity for targeted remediation projects that will potentially reconnect the greatest extent of stream networks. We illustrated the utility of our framework as a first step towards a targeted conservation/restoration strategy. This desktop exercise is significantly quicker than field surveying each crossing, e.g. our study area of 3,223 km² and >1,000 potentially crossings took approximately 150 hours, plus the time for data preparation and creation of road and stream networks. The skill of the user will also affect time to completion. Our relatively simple modelling exercise details how managers can assess barriers and connectivity at the landscape-scale for targeted restoration activities, for example specific fish species, and presents future opportunities to maximize restoration efficiencies. For instance, collating barrier data with landscape-scale models of habitat (O’Sullivan et al. 2021) and temperature (Jackson et al. 2017) would be an improvement in accuracy and efficiency towards reconnecting habitats that are not both hydraulically and thermally resilient (Daigle et al. 2015). Such a targeted approach has myriad applications for land users and managers.

4.1. Limitations

The framework is designed for landscape-scale analyses using remotely collected data and therefore, inherent assumptions exist. LiDAR and other imagery are static data, and don’t reflect temporal variability of the landscape, e.g. culvert repairs or additions over time or natural dynamic processes of riverine systems that alter habitats and accessibility over time (Fullerton et al. 2010; Ogren and Huckins 2015). Other uncertainties exist such as the false detections and model-determined sites with no channel upstream we observed and occurred in first and second order streams (15%). These results could stem from the DEM, its hydrological conditioning, and the flow initiation threshold chosen to derive surface flow channels for our stream network. Wall et al. (2015) demonstrated that flow accumulation can greatly alter the sensitivity in the Hydrocutter model if mapping ephemeral streams or closed depressions in the landscape. We chose a high flow accumulation threshold (40 ha) to derive our surface flows, as compared to the National Hydrographic Network, and this resulted in an increased number of crossings in the DEM. Zahra et al. (2017) explains that a significant range in stream gradients can cause an increase in detection of “false” barriers. Our study area in the Restigouche River catchment ranges from 0-700 MASL which is the likely cause of the many false detections. Overall, our approach led to false detections only in small, 1–2 order streams and this should be considered by managers where smaller catchments are more relevant to their stream network management goals. The flow accumulation threshold can be adjusted to capture smaller catchments and potential crossings. The caveat of this approach is that there could be crossings that are present for ephemeral streams, which would result in a non-detection of the crossing.

An additional consideration arises from the existing definitions of passability for fish at a culvert. Currently, passability is based on a culvert slope and the outlet drop that are assumed to be surrogates for water velocity and therefore, thresholds of swimming ability and jumping ability (e.g. Love and Taylor 2003; MacPherson et al. 2012). In part, this reflects the paucity of research assessing passability for fishes (e.g. Warren and Pardew 1998, Peake 2008, Mahlum et al. 2014, Gautreau 2017). Others have highlighted additional problems with fish passability assumptions (Diebel et al. 2010; Kemp and O'Hanley 2010). For example, high velocity barriers can impede fish passage (Peake 2008, Goerig et al. 2016), but high velocity or culverts with outlet drops can be passed by fish under certain water conditions (Kemp and O'Hanley 2010; Anderson et al. 2012). Our culvert drop of 0.3 m was static and this will change as we learn more about fish behaviour in the field. As passability characteristics thresholds for species improve with new studies, e.g. better slope thresholds and culvert design differences (internal baffles/smooth pipe/corrugated pipe, e.g. Bouska and Paukert 2010), the improved thresholds and criteria can quickly and easily be inserted into the framework to improve overall predictability for managers.

Additionally, our model can only detect potential barriers based on topographic position and doesn’t differentiate between functional/non-functional culverts. Culverts can be collapsed or clogged with debris which is common due to lack of maintenance (Ward and Stanford 1995; Gautreau 2017). The only de facto process to identify functional culverts is to visit and assess their status. While our framework can’t achieve a complete assessment of functionality at all individual road crossings, it can significantly shorten the workflow leading to priority assessments as, for example, opening of potentially new habitats (Erkinaro et al. 2017).

5. Conclusion

The goal of this research was to present a cost effective, desktop framework and model to locate and identify road stream crossings, assess their potential to be barriers, in our example passability for fish, and prioritize restoration to overcome riverscape fragmentation. We correctly identified the type of crossing visually from LiDAR DEM and orthophotography with >90% accuracy and 100% for stream orders >2. Additionally, the use of the lowest elevation cell on each side of a road embankment was a repeatable process and produced an accurate measurement of a culvert’s length and slope. As a result, the model was well suited for a landscape-scale analysis to remotely locate, identify, and evaluate stream crossings, and to be cost-effective if the LiDAR data is available. As LiDAR technology continues to advance and such data becomes more readily available, land and resource managers and conservation organizations can be equipped with quick, efficient, and accurate landscape-scale tools for many management decisions such as our stream connectivity analysis.

In general, it is extremely difficult to correctly characterize fish habitats in smaller streams at a landscape-scale using remote sensing methods. The framework builds on the use of stream gradient as an indicator of potential habitat (e.g. Januchowski-Hartley et al. 2014), and then to demonstrate the framework’s functionality for decision-making tools, examined stream length of acceptable gradients as a surrogate metric of habitat added by removing a barrier (e.g. Erkinaro et al. 2017). It is a demonstration of a simple, desktop-derived metric to highlight the functionality and adaptability of the framework. The ability to assess actual aquatic habitats from remotely sensed images is improving (e.g. Praskievicz and Buege 2017, O’Sullivan et al. 2021) and we believe managers will eventually have on their desktops, both habitat and barrier maps at the landscape-scale derived from remotely sensed data that can be used effectively in decision-making.

Acknowledgements

Thank you to the Atlantic Salmon Conservation Foundation, MITACS, NSERC (RGPIN-2018-06015 – RAC), and Collaboration for Atlantic Salmon Tomorrow for funding support, staff of the Restigouche River Watershed Management Council for initiating this effort and providing field support, and J.D. Irving. Limited for the improved road network data.

Data availability statement

The Provincial data used in the study are openly available from Service New Brunswick at http://www.snb.ca/geonb1/e/index-E.asp. The industry’s road network was provided privately via a research agreement.

Disclosure statement

No potential conflict of interest was reported by the authors.