A novel approach for designing large-scale river temperature monitoring networks

Rising water temperatures (Tw) have the potential to alter the thermal suitability of rivers for freshwater fish, which are frequently the focus of management (Mohseni et al. 2003; Isaak et al. 2010, 2012). Cold water fish such as salmonids are highly sensitive to river temperature which affects growth, metabolism, performance, survival and demographic characteristics (Elliott 1994; Gurney et al. 2008). Atlantic salmon (Salmo salar) and, to a lesser extent, brown trout (Salmo trutta) have a high economic (Radford et al. 2004), recreational and conservation value (Anon 2009). Consequently, there are strong socio-economic drivers for understanding the spatio-temporal dynamics of thermal regimes, their sensitivity to drivers of change and opportunities for management or mitigation of thermal extremes (Malcolm et al. 2008; Hrachowitz et al. 2010). In recognition of the importance of these issues, CAMERAS (Coordinated Agenda for Marine, Environment and Rural Affairs Science), an umbrella group of Scottish Government departments and agencies, prioritised the development of a strategic national water temperature network in their recent freshwater monitoring action plan.

Large-scale Tw networks are required to characterise and understand temperature variability and make predictions of current and future Tw in monitored and unmonitored rivers (Hrachowitz et al. 2010; Deweber et al. 2014). However, globally there are relatively few quality controlled long-term networks and even fewer large-scale (at least regional and >100 km²) planned and coordinated river temperature monitoring networks (Table 1). Many monitoring networks are produced ad hoc, evolving over time with poorly defined objectives or represent aggregations of numerous data sets, spanning multiple regions (e.g., US Geological Survey (USGS), NorWest) or countries (e.g., GEMS/Water). Data are often collected with a range of earlier aims, at varying sampling frequencies, with varying deployment approaches and equipment (Table 1). This can result in spatial and temporal biases. As such, aggregated networks arguably do not provide the consistency necessary for use over wider spatial domains. This is especially important in the context of understanding environmental change where temperature trends may be small relative to measurement bias.

A lack of strategic planning can potentially limit the value of networks if sites are not representative of the parameter of interest or the processes or landscape characteristics that control the parameter (Parr et al. 2002; Deweber et al. 2014). Where a network contains numerous monitoring sites with similar characteristics, or where a network provides incomplete coverage of process or landscape controls (including spatial coverage) then highly uncertain or biased model fits and predictions will result (Marsh & Anderson 2002; Deweber et al. 2014).

Recent advances in geographical information systems (GIS), spatial statistics and inexpensive temperature data loggers now allow the strategic design and deployment of large-scale monitoring networks, making monitoring and modelling of river temperatures increasingly possible (McCullough et al. 2009; Thomassen et al. 2010; Sowder & Steel 2012). The need for larger-scale monitoring and understanding has been identified by a number of researchers including Tetzlaff et al. (2008), who highlighted the need to upscale studies from small (∼1 km²) to larger catchments (∼10² km²) to facilitate management. In the context of river temperature, it is not possible to directly upscale process-based energy budget studies to larger spatial scales due to high costs, logistical challenges and the computational burden of such work (Hannah et al. 2008; Malcolm et al. 2008; Hilderbrand et al. 2014).

However, large-scale statistical modelling of Tw using landscape characteristics that are proxies for energy exchange processes or controls show significant potential to inform effective environmental management (Isaak & Hubert 2001; Hrachowitz et al. 2010; Chang & Psaris 2013). Landscape data provide a cost-effective method of generating environmental data across large spatial scales (Wehrly et al. 2009). GIS analysis can be used to determine landscape characteristics at any point on a river network, without the expense of field survey. Furthermore, the availability of inexpensive data loggers has dramatically increased Tw monitoring (Sowder & Steel 2012) to the extent that staff time, quality control and appropriate data storage are greater constraints on logger deployment than the cost of instrumentation.

Despite these advances, relatively few studies have modelled temperature distributions across whole basins in relation to environmental and landscape controls (Hrachowitz et al. 2010). Additionally, as far as the authors are aware, there have been no attempts to establish a large-scale strategically designed network that meets the requirements for modern spatio-temporal statistical modelling, which include appropriate coverage relative to landscape predictors (covariates), calibration, quality control and data storage. Such a network and associated modelling have the potential to answer critical management questions about the spatial variability in river temperature and its controls, the effects of changing landuse and the likely impacts of climate change. The current paucity of such networks demonstrates a challenge to understanding thermal regimes at multiple spatial and temporal scales (Garner et al. 2014) and to informing appropriate management of rivers.

This paper aims to develop a novel methodology for the design of a large-scale water quality monitoring network using the Scotland River Temperature Monitoring Network (SRTMN) as a case study. This initiative aims to produce a network which avoids common limitations exhibited by large-scale networks and has the potential to provide data appropriate for spatio-temporal analysis.

It is crucial that the objectives of any monitoring network are defined a priori (Parr et al. 2002). In the case of the SRTMN, the network is required to provide data to: (1) characterise spatial and temporal variability in thermal regimes across Scotland; (2) assess climatically sensitive rivers; (3) improve understanding of landscape controls on water temperature and their ability to buffer water temperature changes and extremes; and (4) develop models to predict spatio-temporal variability in river temperature, including in unmonitored catchments and locations. Using the information from 3 and 4, mitigation and adaptation strategies for high temperature may be assessed. Finally, and importantly, the network is expected to provide long-term monitoring of river temperature, covering the range of expected environmental responses across Scotland.

To address these objectives, it was crucial that the sites cover the environmental range and combinations of landscape controls observed in Scotland's rivers. Given the importance of Atlantic salmon as a target species for management and conservation, the environmental range was constrained to accessible rivers using the map of Atlantic salmon distribution originally developed by Gardiner & Egglishaw (1986). In practice, this constrained the altitudinal range from 0 to ca. 700 m, above which any long-term monitoring would also have been impractical. As the SRTMN is a new network, strategically planned from the start, sites could be selected to cover landscape attributes appropriate to the objectives. By covering the range and combination of controlling variables it will not be necessary to predict outside the range of observations. Furthermore, because the landscape covariates are process based, informed by the literature, the resulting statistical models will be transferable, avoiding spurious correlations. Taken together, this should ensure accurate and unbiased predictions for unmonitored locations. Finally, the SRTMN seeks to minimise the amount of changes to monitoring sites post-deployment (cf. Hrachowitz et al. 2010) thereby reducing staff costs as loggers are deployed once and then downloaded bi-annually, across a consistent network.

It is also important to define the available resource which controls the number of monitoring sites and influences the number of covariates upon which sites could be selected. For the purposes of this network, 200 loggers were identified as the preferred number for deployment, rising to a maximum of 250. The additional loggers were held in reserve in case the preliminary deployment plan was unable to cover all environmental combinations and, importantly, to allow batches of loggers to be recalibrated.

To maximise the value of the network to government and environment agencies, environmental designations such as Special Areas of Conservation (EU Habitats Directive 92/43/EEC) were considered in the selection of study catchments (where consistent with the required experimental design) so that resulting data could be used for reporting environmental status. Catchments impacted heavily by hydropower and storage schemes were avoided to ensure the network provides an understanding of near-natural spatial variability and long-term change. However, it is important to recognise that truly natural mid/lower rivers in Scotland can be difficult to find due to management modifications (Gilvear et al. 2002).

Elevation can be used as a surrogate for air temperature (Ta), as adiabatic lapse rates reduce Ta with altitude, which influences Tw (Hrachowitz et al. 2010; Hill & Hawkins 2014). Therefore, elevation is a significant predictor of Tw, influencing mean monthly and maximum stream temperatures (Imholt et al. 2011; Chang & Psaris 2013) and its importance changes over time. For example, Chang & Psaris (2013) found mean elevation to be the only significant predictor of water temperature during wet winter months.

Discharge (Q) is related to thermal capacity and affects rates of heating and cooling. Previous studies have shown an inverse relationship between Q and Tw and increasing rates of warming at lower discharges. For example, van Vliet et al. (2011) showed increasing Q by 20% decreased Tw and decreasing Q by 20 and 40% caused an increase in Tw. Unfortunately, it would be impractical to measure discharge at all potential monitoring locations so GIS derived landscape proxies were required. These were derived from a digital elevation model (Ver Hoef et al. 2006). Upstream catchment area can be used as a proxy for river discharge (Ver Hoef et al. 2006; Hannah et al. 2008; Johnson et al. 2014) especially where interactions are considered with other covariates that reflect climatological gradients (e.g. latitude and longitude). Downstream reaches have larger discharges, thus a larger thermal capacity moderating maximum temperatures compared to upland reaches (Imholt et al. 2011).

The land use, geology and the presence of standing water within a catchment also has a key role in controlling thermal regime through effects on catchment responsiveness and residence times, which influence time available for energy exchange processes (Hill et al. 2013). Furthermore, groundwater–surface water interactions at the catchment scale can be inferred from catchment geology. For example, Hill et al. (2013) attributed warmer mean winter Tw to geologic permeability, which was associated with groundwater flow in the catchment.

Catchment topography governs the amount of shading provided by the landscape (herein hillshading) which strongly influences solar radiation, particularly during periods of low solar angle, like winter (Hrachowitz et al. 2010). Hillshading was therefore calculated for both summer and winter, to encompass the annual variability due to changes in azimuth, zenith angles and day length. As a result, the maximum (winter) and minimum (summer) amount of hillshading were found for each point. The amount of hillshading was calculated for six time points (06:00, 09:00, 12:00, 15:00, 18:00, 21:00); the resulting hillshading values were then weighted depending upon the amount of incoming radiation (Fu & Rich 1999) and averaged to give an overall summer (winter) hillshading value for each point. Azimuth and solar altitude values were found for the centre point of each catchment and this was used for all locations.

Finally, there are numerous controls that affect energy exchange and thus stream temperature at the reach scale. Topographic controls have been correlated with temperature variability, particularly when characterised for intermediate scales (1 km) around monitoring locations (Hrachowitz et al. 2010; Isaak et al. 2010; Chang & Psaris 2013). Stream width controls the surface area available for energy exchange and for a given catchment area strongly influences width-depth ratios (Imholt et al. 2011). Wide rivers are also characterised by relatively lower topographic and vegetative shading, increasing sensitivity to solar inputs (Malcolm et al. 2004, 2008; Imholt et al. 2011; Li et al. 2012; Chang & Psaris 2013; Ryan et al. 2013). Upstream catchment area can also be a proxy for stream width where this cannot be assessed directly from a GIS, for example, where rivers are represented as lines and have no area attributes (Imholt et al. 2011; Peterson & Ver Hoef 2014).

Channel orientation is important for receipt of solar radiation (Guan et al. 2013). It affects the amount of solar radiation reaching the stream and the shading effects of banks and vegetation, with north/south channels experiencing maximum exposure to incoming radiation and east/west channels the minimum (Malcolm et al. 2004). Orientation is particularly important when considering the effects of land-use as it affects channel shading from riparian vegetation (Ryan et al. 2013).

Transit times, bed friction and channel morphology are influenced by channel slope. This can alter the amount of time available for energy exchange processes and also the degree of hyporheic or groundwater exchange. Furthermore, steep channel slopes often result in greater channel incision and thus greater topographic shading (Moore et al. 2013).

Riparian planting has been suggested as a potential tool to manage high temperatures and mitigate against the effects of climate change (Malcolm et al. 2008; Hrachowitz et al. 2010; Garner et al. 2014). Numerous studies have found the percentage of forest cover to be an important predictor of maximum river temperature at all the scales assessed (Hrachowitz et al. 2010; Imholt et al. 2011; Chang & Psaris 2013). ET GeoWizard tools were used to create individual buffers upstream of each potential monitoring site. The percentage of woodland was calculated within each buffer using OS MasterMap for land cover information. It is important to recognise that forestry operations may have altered some of the land cover since the creation of the MasterMap data set in 2012. Consequently, ground truthing riparian vegetation characteristics will be important during network deployment.

GIS-derived metrics were calculated, every 500 m across the river network, using tools within ArcGIS10 and the RivEX river networks tool (Hornby 2015) to generate a data set of potential monitoring locations. To simplify the site selection process all covariates were standardised to range between 0 and 1. To ensure that the ‘chosen’ monitoring locations represented the environmental range of ‘potential’ locations a regular grid was placed over the environmental characteristics of all the potential sites. This is akin to (although not as rigorous as) the concept of Latin squares, which ensures coverage of all potential environmental combinations (Gao 2005). Latin square designs ensure that each characteristic occurs only once in the experimental design and that no combinations of characteristics are missing (Gao 2005). Given the spatial coverage of the SRTMN, it would be impossible to replicate all covariate combinations at all geographic locations (Martin 2001) and the limited number of data loggers and upkeep required meant full implementation of a Latin square design was not possible, but the principles remain similar.

Data sets collated from multiple organisations can contain high levels of redundancy due to poor coordination between agencies (duplication of effort) or offer only limited opportunities for direct comparisons, for example, where loggers are deployed for differing time periods (e.g., limited to summer months) or with different sampling frequencies (Isaak et al. 2012; Mauger et al. 2015). To get the maximum value from existing resources and deployments, the characteristics of current water temperature monitoring locations (Marine Scotland Science (MSS), Scottish Environment Protection Agency (SEPA) and River and Fisheries Trust Tw loggers) and infrastructure (SEPA gauge locations) were calculated. Current Tw monitoring sites were overlaid on the covariate selection plots (Figure 5, squares) and where characteristics were already covered, ‘chosen sites’ were removed from the deployment plan. For example, in the SRTMN, sites in the Upper Dee were removed as current monitoring undertaken by MSS and the River Dee Trust already covered the required environmental range and conformed to SRTMN quality control procedures and protocols. Where current monitoring sites did not conform to SRTMN protocols they were brought to the same standard to ensure comparability of data sets (see ‘Stage 4: quality control’ below).

Similarly, there was a desire to make use of existing telemetry infrastructure even where temperature monitoring did not currently exist. For example, the location of SEPA gauging stations was also overlain on the selection grid and used to replace SRTMN selected sites where they lay within the defined point radius. The addition of gauging station locations to the network had the added benefit of providing discharge data that are potentially useful for understanding river thermal regimes, even at a small number of sites.

The logistics and health and safety involved in maintaining a continuous monitoring site must be considered (Laize 2004). Costs are increased if data loggers are in areas of limited accessibility, involving time-consuming hikes to isolated streams for downloading. This can create problems in maintaining that particular site long-term and requires additional safety considerations. In addition, the loss of equipment, for example, through vandalism, can also affect the long-term viability of a network and the quality of data collected from it. Consequently, local knowledge from collaborators was used to assess the risk of loss, drying out or vandalism. Where this was likely, alternative locations for the originally chosen point were found from other points within a radius (diamonds in Figure 5) of the desired characteristics (black grid points in Figure 5).

Large-scale networks can be complicated and confounded by differences in monitoring equipment, measurement drift, quality control procedures and data archiving (Parr et al. 2002; Hannah et al. 2011; Mauger et al. 2015). United Nations Environment Programme (UNEP) highlights that robust quality assurance and quality control is crucial when collating data from numerous sources, yet this is more easily stated than achieved (UNEP 2014). To prevent any inconsistencies standard operating procedures were created following Joint Code of Practice (JCoPs) (Anon 2012) to create a robust and comparable data set (Mauger et al. 2015).

There are a variety of methods for anchoring and positioning loggers in the stream, which could influence uniformity of deployment across a large network, creating issues when comparing data. The deployment method of the SRTMN states that loggers should be placed in similar stream areas and should ensure the location is representative of the reach, in a well-mixed water column that will not stratify and is unlikely to dry out during warm periods (Imholt et al. 2011). Loggers were anchored using metal stakes in the stream, where they will remain throughout the monitoring period (Imholt et al. 2011). An alternative deployment method using a land anchor and weighted logger was used for deep rivers where it was deemed potentially hazardous to have staff members in the water.

Methods of sensor shading/shielding also vary between published studies and uncertainty remains about how much effect solar radiation has on temperature measurements (Johnson & Wilby 2013; Johnson et al. 2014). The SRTMN uses the most common technique of placing loggers inside PVC tubes to prevent exposure to direct sunlight (Hrachowitz et al. 2010; Imholt et al. 2011; Isaak et al. 2013). These tubes are white and have a high albedo to reflect radiation, but are also large enough to allow water to circulate around the logger ensuring the temperature is representative (Brown & Hannah 2008).

All loggers were calibrated against two reference loggers, which were, in turn, calibrated by a UKAS accredited laboratory 0794 (SGS UK Ltd). The reference loggers were calibrated by comparison to Semi-Standard Platinum Resistance Thermometers, in a stirred water bath, with measurements accurate to ±0.0025 °C. Recalibration of deployed loggers will be undertaken no greater than every three years, to ensure data quality is retained. Reference loggers are calibrated annually. To permit a rolling programme of calibration and uninterrupted data collection, a store of spare loggers was retained at MSS. The size of this store reflected the maximum number of loggers (40) found in a single catchment. This allows a batch of loggers to be sent to a collaborator for them to rotate into the network; the previously deployed loggers are then returned to MSS for recalibration. Once recalibrated these loggers form the next batch of loggers to be deployed and the process is repeated in the next catchment in the schedule.

It is also important that data are traceable, quality controlled and have appropriate metadata (Hannah et al. 2011; Mauger et al. 2015); consequently, a centralised database (FleObs) was created to house the data collected by the SRTMN. The FleObs database stores information on loggers, sensors, sites and their various combinations and also stores information on the calibration which can be used to correct raw data on export and provide a measure of uncertainty in the resulting measurements. Calibration equations have valid periods associated with them, and new equations (with their associated valid period) are added when loggers are recalibrated. Therefore, when data are exported from the database they can be corrected using the calibration equation associated with that time period, for that logger. This ensures that all calibration, data quality control and correction are undertaken to traceable and recognised national standards. This is extremely important for a long-term network where data are collected by numerous collaborators.

Following the first year of data collection (Figure 3) and associated modelling it is useful to assess the success of this proposed network design. In brief, this will involve model validation and power analysis to guide network revision. Leave-one-out cross validation, removing single sites or entire catchments, can be utilised for model validation (Hrachowitz et al. 2009). This will allow prediction error to be estimated and demonstrate the ability of the models to make predictions for new sites and catchments (Hrachowitz et al. 2009). Power analysis can then be used to assess the magnitude of temperature effects that could be detected by different covariates (Isaak et al. 2014). Sites may be removed from the network if fitted models suggest that characteristics are not significant predictors of temperature. Alternatively, additional loggers may be deployed to improve estimates of the effects of individual variables including spatial autocorrelation (Isaak et al. 2014).

A perfectly gridded coverage, as exemplified in Figure 4, could not be expected for the chosen sites. However, Figure 6 demonstrates a good coverage of the environmental range across all combinations of potential controlling variables. If a network were biased to particular characteristics, uncertainty will be increased for any extrapolation from monitored to unmonitored locations (Wagner et al. 2014). As the selected sites for SRTMN cover the environmental range and combinations of variables (Figure 6), this suggests that the relative importance of different controls should be isolated in future modelling. Consequently, the network should deliver the necessary data requirements to improve understanding of the controls on water temperatures at different spatial and temporal scales. In addition, it is expected that predictions may be made for unmonitored catchments as issues of monitoring deficiency will be avoided (Laize 2004).

Integrating current monitoring sites and infrastructure into the SRTMN avoids duplication and unnecessary maintenance costs (Parr et al. 2002; Mishra & Coulibaly 2009). To enable current monitoring sites to be integrated into the SRTMN, sensors and data loggers needed to be calibrated and deployed following the standard operating procedures created for the SRTMN. As a result, any issues of inter-institutional or equipment bias, lack of cross calibration, variability in sensor anchoring and shielding may be avoided. By avoiding duplication of site characteristics, and thus data redundancy, each site will contribute unique information into statistical models, consequently making a useful and cost-effective network (Mishra & Coulibaly 2009).

This paper described and evaluated a potential methodology for the design of a new monitoring network. The approach was illustrated using the SRTMN as a real-world, practical case study. The method characterised the environmental characteristics of potential monitoring sites to cover the environmental range of controlling variables, required to: (1) characterise spatial and temporal variability in thermal regimes across Scotland; (2) identify climatically sensitive locations; (3) improve understanding of controls on Tw; (4) develop models to predict future river temperatures and predict thermal regimes in unmonitored rivers; (5) assess mitigation and adaptation strategies for high temperature; and (6) provide long-term monitoring of thermal regimes. The network is strategically planned to ensure the desired coverage of controlling characteristics rather than spatially balanced or randomly located sites which are often the focus of previous networks (Isaak et al. 2014 considers). It is therefore anticipated that the network will have minimal redundancy and high levels of statistical power and meet the objectives identified at the start of the network design process.

From the development of this method for the SRTMN, the following key recommendations can be made for designing other large-scale monitoring networks:

• Begin with clear network aims and objectives that identify data requirements.

• Where large-scale spatial statistical models are required, undertake a literature review to determine process drivers and more readily obtained proxies (e.g., GIS or remote sensing data) to represent these processes.

• Assess the amount of resource available and consequently the number of sampling sites or samples that can be planned.

• Select sites to cover the range of environmental characteristics which influence the parameter of interest; an adaptation of Latin squares principles may be used.

• Develop comprehensive standard operating procedures and data storage facilities for data quality control.

• Where possible, integrate current monitoring sites and existing infrastructure to make best use of collective resource.

The merits of this network design will be tested further when data are returned and analysis undertaken. Further research could involve implementing the principles of this approach in other large-scale network designs with different research objectives and target parameters (e.g., water chemistry, fish abundance). The principles identified here are likely to be applicable across different large-scale monitoring networks, due to the values of GIS for assessing landscape characteristics at large spatial scales. Upscaling process-based knowledge to larger spatial scales is a major challenge across disciplines and is required to inform appropriate management, but critically requires large-scale high quality monitoring networks such as SRTMN. Finally, adjustments to this methodology could also be used to assess and revise current monitoring networks that have grown organically and potentially contain redundancy.

This project is part of Marine Scotland Science (MSS) Freshwater Fisheries Laboratories Service Level Agreement FW03N and it is supported by a NERC Open CASE studentship (University of Birmingham and MSS) held by Faye Jackson; grant reference NE/K007238/1. Some map features are based on digital spatial data licensed from the Centre of Ecology and Hydrology, © NERC © Crown Copyright and database right (2014). All rights reserved. Ordnance Survey License number 100024655. Catchment boundaries were from SEPA (2009). Special thanks are given to Hilary Anderson for her assistance with GIS analysis. The authors also thank the SRTMN collaborators for their feedback on the proposed sites and assistance in deploying and maintaining the SRTMN. Two anonymous reviewers provided helpful comments to improve the manuscript. The authors declare that they have no conflict of interest. This article does not contain any studies with human participants or animals performed by any of the authors.