Abstract

The ability to monitor water temperature is important for assessing changes in riverine ecosystems resulting from climate warming. Direct in situ water temperature collection efforts provide point samples but are cost-prohibitive for characterizing stream temperatures across large spatial scales, especially for small, remote streams. In contrast, satellite thermal infrared imagery may provide a spatially extensive means of monitoring riverine water temperatures; however, researchers do not have a good understanding of the accuracy of these remotely sensed temperatures for small streams. Here, we investigated the utility of Landsat 8 thermal infrared imagery and both local and regional environmental variables to estimate subsurface temperatures in high-latitude small streams (2–30 m wetted width) from a test watershed in southcentral Alaska. Our results suggested that Landsat-based surface temperatures were biased high, and the degree of bias varied with hydrological and meteorological factors. However, with limited in-stream validation work, results indicated it is possible to reconstruct average in situ water temperatures for small streams at regional scales using a regression modelling framework coupled with publicly available Landsat or air temperature information. Generalized additive models built from stream stage information from a single gage and air temperatures from a single weather station in the drainage fit to a limited set of in situ temperature recordings could estimate average stream temperatures at the watershed level with reasonable accuracy (root mean square error = 2.4°C). Landsat information did track closely with regional air temperatures and we could also incorporate it into a regression model as a substitute for air temperature to estimate in situ stream temperatures at watershed scales. Importantly, however, while average watershed-scale stream temperatures may be predictable, site-level estimates did not improve with the use of Landsat information or other local covariates, indicating that additional information may be necessary to generate accurate spatially explicit temperature predictions for small order streams.

Introduction

Water temperature is a key physical parameter structuring aquatic ecosystems (Allan and Castillo 2007), influencing water chemistry (e.g., dissolved oxygen; LeBosquet and Tsivoglou 1950) and driving metabolic rates of organisms (Griffiths and Schindler 2012). Riverine water temperature regimes can drive species distributions (Buisson et al. 2008), species growth rates (Holtby 1988; Armour 1991), and the timing of migration and spawning events (Hodgson and Quinn 2002) and can impact aquatic species interactions (De Staso and Rahel 1994). Broad-scale stressors such as climate change have the potential to influence thermal regimes across watersheds (Isaak et al. 2010). Additionally, anthropogenic activities can alter stream temperatures (Webb et al. 2008). For example, logging, the construction of dams, and the removal of riparian vegetation can influence stream hydrology, sediment inputs, and riparian shading (National Research Council 1996; Holtby 1988; Poole and Berman 2001). To predict how riverine species will respond to changing thermal conditions and to inform management, spatially extensive water temperature information is critical (Handcock et al. 2012; Dugdale 2016). For instance, researchers predict increases in water temperature associated with climate change to differentially impact multiple native and nonnative trout species in the United States (Wenger et al. 2011). Similarly, high river temperatures may preclude Chinook Salmon Oncorhynchus tshawytscha recovery in the state of Washington by rendering some habitats as unsuitably warm (Lisi et al. 2013). Rapidly changing temperature regimes in riverine systems threaten to further impact economically and culturally important species, necessitating that we monitor temperatures in these systems.

While extensive temperature datasets are available in some regions (e.g., Isaak et al. 2017), the availability of water temperature data in remote watersheds is sparse, particularly for small streams (McPhee et al. 2016; Cunningham et al. 2018). This is, in part, because monitoring stream temperatures in remote watersheds is costly in terms of labor, travel, and instrumentation. Therefore, available in situ water temperature data (sampled directly from the stream) typically only represent a small portion of a watershed's thermal profile. Instead, efforts to estimate stream temperature from other publicly available archival or remote sensing data sources may offer a more efficient and low-cost alternative for collecting stream temperature data (Handcock et al. 2012; Sohrabi et al. 2017).

Previous efforts have successfully estimated water temperature with relative accuracy using local information on air temperature and water flow rates (e.g., Caissie et al. 1998, 2001; Mohseni et al. 1998; Benyahya et al. 2007; Sohrabi et al. 2017). These studies often take advantage of the typically tight coupling between air and water temperature and may also incorporate other environmental variables to improve model predictions (Benyahya et al. 2007). However, it appears that utilizing these environmental covariates can lead to poor water temperature estimates in some cases; for instance, the propensity for snowmelt in the late spring and summer may result in large estimation errors (Sohrabi et al. 2017). Additionally, the correlation between air temperature and stream temperature is extremely complex as this relationship can be modified by geological and riparian environments, making the estimation of water temperature difficult (Johnson 2003).

As an alternative to water temperature estimation via environmental correlates, researchers have used thermal infrared (TIR) imagery to remotely sense temperatures of water bodies, including the use of thermal sensors attached to helicopters and fixed-wing aircraft (Hick and Carlton 1991; Torgersen et al. 2001; Cristea and Burges 2009) as well as satellite-derived TIR imagery. Given that thermal imagery has a broad range of remote sensing applications, researchers have deployed infrared sensors on satellites since the 1970s (USGS 2015a). Currently, Landsat 7 and Landsat 8 are in orbit, the latter of which provides surface-temperature information by remote sensing TIR radiation at a 100-m resolution, which is then resampled to 30-m pixels using cubic convolution interpolation (Keys 1981; USGS 2015a).

Previous efforts have utilized TIR data from satellites to map surface thermal heterogeneities for large rivers and streams in the United States (Kay et al. 2005; Handcock et al. 2006), France (Wawrzyniak et al. 2012),as well as thermal pollution from dams in China (Ling et al. 2017). However, researchers deemed satellite TIR data provided suitable spatial resolution for predicting water temperatures only for large to midsize rivers (Cherkauer et al. 2005; Handcock et al. 2006; Wawrzyniak et al. 2012). For example, Handcock et al. (2006) tested the ability of both airborne and satellite TIR imagery (using Landsat 7 imagery with a pixel size of 60 m and advanced spaceborne thermal emission and reflection radiometer [ASTER] images at a resolution of 90 m) to estimate stream temperature in the Pacific Northwest across streams with a range of wetted widths (10–500 m), finding satellite thermal imagery would only be suitable for predicting temperatures in very large rivers, roughly 180 m across or more (average bias for large rivers was low at + 1.2°C). Similarly, working in the Rhône River, Wawrzyniak et al. (2012) concluded Landsat 7 thermal imagery data would only be suitable for predicting river temperatures for reaches 60 m or greater in width (i.e., > 1 Landsat 7 pixel wide). Lalot et al. (2015) found that Landsat 7 satellite imagery may be capable of estimating temperatures in rivers less than 180 m across, but that the accuracy of these estimations is subject to large fluctuations.

While the ability to assess large river water temperatures with satellites is promising, smaller-order streams make up a large component of temperate freshwater habitats; roughly half of the running waters in the contiguous United States are smaller first-order streams (Allan and Castillo 2007). For smaller rivers and streams, the influence of terrestrial ground cover and differences in material emissivity may interfere with attempts to measure water temperatures with satellite thermal imagery due to a mismatch between the spatial grain of satellite data and small stream widths.

Utilizing spatial sharpening algorithms (Gustafson et al. 2003), additional initiatives have successfully improved the resolution of satellite imagery (i.e., from ASTER images) from 90 to 30 m (Teggi 2012) and have found that water temperature estimation may be possible in rivers larger than 60 m (Despini and Teggi 2013). To our knowledge, however, no one has tested Landsat 8 TIR imagery, which is publicly available at a pixel resolution of 30 m (after interpolation by cubic convolution), for use in estimating temperatures in streams smaller than 1 pixel (30 m). Our study aimed to satisfy two primary objectives: 1) quantify the ability of Landsat 8 TIR imagery and local and regional environmental data, including air temperature and stream flow information, to predict in situ water temperatures and 2) determine factors that influence the accuracy of Landsat 8 TIR imagery to predict in situ water temperature in order to provide practitioners with guidelines for its use. To evaluate the potential for estimating small-stream temperatures via environmental covariates and remotely sensed TIR data, our study used a set of validation sites on streams with wetted widths ranging from 2 to 30 m in a temperate watershed in southcentral Alaska as a test case.

Methods

We collected in situ stream temperatures with data loggers over an open-water season in a temperate watershed in southcentral Alaska. We extracted a combination of land cover, stream stage, and air temperature information and derived remotely sensed temperatures from calibrated Landsat 8 images to assess their ability to estimate small-stream water temperature. We generated and assessed the predictive ability of simple models to estimate stream temperature for use by practitioners. Finally, we determined whether environmental conditions could help explain variation in the accuracy of Landsat-derived temperature to estimate stream temperatures at both the watershed and local scales.

Study area and data collection

The study took place within the Anchor River watershed on the Kenai Peninsula in southcentral Alaska (Figure 1). The total watershed area is 583 km2 and previous researchers have classified it as a peat wetland–supported system (Rinella et al. 2009). Maximum stream discharge occurs from August to November with low flows occurring during June and July (Mauger 2005). We collected in situ stream temperatures from May through September 2015 at 10 validation sites, including 3 on the South Fork Anchor River, 5 on the North Fork Anchor River, and 2 on the Chakok River, a North Fork tributary (Figure 1, Data S1, Supplemental Material). To avoid thermal input from infrastructure heat sinks, we chose validation sites at least 30 m from bridges, roads, or buildings. We recorded temperatures every hour using HOBO TidbiT v2 loggers (Onset Computer Corp., Bourne, MA; Hoboware 2014) placed midway between the thalweg and wetted edge at deployment sites and anchored 5 cm from the streambed. We recorded data logger locations with a survey-grade GPS unit (model Geo7X; Trimble Inc., Sunnyvale, CA).

Figure 1.

The Anchor River watershed and locations of in situ temperature collection validation sites used to understand the ability of various environmental covariates and Landsat data to estimate small-stream water temperature. The U.S. Geological Survey stream discharge gage is annotated as USGS (gage number: 15239900; 59°26′42″N, 151°27′3.96″W), Natural Resources Conservation Service Snotel (site number: 1062; 59°51′34.92″N, 151°18′39.96″W) air temperature data collection site is annotated as “SNTL”. Validation sites labeled SF are on the South Fork Anchor River, NF are on the North Fork Anchor River, and CR are on the Chakok River.

Figure 1.

The Anchor River watershed and locations of in situ temperature collection validation sites used to understand the ability of various environmental covariates and Landsat data to estimate small-stream water temperature. The U.S. Geological Survey stream discharge gage is annotated as USGS (gage number: 15239900; 59°26′42″N, 151°27′3.96″W), Natural Resources Conservation Service Snotel (site number: 1062; 59°51′34.92″N, 151°18′39.96″W) air temperature data collection site is annotated as “SNTL”. Validation sites labeled SF are on the South Fork Anchor River, NF are on the North Fork Anchor River, and CR are on the Chakok River.

All validation sites were remotely sensed using Landsat-derived information available at a resolution of 30-m pixel size. Given that these 30 30-m pixels were resampled from an original 100-m pixel size, extracted temperatures represent averaged values of nearby pixels. We extracted Landsat 8 TIR (band 10) images from six cloud-free dates to minimize atmospheric influence (USGS 2015a); however, failure of some sensors resulted in a total of 47 data points. We calibrated images using quantized-scaled digital numbers (Formula S1; USGS 2016).

Validation site delineations

To understand how local environmental conditions may impact the accuracy of Landsat-derived temperatures, we characterized the validation sites (30 × 30 m) based upon a suite of stream and riparian attributes. 1) We used the Rosgen stream classification–level I system (Rosgen 1994) to define the stream type at each site (referred to here at Rosgen stream class). This classification system has been widely used among aquatic science professionals and provides a simple habitat characterization framework without the need for practitioners to record extensive stream reach measurements (Kasprak et al. 2016). Anchor River sites reflected types B (moderate gradient reaches with low sinuosity), C (reaches with defined floodplains and channel meandering), and E (low slope reaches with high channel sinuosity) stream classes. 2) We estimated percentage of water cover for each site in ArcGIS (ESRI, Redlands, CA; ESRI 2014) using high-resolution aerial imagery collected in 2010 and 2012 by Cook Inletkeeper, Watershed Sciences Inc., by digitizing the wetted portion of the area captured within the bounds of the 30-m Landsat pixel. 3) We used a forward-looking infrared (FLIR) camera dataset (provided by Cook Inletkeeper, Watershed Sciences Inc. 2010, 2012) collected simultaneously with the imagery, and then we identified three unique qualitative categories of ground cover (everything not water) landscape features as light (i.e., cool) regions in the image (sand bar, exposed substrate, or barren soil), medium (low shrubs, dwarf shrubs, and herbaceous cover), and dark (i.e., warm; white spruce, black spruce, or other deciduous forests). We identified light, medium, and dark regions based upon their relative thermal signatures: they were clearly unique given the different substrate types they represented. Using the digitized polygons, we calculated the proportion of each ground cover category in a Landsat pixel. Given that Landsat 8 images were not of sufficient resolution to estimate percentage of water and ground cover values, we instead used FLIR imagery as a method for characterizing the general habitat composition at each site. Selected sites were highly unlikely to experience significant fluctuations in water cover due to the stream morphology of the Anchor River watershed, permitting the use of single estimations for habitat composition by site.

Additionally, we assessed environmental conditions at the watershed level using United States Geologic Survey (USGS) stream gage data (USGS gage number: 15239900; Figure 1; Data S3, Supplemental Material; USGS 2016) and Natural Resources Conservation Service Snotel air temperature data (site number: 1062; 59°51′34.92″N, 151°18′39.96″W; Figure 1; Data S2, Supplemental Material; Natural Resources Conservation Service 2015) which are recorded at 15-min and hourly intervals, respectively. Stream stage (i.e., height in feet above the stream bed) variables included 1) average stream stage from the 24 h preceding a given Landsat TIR pass-over date and 2) the difference between the 24-h average stage and the average stream stage over the 7 d preceding the Landsat TIR image date, henceforth referred to as gage daily and gage difference, respectively. Air temperature variables included 1) average air temperature from the 24 h preceding a given Landsat TIR pass-over date and 2) the difference between the 24-h average temperature and the average air temperature over the 7 d preceding the Landsat TIR image date, henceforth referred to as air daily and air difference, respectively. Gage difference and air difference provide context for fluctuations or stability within the river's flow regime and thermal environment in the week preceding the Landsat TIR image.

Data analysis

Prior to assessing the ability of environmental covariates to predict stream temperatures, we first compared raw surface temperatures from Landsat to in situ temperatures to explore whether we could use Landsat information, not incorporated into a regression model, as a proxy for stream temperatures (throughout the paper, we refer to Landsat information not incorporated into a regression model as raw Landsat). We then assessed the ability of Landsat 8 TIR imagery-derived and environmental variables to predict in situ stream temperature by regressing Landsat-derived temperatures, water cover, riparian land cover, Rosgen stream class, air temperature, and stream stage data against in situ stream temperature using generalized additive mixed models (GAMMs), which allow for potentially nonlinear relationships between explanatory variables and response data (Wood 2004; 2006). We treated each combination of a validation site and Landsat flyover date with usable data as a data point for subsequent temperature accuracy assessment. We constructed a global GAMM model using a combination of nonsmoothed terms (Rosgen stream class, percentage of water cover, and percentage of dark ground cover), five smoothed terms (Landsat, gage daily, gage difference, air daily, and air difference), and a normal random effect term for site. Note that we did not include light and medium ground cover as covariates given that they are perfectly collinear with percentage of water and dark ground cover (i.e., they all add up to 100%). We selected dark ground cover for inquiry because we hypothesized that dark substrate would absorb more heat and potentially have a high surface temperature. The random effect term was included to account for repeated measurements at validation sites. Given the small number of unique values across the range of smoothed covariates, we used a modest maximum basis size (i.e., k = 5) with cubic splines to model the smoothed covariates. We checked model specifications through an examination of residual plots and validated basis dimension choices using k indices and P values following Wood (2006).

We modeled in situ data as normally distributed (Gaussian GAMMs) and fitted them with the mgcv package (Wood 2001) in R (R Core Team 2020). We assessed the global model for goodness of fit and for patterns in residuals that would suggest structural problems in the model. Pending structural adequacy of the global model, we conducted multimodel inference using Akaike's information criterion corrected for small sample sizes (AICc) to assess the relationship between stream and riparian explanatory variables and stream temperature (Burnham and Anderson 2002). Models included all combinations of explanatory variables contained in the global model, without interaction terms, and with the site random effect term included in all regressions. We selected the 95% confidence set of models (i.e., sum of ordered Akaike weights is 0.95; Burnham and Anderson 2002), which we then ranked by AICc scores; we used AICc variable importance to summarize support for covariates. We summarized model goodness of fit with adjusted R2 using empirical degrees of freedom to represent the fitted model complexity penalty (Wood 2001).

To develop options for predicting stream temperatures, we tested the predictive ability of a suite of reduced GAMM models. We then implemented a leave-one-out cross-validation routine to generate root mean square error (RMSE) values as a measure of predictive performance in reconstructing stream temperatures from Landsat and environmental data. To understand whether we could use regression models to predict temperatures at new sites, leave-one-out cross-validation predictions only included the fixed-effects portion of the GAMM (i.e., without site), but uncertainty estimations (i.e., RMSE) included random draws of the random effect variance term and residual variance term.

Given that the covariates we explored were either recorded at the local level (i.e., site-specific) or at the regional level (i.e., at one location for the watershed), models with alternative compositions of covariates have variable abilities to estimate spatially explicit vs. average watershed stream temperatures. Regional-level covariates included gage daily, gage difference, air daily, and air difference, while we recorded all other covariates at each site. As such, GAMMs built only from regional-scale data (i.e., no site-level covariates) can only estimate average stream temperature across the region. On the other hand, we can potentially use GAMMs including site-level covariates (e.g., Landsat, dark ground cover, etc.) to predict local stream temperatures.

To understand whether the accuracy of Landsat 8 TIR imagery-derived water temperatures varies with environmental conditions, we quantified the accuracy of Landsat temperatures as the difference between the remotely sensed Landsat temperature measurement, TR and the in situ stream temperature, TS, at a point in space and time, henceforth referred to as Landsat thermal offset (LTO): LTO = TRTS. We again constructed a global GAMM model, but with LTO as the response variable and used a combination of nonsmoothed terms (Rosgen stream class, percentage of water cover, and percentage of dark ground cover), smoothed terms (gage daily, gage difference, air daily, and air difference), and a normal random effect term for site. As before, we compared models using AICc with all combinations of explanatory variables contained in the global model, without interaction terms, and with the site random effect term included in all regressions. Using a 95% confidence model set, as ranked by AICc scores, we examined AICc variable importance to summarize support for covariates that may explain the accuracy of Landsat 8 TIR imagery-derived water temperatures.

Results

Objective 1. Quantify the ability of Landsat 8 TIR imagery and local and regional environmental data to predict in situ water temperatures

Results revealed that raw surface temperatures from Landsat failed to accurately represent stream temperature across the sampling period (Figure 2). The propensity for Landsat information to mischaracterize observed stream temperatures increased with rising temperatures, and the overall RMSE for Landsat compared to in situ temperature was high, with a RMSE of 5.3°C. Next, we incorporated environmental covariates into a global model. We checked model specifications via an examination of k indices and P values for all smoothed terms, which indicated the basis dimension for smooth terms (k = 5) was adequate (Landsat: P = 0.75, k index = 1.12; gage daily: P = 0.67, k index = 1.10; gage difference: P = 0.83, k index = 1.15; air daily: P = 0.76, k index = 1.12; air difference: P = 0.73, k index = 1.09). The global model had a high goodness of fit ( = 0.75) and lacked obvious patterns in residuals (Figure S1, Supplemental Material); thus, we proceeded with AICc multimodel inference.

Figure 2.

Stream in situ temperature plotted across Landsat 8 thermal infrared imagery collection dates and all recorded variables used to explore potential predictors of stream temperature. Covariates displayed in this plot represent variable temporal and spatial resolutions: percentage of water cover, ground cover variables, and Rosgen stream class are spatially explicit, but time invariant (i.e., estimates taken once over the course of the study). We captured air and stream gage variables at a single location in 2015 but are temporally explicit, such that they are continually recorded throughout the study. Finally, Landsat is both space- and time-explicit.

Figure 2.

Stream in situ temperature plotted across Landsat 8 thermal infrared imagery collection dates and all recorded variables used to explore potential predictors of stream temperature. Covariates displayed in this plot represent variable temporal and spatial resolutions: percentage of water cover, ground cover variables, and Rosgen stream class are spatially explicit, but time invariant (i.e., estimates taken once over the course of the study). We captured air and stream gage variables at a single location in 2015 but are temporally explicit, such that they are continually recorded throughout the study. Finally, Landsat is both space- and time-explicit.

Results indicated support for models of in situ temperatures that included the following environmental covariates: air daily, gage daily, and gage difference (adjusted R2 = 0.716, RMSE = 1.32, effective degrees of freedom = 12.26 as estimated from the GAMM fit, which included penalizations for the intercept and both fixed and smoothed terms; Wood 2006). The AICc variable importance scores for air daily, gage daily, and gage difference were 0.82, 0.71, and 0.59, respectively. The top model did not include Rosgen stream class, percentage of dark ground cover, percentage of water cover, air difference, and Landsat. Overall, the model performed well (Figures S2 and S3, Supplemental Material) and revealed that in situ temperatures generally increased with rising daily air temperature (P < 0.001), with declining gage daily (P < 0.001), and with rising gage difference (P = 0.030; Figure S2, Supplemental Material). Further examination revealed moderate concurvity (i.e., the GAMM equivalent of collinearity) between air daily, gage daily, and gage difference covariates (concurvity estimates: 0.61–0.86), suggesting that either covariate could possibly be used by practitioners to predict stream temperatures.

To further explore the potential for predicting stream temperature with environmental variables and Landsat data, we analyzed a subset of three additional simple models using AICc-supported covariates and Landsat, a non–AICc-supported covariate, (plus site included as a random effect) to assess the degree to which stream temperatures could be predicted with different forms of information available to practitioners: top model + Landsat, air daily only, Landsat only. The expanded top model including the addition of Landsat (air daily + gage daily + gage difference + Landsat) as a covariate explained in situ water temperatures only slightly better than the top model without Landsat (adjusted R2 = 0.749, RMSE = 1.15, effective degrees of freedom = 17.23; see Figures S4 and S5, Supplemental Material). A simple model with just air daily as a covariate performed similarly to the top model (adjusted R2 = 0.716, RMSE = 1.31, effective degrees of freedom = 12.83; see Figure S6 and S7, Supplemental Material). Finally, while Landsat was strongly correlated with air daily in our dataset (concurvity estimate = 0.70), a simple model with just Landsat as a covariate explained less variation in water temperature compared to the top model + Landsat or the air daily only models we explored (adjusted R2 = 0.662, RMSE = 1.42, effective degrees of freedom = 12.94; see Figures S8 and S9, Supplemental Material).

Leave-one-out cross-validation revealed relatively strong predictive performance of the GAMM run with the three covariates air daily, gage daily, and gage difference (RMSE = 2.4°C), while the addition of Landsat did not add any predictive power (RMSE = 2.4°C). The models with individual covariates performed comparably well (air daily only; RMSE = 2.5°C, Landsat only; RMSE = 2.6°C), indicating that average watershed-scale temperatures can be reconstructed using regional data sources with reasonable accuracy. Generally, these models suggest that GAMMs with the covariate air daily appear to provide reasonable estimates of average in situ stream temperatures (Figures 3b–3d). Importantly, these models successfully account for the temporal patterns in stream temperatures at the watershed scale, while the raw surface temperatures from Landsat (i.e., not incorporated in a regression model) did not accurately represent stream temperature across the sampling period. (Figure 3a). However, likely due to the strong correlation between daily air temperature and Landsat data (R = 0.80), Landsat data incorporated into a regression model can also provide reasonable estimates of stream temperature (Figure 3e).

Figure 3.

In 2015, we explored whether we could use various environmental variables and Landsat data to estimate the temperatures of small streams. Here we assessed the predictive accuracy of generalized additive mixed models in estimating water temperatures for small streams as estimated using leave-one-out cross-validation. Observed site-level temperatures and predicted site-level temperatures are linked by thin gray lines. Thick lines intersect the mean temperature value (averaged across sites) at each date. (a) Raw surface temperatures from Landsat (blue line) compared to in situ temperatures (black line). (b–e) In situ temperatures (black line) compared to leave-one-out cross-validation model predictions for a subset of four GAMMs (red line). All models included site (Site) as a random effect. Points are jittered along the x-axis to show overlapping in situ vs. model-predicted temperatures. Thin black lines connect points at each site by date.

Figure 3.

In 2015, we explored whether we could use various environmental variables and Landsat data to estimate the temperatures of small streams. Here we assessed the predictive accuracy of generalized additive mixed models in estimating water temperatures for small streams as estimated using leave-one-out cross-validation. Observed site-level temperatures and predicted site-level temperatures are linked by thin gray lines. Thick lines intersect the mean temperature value (averaged across sites) at each date. (a) Raw surface temperatures from Landsat (blue line) compared to in situ temperatures (black line). (b–e) In situ temperatures (black line) compared to leave-one-out cross-validation model predictions for a subset of four GAMMs (red line). All models included site (Site) as a random effect. Points are jittered along the x-axis to show overlapping in situ vs. model-predicted temperatures. Thin black lines connect points at each site by date.

We explored leave-one-out cross-validation prediction residuals for each model by validation sites to assess the ability of alternative models to predict spatially explicit local-scale temperatures (as opposed to predicting an overall mean watershed-scale temperature). Results indicate that site-specific covariates did not improve local-scale stream temperature prediction performance (Figure 4). While at some sites, prediction accuracy increased with the addition of Landsat data, such as at site CR-2 where the inclusion of Landsat as a site-level covariate to the model “air daily + gage daily + gage difference” reduced prediction error closer to zero, these improvements were not consistent across space, indicating poor predictability for spatially explicit small-order stream temperatures using Landsat data or other local-scale environmental variables tested here.

Figure 4.

To understand the predictive accuracy of models with various environmental variables and Landsat data, in 2015, we inspected the boxplots of residuals for each generalized additive mixed model in estimating water temperatures for small streams. Each panel shows results for unique sites. Colors correspond to different models. Sites labeled SF are on the South Fork Anchor River, NF are on the North Fork Anchor River, and CR are on the Chakok River.

Figure 4.

To understand the predictive accuracy of models with various environmental variables and Landsat data, in 2015, we inspected the boxplots of residuals for each generalized additive mixed model in estimating water temperatures for small streams. Each panel shows results for unique sites. Colors correspond to different models. Sites labeled SF are on the South Fork Anchor River, NF are on the North Fork Anchor River, and CR are on the Chakok River.

Objective 2: Determine factors that influence the accuracy of Landsat 8 TIR imagery

LTO values ranged from −6.4 to 13.1°C, with a mean of 3.0°C and a standard deviation of ± 4.4°C. The mean LTO across validation sites was highest on June 15, 2015 (8.5°C) and closest to zero on July 1, 2015 (0.1°C; Figure 3). The accuracy (i.e., LTO) of Landsat-derived temperatures worsened as Landsat-derived temperatures increased (Figure 5). We then incorporated environmental covariates into a global model. We checked model specifications via an examination of k indices and P values for all smoothed terms, which indicated the basis dimension for smooth terms (k = 5) was likely adequate (gage daily: P = 0.34, k index = 0.94; gage difference: P = 0.53, k index = 1.02; air daily: P = 0.26, k index = 0.94; air difference: P = 0.29, k index = 0.93). The global model had a high goodness of fit ( = 0.79) and lacked obvious patterns in residuals (Figure S10, Supplemental Material); thus, we proceeded with AICc multimodel inference. Assessment of this global GAMM revealed that local and regional covariates also predicted LTO.

Figure 5.

In 2015, we explored that the accuracy of Landsat derived water temperatures in small streams. Here, we assessed accuracy as the Landsat thermal offset (i.e., LTO), plotted across Landsat 8 thermal infrared imagery collection dates and all recorded environmental variables. Covariates displayed in this plot represent variable temporal and spatial resolutions: percentage of water cover, ground cover variables, and Rosgen stream class are spatially explicit, but time invariant (i.e., estimates taken once over the course of the study). We captured air and stream gage variables at a single location but are temporally explicit, such that they are continually recorded throughout the study. Finally, Landsat is both space- and time-explicit.

Figure 5.

In 2015, we explored that the accuracy of Landsat derived water temperatures in small streams. Here, we assessed accuracy as the Landsat thermal offset (i.e., LTO), plotted across Landsat 8 thermal infrared imagery collection dates and all recorded environmental variables. Covariates displayed in this plot represent variable temporal and spatial resolutions: percentage of water cover, ground cover variables, and Rosgen stream class are spatially explicit, but time invariant (i.e., estimates taken once over the course of the study). We captured air and stream gage variables at a single location but are temporally explicit, such that they are continually recorded throughout the study. Finally, Landsat is both space- and time-explicit.

Results indicated that a model with air difference and Rosgen stream class (relative variable importance = 0.85 and 0.74, respectively) best explained LTO (adjusted R2 = 0.789, RMSE = 1.32, effective degrees of freedom = 8.49; Figures S11 and S12, Supplemental Material). Percentage of dark ground cover, percentage of water cover, air daily, gage difference, and gage daily were not included in this top model. Results suggest that the accuracy of raw Landsat-derived temperatures declines with air temperature variability (P < 0.001) and is worse for Rosgen class E (low slope reaches with high channel sinuosity) type streams (P = 0.005).

Discussion

A paucity of stream temperature data exists for remote watersheds, particularly for small streams, creating challenges for researchers and practitioners that seek to understand the implications of a changing climate and anthropogenic influences on freshwater ecosystems. Results from this study indicate that raw Landsat-based measurements generally overpredict water temperatures in temperate small streams and may not be suitable as proxies for stream temperature without being incorporated into regression models. Here, validation work using generalized additive modeling and regionally specific air temperature data, independently of Landsat information, was able to capture watershed-scale changes in small-stream water temperature. We also found that temperatures derived from satellite-based thermal imagery tracked closely with regional air temperatures. Therefore, researchers could exchange air temperature data for satellite-derived temperatures, which they could also integrate into a modeling framework to estimate stream temperature at the watershed scale, albeit with reduced performance compared to models with air temperature. Practitioners should apply caution, as both meteorological and hydrological conditions appeared to influence the accuracy of Landsat temperatures. Importantly, however, local-scale predictions were more variable and did not improve with the use of Landsat or other site-specific covariates. Given that the data used here were insufficient for reconstructing spatially explicit stream temperatures, there is a need to identify other data sources or modeling approaches that could help improve prediction performance to enable small-stream temperature reconstruction at finer resolutions.

Results here indicated that several processes combine to influence the accuracy of model-based estimations of water temperatures for small streams. Riverine water temperatures represent a balance between absorption of heat from solar radiation, heat exchange with air and land, evaporative cooling, and the temperature of water inputs and hyporheic flow (Brown 1969; Sinokrot and Stefan 1993; Evans and Petts 1997; Webb and Zhang 1997). As streams move through the landscape, water mixes through turbulent flow and exchanges heat with the air and riparian environment (Webb and Zhang 1997; Torgersen et al. 2001), which likely drives the equilibration of temperatures between land and water. Accordingly, we found that air temperature, even at the regional scale, offers a good proxy for subsurface temperatures in small stream systems. However, practitioners should consider their study objectives and the associated stream temperature estimation error that they are willing to accept. Even the best performing model resulted in a root mean square error of 2.4°C. While this may be acceptable for studies exploring broad trends, this may misalign with the resolution of temperature estimations required of ecological research on the thermal tolerances of freshwater fishes.

The failure of Landsat data to add predictive power and spatial accuracy to these models is likely, in part, attributable to the fact that the resolution of Landsat 8 pixels is 30 m, and that these remotely sensed temperatures represent a mix of stream and riparian habitat surface temperatures. This was supported through our assessment of the accuracy of Landsat-derived temperatures, which were moderately correlated with the percentage of dark ground cover among pixel composition variables. This may indicate that conductive heat exchange is a driver of stream temperatures in this high-latitude system. Additionally, the geographic positioning of Landsat images changes slightly with each pass, such that the influence of ground and water cover may also vary over time. Given that our approach using high-resolution FLIR imagery to estimate pixel composition before the study period would have missed this variability, other methodologies may find increased predictive power of ground cover and water cover variables.

Additionally, we found that when the air temperature was stable, LTO was low, which we interpret as an indication of equilibration between land and water temperatures under stable environmental conditions. For example, on the July 1, 2015, sampling date, the air temperature was extremely stable in the week before the Landsat flyover (Figure S13, Supplemental Material), potentially leading to highly accurate stream temperature predictions with a mean LTO of 0.1°C across validation sites. Further complicating the relationship between satellite-derived and in situ temperatures is that remotely sensed thermal imagery only captures surface temperatures. The effect of thermal stratification is typically minimal in small streams, where surface waters are warmer than deeper waters, but this may have contributed to these discrepancies (Torgersen et al. 2001; Handcock et al. 2006). For example, we found study reaches classified as type-E Rosgen stream class (low slope reaches with high channel sinuosity) had the highest average LTO values over the study period (Figure 5), possibly reflecting water stratification and surface warming effects in these proportionally deeper and narrower stream habitats. These study reaches were in the Chakok River, reported by previous studies to maintain depths up to 1.5 m, potentially allowing for thermal stratification (Cieutat et al. 1991). Exacerbating stream water stratification, low stream flows reduce stream mixing and can allow thermal boundaries to set up more widely in lotic habitats, particularly if coupled with air temperature changes (Torgersen et al. 2001). During this study, peak daily air temperature on a Landsat flyover date occurred on June 15 (18.4°C). This high temperature is uncharacteristic for temperate climates such as the Anchor River basin, which has a mean historic air temperature of 14°C for June 15 (NWS 2016). Rapid air warming during this week, concomitant with changes in water flow, may have contributed to the highest mean LTO observed, 8.5°C (Figure S14, Supplemental Material). Importantly, however, regression models were able to capture this spike in regional stream temperatures, highlighting the value of modeling approaches for estimating watershed-scale stream temperature fluctuations.

Overall, results from the Anchor River test watershed provide insight for practitioners who wish to estimate water temperatures for small streams. Our study demonstrates that regression modeling approaches utilizing stream stage and air temperature data, which are frequently available via archival instrumentation stations, may be useful for predicting the average subsurface temperature of small streams in a watershed. If this information is not available, Landsat information may also be suitable if practitioners are willing to accept reduced estimation accuracy. However, the margin of error of Landsat TIR imagery–derived temperatures for small streams appears to be highly context-dependent. Results from the Anchor River indicated that Landsat TIR imagery predictions of small-stream temperatures were most accurate during periods of air temperature stability associated with consistent weather conditions. We caution that practitioners need additional validation efforts to understand under what conditions raw Landsat TIR imagery provides a valid characterization of stream temperatures, given the temporal and spatial limitations of this case study.

Developing methodologies to assess water temperatures is critical to our understanding of freshwater ecosystems and will aid in our ability to monitor the effects of natural and anthropogenic stressors across landscapes. Continued validation efforts in other small-stream systems will be important for testing the robustness of methods to use environmental data and Landsat TIR imagery-based to estimate small stream temperatures. While results presented herein for the Anchor River watershed suggest that remotely sensed data could potentially be integrated into regression models depending upon the research question's required margin of error, assessing the accuracy of Landsat TIR imagery for predicting small-stream temperatures will benefit from future efforts to understand whether validation efforts can be applied across time, seasons, or watersheds of similar latitude and habitat composition. Finally, results here assessed stream temperatures during the late spring through early fall periods. Given that thermal stratification is likely to occur most during the summer season when flow is reduced and air temperatures are high (Nielsen et al. 1994), future studies that assess the effects of these covariates on LTO across an annual cycle will further improve our understanding about the use of remotely sensed temperature in small streams.

Supplemental Material

Data S1. The final dataset of site-specific and regional environmental variables, Landsat-derived temperatures, and in situ temperatures in the Anchor River watershed, southcentral Alaska, 2015.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S1 (8 KB CSV).

Data S2. Air temperature data from the regional sensor of the Anchor River watershed via Natural Resources Conservation Service Snotel (site number: 1062; 59°51′34.92″N, Longitude −151°18′39.96″W) air temperature data collection site, 2015.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S2 (72 KB CSV).

Data S3. Stream gage height data from the regional sensor of the Anchor River watershed via U.S. Geological Survey stream discharge gage in 2015 (gage number: 15239900; 59°26′42″N, 151°27′3.96″W).

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S3 (318 KB CSV).

Figure S1. Residual diagnostic plots for the global model used to explore potential predictors of in situ temperatures in small streams in Alaska in 2015. The global generalized additive mixed model used combination of nonsmoothed terms (Rosgen stream class, percentage of water cover, and percent of dark ground cover), five smoothed terms (Landsat, gage daily, gage difference, air daily, and air difference), and a normal random effect term for site.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S4 (48 KB JPG).

Figure S2. The top-ranked generalized additive mixed model (according to Akaike's information criterion corrected for small sample sizes) output with response variable, in situ temperature of small streams in the Anchor River watershed, southcentral Alaska in 2015 (air daily, gage daily, and gage difference included as covariates plus site included as a random effect) displaying marginal relationships between (a) in situ temperatures and air daily, (b) in situ temperatures and gage daily, (c) in situ temperatures and gage difference, and (d) model-predicted in situ temperatures compared to observed in situ temperatures.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S5 (85 KB JPG).

Figure S3. Residual diagnostic plots for the top-ranked generalized additive mixed model (according to Akaike's information criterion corrected for small sample sizes) with response variable, in situ temperature of small streams in the Anchor River watershed, southcentral Alaska in 2015 (air daily, gage daily, and gage difference included as covariates plus site included as a random effect).

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S6 (53 KB JPG).

Figure S4. The generalized additive mixed model output with response variable, in situ temperature of small streams in the Anchor River watershed, southcentral Alaska, 2015, and air daily, gage daily, gage difference, and Landsat included as covariates plus site included as a random effect, displaying marginal relationships between (a) in situ temperatures and air daily, (b) in situ temperatures and gage daily, (c) in situ temperatures and gage difference, (d) in situ temperatures and Landsat, and (e) model-predicted in situ temperatures compared to observed in situ temperatures.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S7 (161 KB JPG).

Figure S5. Residual diagnostic plots for a generalized additive mixed model with response variable, in situ temperature of small streams in the Anchor River watershed, southcentral Alaska, in 2015, and air daily, gage daily, gage difference, and Landsat included as covariates plus site included as a random effect.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S8 (54 KB JPG).

Figure S6. The generalized additive mixed model output with response variable, in situ temperature of small streams in the Anchor River watershed, southcentral Alaska, in 2015, and air daily included as a covariate plus site included as a random effect, displaying marginal relationships between (a) in situ temperatures and air daily and (b) model-predicted in situ temperatures compared to observed in situ temperatures.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S9 (67 KB JPG).

Figure S7. Residual diagnostic plots for a generalized additive mixed model with response variable, in situ temperature of small streams in the Anchor River watershed, southcentral Alaska, in 2015, and air daily included as a covariate plus site included as a random effect.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S10 (58 KB JPG).

Figure S8. The generalized additive mixed model output with response variable, in situ temperature of small streams in the Anchor River watershed, southcentral Alaska, in 2015, and Landsat included as a covariate plus site included as a random effect, displaying marginal relationships between (a) in situ temperatures and Landsat and (b) model-predicted in situ temperatures compared to observed in situ temperatures.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S11 (71 KB JPG).

Figure S9. Residual diagnostic plots for a generalized additive mixed model with response variable, in situ temperature of small streams in the Anchor River watershed, southcentral Alaska, in 2015, and Landsat included as a covariate plus site included as a random effect.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S12 (70 KB JPG).

Figure S10. Residual diagnostic plots for the global model used to explore potential predictors of Landsat thermal offset (LTO; i.e., the accuracy of Landsat data not incorporated into a regression model) in small streams in Alaska, in 2015. The global generalized additive mixed model used a combination of nonsmoothed terms (Rosgen stream class, percentage of water cover, and percentage of dark ground cover), four smoothed terms (gage daily, gage difference, air daily, and air difference), and a normal random effect term for site.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S13 (59 KB JPG).

Figure S11. The top-ranked generalized additive mixed model (according to Akaike's information criterion corrected for small sample sizes) output with response variable, Landsat thermal offset (LTO; air difference and Rosgen stream class included as covariates plus site included as a random effect) displaying marginal relationships between (a) in situ temperatures of small streams in the Anchor River watershed, southcentral Alaska and air difference, (b) in situ temperatures and Rosgen stream class, and (c) model-predicted in situ temperatures compared to observed in situ temperatures. Data was collected in 2015.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S14 (102 KB PNG).

Figure S12. Residual diagnostic plots for the top-ranked generalized additive mixed model (according to Akaike's information criterion corrected for small sample sizes) with response variable, Landsat thermal offset (LTO; i.e., the accuracy of Landsat data not incorporated into a regression model) used to explore potential predictors of LTO in small streams in Alaska, in 2015. Covariates included Rosgen stream class and air difference plus site included as a random effect).

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S15 (62 KB JPG).

Figure S13. We explored whether we could use various environmental variables and Landsat data to estimate the temperatures of small streams in the Anchor River watershed, southcentral Alaska, in 2015. Here we have plotted the two environmental variables, stream discharge (i.e., gage height in feet) and air temperature preceding the July 1, 2015, Landsat 8 thermal infrared (TIR) imagery collection.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S16 (930 KB TIFF).

Figure S14. We explored whether we could use various environmental variables and Landsat data to estimate the temperatures of small streams in the Anchor River watershed, southcentral Alaska, in 2015. Here we have plotted the two environmental variables, stream discharge (i.e., gage height in feet) and air temperature preceding the June 15, 2015, Landsat 8 thermal infrared (TIR) imagery collection.

Found at DOI: https://doi.org/10.3996/JFWM-20-048.S17 (948 KB TIFF).

Acknowledgments

The fieldwork, in situ data collection and management, Landsat image processing and early drafts of this work were conducted by J. Hagan and T. Smeltz as part of Hagan's graduate thesis research. Jason Geck (Alaska Pacific University [APU]) provided critical analytical and technical support for Landsat image processing and GIS construction. Funding support was provided by the At-Sea-Processors Association, the Alaska Space Grant Program, and the United States Fish and Wildlife Service–Western Alaska Landscape Conservation Cooperative. Technical and scientific support came from Sue Mauger (Cook InletKeeper), Mike Booz (Alaska Department of Fish and Game), Dr. Mike Loso (APU), Dr. Nathan Wolf (APU), Aileen Nimick (APU), Sarah Webster (APU), Dr. Chris McGonigle (Ulster University), Jay Calvert (Ulster University), Tom Harris (Anchor Point resident), and Dr. Rebecca Handcock (Murdoch University). We also thank the reviewers and editors of this manuscript, all of whom provided valuable feedback that resulted in a stronger manuscript.

Any use of trade, product, website, or firm names in this publication is for descriptive purposes only and does not imply endorsement by the U.S. Government.

References

Allan
JD,
Castillo
MM.
2007
.
Stream ecology: structure and function of running waters
.
Dordrecht, Netherlands
:
Springer Science & Business Media
.
Armour
CL.
1991
.
Guidance for evaluating and recommending temperature regimes to protect fish
.
Washington, D.C
.:
US Department of the Interior, Fish and Wildlife Service
.
Benyahya
L,
Caissie
D,
St-Hilaire
A,
Ouarda
T,
Bobée
B.
2007
.
A review of statistical water temperature models. A review of statistical water temperature models
.
Canadian Water Resources Journal
32
:
179
192
.
Brown
GW.
1969
.
Predicting temperatures of small streams
.
Water Resources Research
5
:
68
75
.
Buisson
L,
Blanc
L,
Grenouillet
G.
2008
.
Modelling stream fish species distribution in a river network: the relative effects of temperature versus physical factors
.
Ecology of Freshwater Fish
17
:
244
257
.
Burnham
KP,
Anderson
DR.
2002
.
Model selection and multimodel inference: a practical information-theoretic approach
.
New York
:
Springer Science & Business Media
.
Caissie
D,
El-Jabi
N,
Satish
M.
2001
.
Modelling of maximum daily water temperatures in a small stream using air temperatures
.
Journal of Hydrology
251
:
14
28
.
Caissie
D,
El-Jabi
N,
St-Hilaire
A.
1998
.
Stochastic modelling of water temperatures in a small stream using air to water relations
.
Canadian Journal of Civil Engineering
25
:
250
260
.
Cherkauer
KA,
Burges
SJ,
Handcock
RN,
Kay
JE,
Kampf
SK,
Gillespie
AR.
2005
.
Assessing satellite-based and aircraft-based thermal infrared remote sensing for monitoring Pacific Northwest river temperature
.
Journal of the American Water Resources Association
41
:
1149
1159
.
Cieutat
BA,
Goldfarb
RJ,
Bradley
DC,
Roushey
BH.
1991
.
Placer gold of the Kenai lowland
.
Pages
23
29
in
DC,
Bradley
Dusel-Bacon,
C,
editors.
Geologic studies in Alaska. Fairbanks, Alaska: U.S. Geological Survey Bulletin 2041
.
Cristea
NC,
Burges
SJ.
2009
.
Use of thermal infrared imagery to complement monitoring and modeling of spatial stream temperatures
.
Journal of Hydrologic Engineering
14
:
1080
1090
.
Cunningham
CJ,
Westley
PA,
Adkison
MD.
2018
.
Signals of large scale climate drivers, hatchery enhancement, and marine factors in Yukon River Chinook Salmon survival revealed with a Bayesian life history model
.
Global Change Biology
24
:
4399
4416
.
Despini
F,
Teggi
S.
2013
.
Analysis of temperature maps of waterbodies obtained from ASTER TIR images
.
International Journal of Remote Sensing
34
:
3636
3653
.
De Staso
J
III,
Rahel
FJ.
1994
.
Influence of water temperature on interactions between juvenile Colorado River Cutthroat Trout and Brook Trout in a laboratory stream
.
Transactions of the American Fisheries Society
123
:
289
297
.
Dugdale
SJ.
2016
.
A practitioner's guide to thermal infrared remote sensing of rivers and streams: recent advances, precautions and considerations
.
Wiley Interdisciplinary Reviews: Water
3
:
251
268
.
ESRI.
2014
.
Arcgis desktop: Release 10.3.2
.
Redlands, CA
:
ESRI
.
Evans
E,
Petts
GE.
1997
.
Hyporheic temperature patterns within riffles
.
Hydrological Sciences Journal
42
:
199
213
.
Griffiths
JR,
Schindler
DE.
2012
.
Consequences of changing climate and geomorphology for bioenergetics of juvenile sockeye salmon in a shallow Alaskan lake
.
Ecology of Freshwater Fish
21
:
349
362
.
Gustafson
WT,
Handcock
RN,
Gillespie
AR,
Tonooka
H.
2003
.
Image sharpening method to recover stream temperatures from aster images
.
Remote Sensing for Environmental Monitoring, GIS Applications, and Geology II
4886
:
72
84
.
Handcock
R,
Gillespie
A,
Cherkauer
K,
Kay
J,
Burges
S,
Kampf
S.
2006
.
Accuracy and uncertainty of thermal-infrared remote sensing of stream temperatures at multiple spatial scales
.
Remote Sensing of Environment
100
:
427
440
.
Handcock
RN,
Torgersen
CE,
Cherkauer
KA,
Gillespie
AR,
Tockner
K,
Faux
RN,
Tan
J,
Carbonneau
P.
2012
.
Thermal infrared remote sensing of water temperature in riverine landscapes
.
Fluvial Remote Sensing for Science and Management
12
:
85
113
.
Hick
P,
Carlton
M.
1991
.
Practical applications of airborne multispectral scanner data for forest, agriculture, and environmental monitoring
.
Pages
60
61
in
24th International Symposium on Remote Sensing of the Environment.
Hoboware.
2014
.
Hoboware pro for desktop. Version 5.1 ed
.
Bourne, Massachussetts
:
Onset Corporation 2001
.
Hodgson
S,
Quinn
TP.
2002
.
The timing of adult sockeye salmon migration into fresh water: adaptations by populations to prevailing thermal regimes
.
Canadian Journal of Zoology
80
:
542
555
.
Holtby
LB.
1988
.
Effects of logging on stream temperatures in Carnation Creek British Columbia, and associated impacts on the Coho Salmon (Oncorhynchus kisutch)
.
Canadian Journal of Fisheries and Aquatic Sciences
45
:
502
515
.
Isaak
DJ,
Luce
CH,
Rieman
BE,
Nagel
DE,
Peterson
EE,
Horan
DL,
Parkes
S,
Chandler
GL.
2010
.
Effects of climate change and wildfire on stream temperatures and salmonid thermal habitat in a mountain river network
.
Ecological Applications
20
:
1350
1371
.
Isaak
DJ,
Wenger
SJ,
Peterson
EE,
Ver Hoef
JM,
Nagel
DE,
Luce
CH,
Hostetler
SW,
Dunham
JB,
Roper
BB,
Wollrab
SP.
2017
.
The norwest summer stream temperature model and scenarios for the Western US: a crowd-sourced database and new geospatial tools foster a user community and predict broad climate warming of rivers and streams
.
Water Resources Research
53
:
9181
9205
.
Johnson
SL.
2003
.
Stream temperature: scaling of observations and issues for modelling
.
Hydrological Processes
17
:
497
499
.
Kasprak
A,
Hough-Snee
N,
Beechie
T,
Bouwes
N,
Brierley
G,
Camp
R,
Fryirs
K,
Imaki
H,
Jensen
M,
O'Brien
G.
2016
.
The blurred line between form and process: a comparison of stream channel classification frameworks
.
PLoS ONE
11
:
e0150293
.
Kay
JE,
Kampf
SK,
Handcock
RN,
Cherkauer
KA,
Gillespie
AR,
Burges
SJ.
2005
.
Accuracy of lake and stream temperatures estimated from thermal infrared images
.
Journal of the American Water Resources Association
41
:
1161
1175
.
Keys
R.
1981
.
Cubic convolution interpolation for digital image processing
.
IEEE Transactions on Acoustics, Speech, and Signal Processing
29
:
1153
1160
.
Lalot
E,
Curie
F,
Wawrzyniak
V,
Baratelli
F,
Schomburgk
S,
Flipo
N,
Piegay
H,
Moatar
F.
2015
.
Quantification of the contribution of the beauce groundwater aquifer to the discharge of the Loire River using thermal infrared satellite imaging
.
Hydrology and Earth System Sciences
19
:
4479
4492
.
LeBosquet
M,
Tsivoglou
EC.
1950
.
Simplified dissolved oxygen computations
.
Sewage and Industrial Wastes
221054
1061
.
Ling
F,
Foody
GM,
Du
H,
Ban
X,
Li
X,
Zhang
Y,
Du
Y.
2017
.
Monitoring thermal pollution in rivers downstream of dams with landsat etm+ thermal infrared images
.
Remote Sensing
9
:
1175
.
Lisi
PJ,
Schindler
DE,
Bentley
KT,
Pess
GR.
2013
.
Association between geomorphic attributes of watersheds, water temperature, and salmon spawn timing in Alaskan streams
.
Geomorphology
.
185
:
78
86
.
Mauger
S.
2005
.
Lower Kenai Peninsula's salmon streams: Annual water quality assessment, July 2004–September 2005. Homer Soil and Water Conservation District and Cook Inlet Keeper
.
McPhee
MV,
Leon
JM,
Wilson
LI,
Siegel
JE,
Agler
BA.
2016
.
Changing growth and maturity in western Alaskan Chinook Salmon Oncorhynchus tshawytscha, brood years 1975–2005
.
North Pacific Anadromous Fish Commission Bulletin
6
:
307
327
.
Mohseni
O,
Stefan
HG,
Erickson
TR.
1998
.
A nonlinear regression model for weekly stream temperatures
.
Water Resources Research
34
:
2685
2692
.
National Research Council.
1996
.
Upstream: salmon and society in the Pacific Northwest
.
Washington, D.C
.:
National Academies Press
.
Natural Resources Conservation Service.
2015
.
Snotel site information and reports: Site 1062
.
[NWS] National Weather Service. Archival daily air temperature: Anchor point, AK.
2016
.
Nielsen
JL,
Lisle
TE,
Ozaki
V.
1994
.
Thermally stratified pools and their use by steelhead in Northern California streams
.
Transactions of the American Fisheries Society
123
:
613
626
.
Poole
GC,
Berman
CH.
2001
.
An ecological perspective on in-stream temperature: natural heat dynamics and mechanisms of human-caused thermal degradation
.
Environmental Management
27
:
787
802
.
R Core Team.
2020
.
R: a language and environment for statistical computing, R Foundation for Statistical Computing, Vienna, Austria, 2015
.
Available: http://www.R-project.org (March 2021).
Rinella
DJ,
Booz
M,
Bogan
DL,
Boggs
K,
Sturdy
M,
Rinella
MJ.
2009
.
Large woody debris and salmonid habitat in the Anchor River basin, Alaska, following an extensive spruce beetle (Dendroctonus rufipennis) outbreak
.
Northwest Science
83
:
57
69
.
Rosgen
DL.
1994
.
A classification of natural rivers
.
Catena
22
:
169
199
.
Sinokrot
BA,
Stefan
HG.
1993
.
Stream temperature dynamics: measurements and modeling
.
Water Resources Research
29
:
2299
2312
.
Sohrabi
MM,
Benjankar
R,
Tonina
D,
Wenger
SJ,
Isaak
DJ.
2017
.
Estimation of daily stream water temperatures with a Bayesian regression approach
.
Hydrological Processes
31
:
1719
1733
.
Teggi
S.
2012
.
A technique for spatial sharpening of thermal imagery of coastal waters and of watercourses
.
International Journal of Remote Sensing
33
:
3063
3089
.
Torgersen
CE,
Faux
RN,
McIntosh
BA,
Poage
NJ,
Norton
DJ.
2001
.
Airborne thermal remote sensing for water temperature assessment in rivers and streams
.
Remote Sensing of Environment
76
:
386
398
.
[USGS] United States Geological Survey. Landsat 8 OLI (operational land imager) and TIRS (thermal infrared sensor) website.
2015
[USGS] United States Geological Survey.
2016
.
Landsat 8 (L8) data users handbook, version 2.0 Sioux Falls, South Dakota: Earth Resources Observation and Science (EROS) Center
.
[USGS] United States Geological Survey.
2019
.
Using the USGS Landsat level-1 data product
.
Watershed Sciences Inc.
2010
.
Airborne thermal infrared remote sensing: Anchor River basin, Alaska
.
Technical report submitted to Cook Inlet Keeper
,
Homer, Alaska
.
Watershed Sciences Inc.
2012
.
Airborne thermal infrared remote sensing: North Fork Anchor River
.
Technical report submitted to Cook Inlet Keeper
,
Homer, Alaska
.
Wawrzyniak
V,
Piégay
H,
Poirel
A.
2012
.
Longitudinal and temporal thermal patterns of the French Rhône River using Landsat etm+ thermal infrared images
.
Aquatic Sciences
74
:
405
414
.
Webb
B,
Zhang
Y.
1997
.
Spatial and seasonal variability in the components of the river heat budget
.
Hydrological Processes
11
:
79
101
.
Webb
BW,
Hannah
DM,
Moore
RD,
Brown
LE,
Nobilis
F.
2008
.
Recent advances in stream and river temperature research
.
Hydrological Processes: An International Journal
22
:
902
918
.
Wenger
SJ,
Isaak
DJ,
Luce
CH,
Neville
HM,
Fausch
KD,
Dunham
JB,
Dauwalter
DC,
Young
MK,
Elsner
MM,
Rieman
BE.
2011
.
Flow regime, temperature, and biotic interactions drive differential declines of trout species under climate change
.
Proceedings of the National Academy of Sciences of the United States of America
108
:
14175
14180
.
Wood
SN.
2001
.
Mgcv: GAMS and generalized ridge regression for R
.
R News
1
:
20
25
.
Wood
SN.
2004
.
Stable and efficient multiple smoothing parameter estimation for generalized additive models
.
Journal of the American Statistical Association
99
:
673
686
.
Wood
SN.
2006
.
Generalized additive models: an introduction with R
.
Boca Raton, Florida
:
Chapman and Hall/CRC
.

Appendix

Quantized and calibrated scaled digital numbers representing the multispectral image data (USGS 2019).
formula
where Lλ = Top of Atmosphere (TOA) spectral radiance (watts/[ m2 · srad · μm]); ML = band-specific multiplicative rescaling factor from the metadata (RADIANCE_MULT_BAND_x, where x is the band number); AL = band-specific additive rescaling factor from the metadata (RADIANCE_ADD_BAND_x, where x is the band number); and Qcal = quantized and calibrated standard product pixel values (digital numbers).
formula
where T = at-satellite brightness temperature (K); Lλ = TOA spectral radiance ( watts/[m2 · srad · μm]); K1 = band-specific thermal conversion constant from the metadata (K1_CONSTANT_BAND_x, where x is the band number, 10 or 11); and K2 = band-specific thermal conversion constant from the metadata (K2_CONSTANT_BAND_x, where x is the band number, 10 or 11).

The findings and conclusions in this article are those of the author(s) and do not necessarily represent the views of the U.S. Fish and Wildlife Service.

Author notes

Citation: Murphy RD Jr, Hagan JA, Harris BP, Sethi SA, Smeltz TS, Restrepo F. 2021. Can Landsat thermal imagery and environmental data accurately estimate water temperatures in small streams? Journal of Fish and Wildlife Management 12(1):12–26; e1944-687X. https://doi.org/10.3996/JFWM-2020-048

Supplementary data