• E-mail
• Print
• Comment
• Font Size
• Digg
• del.icio.us
• Discuss article

Accuracy of Lake and Stream Temperatures Estimated From Thermal Infrared Images1

Posted on: Sunday, 23 October 2005, 03:01 CDT

By Kay, Jennifer E; Kampf, Stephanie K; Handcock, Rebecca N; Cherkauer, Keith A; Et al

ABSTRACT:

Emitted thermal infrared radiation (TIR, λ = 8 to 14 m) can be used to measure surface water temperatures (top approximately 100 m). This study evaluates the accuracy of stream (50 to 500 m wide) and lake (300 to 5,000 m wide) radiant temperatures (15 to 22C) derived from airborne (MASTER, 5 to 15 m) and satellite (ASTER 90 m, Landsat ETM+ 60 m) TIR images. Applied atmospheric compensations changed water temperatures by -0.2 to +2.0C. Atmospheric compensation depended primarily on atmospheric water vapor and temperature, sensor viewing geometry, and water temperature. Agreement between multiple TIR bands (MASTER - 10 bands, ASTER - 5 bands) provided an independent check on recovered temperatures. Compensations improved agreement between image and in situ surface temperatures (from 2.0 to 1.1C average deviation); however, compensations did not improve agreement between river image temperatures and loggers installed at the stream bed (from 0.6 to 1.6C average deviation). Analysis of field temperatures suggests that vertical thermal stratification may have caused a systematic difference between instream gage temperatures and corrected image temperatures. As a result, agreement between image temperatures and instream temperatures did not imply that accurate TIR temperatures were recovered. Based on these analyses, practical accuracies for corrected TIR lake and stream surface temperatures are around 1C.

(KEY TERMS: temperature; remote sensing; lakes; rivers/streams; thermal infrared; atmospheric compensation.)

INTRODUCTION

Monitoring of stream and lake temperatures is important for water quality management, land-use monitoring, and hydrological studies. Throughout the Pacific Northwest, many water bodies are thermally impaired, with summer temperatures often exceeding water quality standards (e.g., 16C for class AA Washington State, USA, streams) (Washington State, 2003). Elevated temperatures can threaten the ecological integrity of water habitats. Salmonids in particular are highly sensitive to temperature, and their temperature tolerance changes through their life cycle (Marine and Cech, 2004). In the absence of cold water rfugia (Ebersole et al., 2003), the combination of low flows and high ambient temperatures can result in elevated stream temperatures that negatively affect salmon. Urbanization can also alter stream temperature patterns. In particular, the removal of riparian vegetation increases direct solar insolation and may elevate temperatures (Poole and Berman, 2001).

Temporal and spatial variability in stream and lake temperatures is typically monitored using instream gages. A low density of gages can, however, limit the accuracy of inferred spatial distributions of temperatures. Remotely sensed thermal infrared (TIR, 8 to 14 m) images directly measure the spatial distribution of water temperatures and they have been used extensively to estimate sea surface temperatures (SST) (Kidder and Yonder Haar, 1995; Donlon et al., 2002; Kumar et al., 2003); however, less work has been done with land surface water bodies, in part because streams and small lakes can be difficult to resolve with remotely sensed images. Thermal infrared images of lake temperatures have been used to identify habitat and restoration areas (e.g., Hk et al., 2004) and as boundary conditions for hydrological models (e.g., Schott et al, 2001). In streams, TIR images have been used to identify thermal refugia important for maintaining fish habitat (e.g., Torgersen et al., 1999), to delineate ground water discharge (Bank et al., 1996), and as a check on the performance of stream temperature models (Boyd et al, 2002).

Previous studies have shown the utility of spatial temperature distributions (Schott et al., 2001; Hk et al., 2004) but also note the importance of extensive calibration and atmospheric compensation to achieve accurate temperatures (Schott et al., 2001). Torgersen et al. (2001) evaluated the accuracy of remotely sensed stream (3 to 25 m wide) temperature measurements using very high resolution, forward looking infrared (FLIR) images (0.2 to 0.4 m/pixel) obtained from a low flying helicopter. They identified several sources of error, including reflected long wave radiation, thermal boundary layer effects, and vertical thermal stratification (especially for deeper rivers on clear, sunny days). Empirically calibrated remotely sensed surface temperatures were within 0.5C of instream kinetic temperature measurements, and remotely sensed temperatures discriminated spatial patterns at a resolution and extent unattainable with instream loggers.

The successful development of high spatial and spectral resolution TIR radiometers, such as the National Aeronautics and Space Administration's MODIS/ASTER Airborne Simulator (MASTER) (Hook et al., 2001), Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) on the EOSTerra platform (Yamaguchi et al, 1998) and ETM+ on Landsat 7 (Goward et al., 2001), provides new opportunities for remote observation of stream and lake temperatures (Table 1). MASTER'S 5 to 15 m spatial resolution and ten TIR bands make it an ideal sensor for water bodies in this study (widths greater than approximately 50 m). ASTER and Landsat ETM+ images have lower spatial resolution (90 and 60 m, respectively) and TIR spectral coverage (5 bands and 1 band), but these data cost less to users, cover more area on the ground, and are acquired on a regular basis (potentially every 16 days for polar orbiting satellites over mid-latitudes).

With this wide range of TIR sensors come many opportunities to measure the temperatures of water bodies. Before TIR data are widely used, temperature retrieval in typical land surface environments should be evaluated. Therefore, the goal of this paper is to conduct an assessment of theoretical versus practical accuracies attained for lake and stream temperature retrieval from satellite and airborne TIR images. This is accomplished by evaluating strategies for temperature retrieval, quantifying the effects of atmospheric compensation, and comparing remotely sensed temperatures to those obtained by a known standard (in situ gages). A companion paper, Cherkauer et al. (2005), evaluates the operational ability of these TIR sensors to resolve streams and their temperature trends in the Pacific Northwest.

TABLE 1. TIR Image Bands.

BACKGROUND

At TIR wavelengths (8 to 14 m), both the water surface and the atmosphere emit radiation as functions of their temperature and emissivity. When atmospheric temperature and emissivity, and water emissivity are known, TIR images can be converted to surface temperatures (see Appendix). Here, strategies for atmospheric compensation and a discussion of additional factors affecting TIR temperature retrievals are presented.

Atmospheric Compensation in the Land Surface Environment

Quantitative estimation of temperatures from TIR images requires compensation for absorption by atmospheric gases (primarily water vapor, but also ozone and carbon dioxide) and upwelling atmospheric emitted radiance (Appendix; Equation A2). Increasing atmospheric absorption (decreasing transmissivity) causes at-sensor radiant temperatures to be lower than true temperatures, whereas increasing atmospheric emission (increasing path radiance) causes atsensor radiant temperatures to exceed true temperatures. Transmissivity and path radiance vary with the path from the target to the detector, target temperature, and wavelength. Interactions between these variables can change both the magnitude and the sign of TIR atmospheric compensation.

The remote sensing community has developed numerous TIR atmospheric compensation methods. Unfortunately, some correction techniques are difficult to implement or do not apply for typical land surface water temperature studies. In split window atmospheric compensation techniques (Kidder and Vender Haar, 1995), atmospheric effects are removed using empirical relationships with in situ measurements and brightness temperature differences between strong and weak atmospheric absorption bands. Split window techniques require a ground truth temperature network (e.g., temperature measuring buoys) to develop empirical processing coefficients. Therefore, split window techniques are difficult to apply in environments where calibration temperature networks are not readily available (e.g., Coll and caselles, 1997). In addition, coefficients may vary with location, as suggested for SST retrieval by Kumar et al. (2003). Regression methods based on radiance differences between bands can be used to remove relative atmospheric effects (Young et al., 2002). However, these in-scene methods require radiative transfer models to set the absolute magnitude of the compensation. Therefore, while these techniques help remove relative atmospheric effects and refine precision across bands, they do not improve the absolute accuracy.

Given the limitations of empirical and in-scene atmospheric compensation techniques, Schott et al. (2001) and Thome et al. (1998) have promoted the use of radiative transfer models, such as moderate resolution transmission (MODTRAN) (Berk \et al., 1998), which require profiles of atmospheric temperature and water vapor. Atmospheric profiles can be obtained from radiosondes, numerical weather models, and inscene techniques. However, accurate and concurrent profiles are often difficult to find, unless radiosondes can be launched during field campaigns.

The most common source of atmospheric data is the National Oceanic and Atmospheric Administration (NOAA) operational radiosonde network. In Washington State (WA), where the present study was conducted, radiosondes are launched daily at 5:00 and 17:00 Pacific Daylight Time (PDT) on the Olympic Coast (Quillayute, Washington) and in Eastern Washington (Spokane). Given the orography of Washington and the timing of satellite overpasses (approximately 11:00 PDT for Terra, Landsat 7), the WA radiosondes are not likely to be representative of any of the field sites used in this study.

Numerical weather forecasts and reanalysis products can also be used to estimate atmospheric profiles. In WA, the Pennsylvania State University/National Center for Atmospheric Research mesoscale model (MM5) (Grell et al., 1994), run operationally by the University of Washington's Department of Atmospheric Sciences, provides estimates of atmospheric total column water vapor (TCW) and vertical profiles of temperature and water vapor every hour at 4 km spatial resolution. In this study, MM5 was the most reliable and accurate source of water vapor and temperature information concurrent with image acquisition.

Gao and Goetz (1990) provided in-scene estimates of TCW using the relative absorption between a weak (1.02 m) and strong (0.94 m) water absorption band. The MODIS radiometer, on the Terra platform with ASTER, has a near infrared water vapor retrieval for TCW (Kaufman and Gao, 1992). Differential absorption techniques for estimating TCW have been extended to hyperspectral TIR data (Gu et al., 2000). However, in-scene TCW estimation does not work for radiometers with limited spectral resolution of water absorption bands (e.g., Landsat ETM+, ASTER, MASTER).

Surface Effects and Subpixel Mixing

Once an image has been corrected for atmospheric and emissivity effects, additional sources of uncertainty remain. Remotely sensed temperatures measure the surface, or 'skin' temperature, so surface effects can influence comparisons with temperature gage measurements. Surface effects include multiple reflection (radiation emitted and reflected multiple times by adjacent scene elements), wind driven evaporative cooling, and preferential heating of the surface skin in stagnant water.

For TIR observations of water, multiple reflection can increase radiant temperatures, especially at oblique viewing geometries where specular reflectance increases (e.g., Ishiyama et al., 1995). In the near bank environment of streams, the forest canopy overhanging the stream emits radiation that is reflected off the surface of the water. Fortunately, specular reflectance is minimal at view angles less than 60 degrees from nadir (vertical), so it should not affect most TIR images.

Thermal infrared temperatures may also be affected by subpixel mixing, which occurs when nonwater materials contribute to observed radiance. The severity of subpixel mixing depends on the relative size of the water body and the detector resolution. Where subpixel mixing occurs, spectral unmixing has been used to isolate the fraction of the water leaving radiance attributed to streams or lakes (e.g., Gillespie, 1992). For comparison of TIR temperatures to in situ sensors, selecting "pure" water pixels is critical if inevitable uncertainties from unmixing are to be avoided.

Even if representative TIR temperatures are obtained, the water skin temperature they detect may differ from the bulk or "kinetic" water temperature. Hook et al. (2003) showed that differences in radiant skin temperature and bulk temperature at Lake Tahoe are related to solar heating and wind speeds. On average, daytime skin temperatures there were 0.11C cooler than daytime bulk temperatures, whereas nighttime skin temperatures were 0.46C cooler than nighttime bulk temperatures. In the flowing water of streams, thermal stratification is likely to be different than in lakes; nevertheless, skin versus bulk temperature effects are important and affect temperature comparisons in this study.

MATERIALS AND METHODS

Study Location and Data

In summer 2001 and 2002, MASTER, ASTER, and Landsat ETM+ data (Table 1) were obtained over the Green River, Yakima River, Columbia River, and several lakes south east of Seattle, WA (Figure 1, Table 2). These data provide radiant temperature measurements (T^sub rs^) of lakes and streams at a variety of spatial resolutions, TIR wavelengths, viewing geometries, and atmospheric conditions. Concurrent radiosonde observations and MM5 output provided input atmospheric values for MODTRAN simulations. During the time of image acquisition, hand held Hanna Checktemp thermometers were used to measure surface 'kinetic' temperatures in the upper 10 cm of the stream and Oakton TempTestr infrared thermometers were used to measure surface radiant temperatures. A network of instream and in- lake loggers (Onset Optic StowAway Temp and StowAway TidbiT) measured water temperatures at a depth of 10 cm in the lakes, and on the streambeds.

Figure 1. Location Map for Study Streams and Rivers in Washington State (USA) (shaded areas in main image delineate watershed boundaries, and shaded areas in the inset show the specified lakes; WA = Washington, ID = Idaho, OR = Oregon).

Image Temperature Estimation: Emissivity and Atmospheric Compensation

For each site, multiple pure water pixels were averaged to obtain a representative value of the at-sensor radiance. Laboratory measurements of water emissivity (Salisbury and D'Aria, 1992) and instrument response functions were used to obtain emissivity compensations. Atmospheric absorption and emission spectra were estimated using a personal computer version of MODTRAN, PcModWin 4.0 (Berk et al, 1998). For each image acquisition, MODTRAN model runs used: (1) concurrent radiosonde profiles (Yakima, August 28, 2001 only), (2) concurrent MM5 water vapor and temperature estimates, and (3) the midlatitude standard atmosphere (Figure 2) scaled by TCW ranging from 0.5 to 3.0 g/cm^sup 3^. MODTRAN outputs were convolved with each sensor's instrument response function and band specific transmissivity, and path radiance values were used to solve for the water temperature (Appendix: Equations A2 and A3). This gave a range of potential atmospheric compensations. For subsequent analyses, atmospheric compensations for each flight line had to meet the following criteria: (1) reasonable agreement between selected compensation TCW, radiosonde profiles, and MM5 TCW and profile predictions; (2) after compensation, all bands predicted similar temperatures ("spectral flatness"); and (3) TIR images obtained at similar times and viewing geometries had similar compensation.

TABLE 2. TIR Images Used in This Study.

Figure 2. Mid-Latitude Summer Atmospheric Profile and Approximate Height Above Sea Level (ASL) of MASTER, ASTER and Landsat ETM+.

Evaluating the Accuracy of Image Temperatures

The accuracy of remotely sensed temperatures was evaluated by comparing persistent (varying by less than 1C in 30 minutes), concurrent (within 10 minutes of overpass), and repeated field measurements of surface kinetic (T^sub k-surf^), surface radiant (T^sub r-surf^), and gage (T^sub k-gage^) temperature with uncorrected and corrected image temperatures.

RESULTS

Sensitivity of Atmospheric and Emissivity Compensation Parameters

Emissivity (Salisbury and D'Aria, 1992), transmissivity, and path radiance values convolved to all ASTER, MASTER, and Landsat ETM+ TIR bands for typical conditions are shown in Figure 3. Emissivity values range from 0.97 at 13 m to 0.99 at 11 m (Figure 3a). Transmissivity and path radiance values are inversely related. For example, transmissivity spectra (Figure 3b) are generally concave downward with values ranging from 20 percent to more than 90 percent, whereas path radiance spectra (Figure 3c) are concave upward with values ranging from 0.5 W/m^sup 2^/sr to more than 4 W/ m^sup 2^/sr. From Figure 3, it is clear that increasing water vapor decreases transmissivity and increases path radiance. MASTER band 45 (9.8 m) has a slight depression (elevation) in transmissivity (path radiance) due to ozone. When estimating TIR image temperatures for each sensor, bands that are least sensitive to emissivity and atmospheric compensation were used ("recommended bands") (Table 1).

Figure 3. Typical Emissivity and Atmospheric Compensation Values: (a) Water Emissivity Values for MASTER, ASTER and Landsat ETM+ Band Effective Wavelengths (λ^sub eff^) (Salisbury and D'Aria, 1992); (b) Atmospheric Transmissivity for MASTER λ^sub eff^; and (c) Atmospheric Path Radiance for MASTER λ^sub eff^. Transmissivity and path radiance values are from MODTRAN simulations for 15 m August 25, 2001, MASTER images.

Table 3 quantifies the effect of wavelength, TCW, water temperature, sensor height above sea level (ASL), surface height ASL, and sensor viewing geometry on the magnitude of atmospheric compensation. The sensitivity of atmospheric compensation to a single parameter of interest was evaluated by keeping all other variables constant at "base case" values (wavelength in the middle of the atmospheric window: MASTER Band 44 -9.2 m; 2 g/cm^sup 2^ TCW; 16C water temperature; 6 km sensor height ASL, typical for 15 m MASTER data; and a nadir viewing geometry). Table 3 demonstrates that both the selected wavelength and TCW greatly influence the magnitude of the atmospheric compensation. For the base case sensor height ASL (6 km), selecting a band in the middle of the atmospheric window reduces atmospheric correction to less than 1C even for large TCW. However,this table also reveals the sensitivity of the magnitude and sign of the correction to the water temperature and the sensor height ASL. Differences related to changes in surface height ASL and viewing angle are not important in this study (both less than 0.2C effect).

TABLE 3. Atmospheric Compensation Sensitivity Study Using MODTRAN.

Evaluation of Compensations Through Data case Studies

Through the above discussion, it is clear that variations in atmospheric absorption and emission can yield a range of image- derived water temperatures. In this section, case studies are used to show the effect of sensor height ASL on temperature retrievals and the effect of atmospheric compensation on temperature spectra.

Effect of Path Lengths - Lake Youngs Example. MASTER images of Lake Youngs with 5 m and 15 m spatial resolution at nadir were acquired at 11:48 PDT and 14:40 PDT, respectively, on August 25, 2001 (Figure 4, Table 2). As the sensor height ASL increases from 2 km (5 m MASTER data) to 6 km (15 m MASTER data) and both images include viewing geometries that deviate from nadir, this pair of Lake Youngs images demonstrates the effects of changing the path length from the water surface to the detector.

Uncorrected MASTER temperatures indicate that Lake Youngs cooled between the acquisition of the 5 m (11:48) and 15 m (14:40) data, whereas corrected images show Lake Youngs warmed by 0.6C. The corrected Lake Youngs image temperatures agree with a temperature gage in Little Soos Creek, an outlet of the lake. The trend in the uncorrected image temperatures is consistent with higher transmissivity and path radiance at 2 km ASL (5 m MASTER data) than at 6 km ASL (15 m MASTER data). When the sensor is 2 km ASL, radiation from the water surface has traveled through 80 percent of the TCW, and the atmospheric temperature is relatively warm (e.g., 12C in Figure 2). When the sensor is 6 km ASL, the water leaving radiance has traveled through 98 percent of the TCW and the atmospheric temperature is much cooler (e.g., -11C in Figure 2). Thus, in the 5 m MASTER data, decreased water vapor absorption yielded increased atmospheric transmissivity, and higher atmospheric temperature yielded a larger path radiance.

Figure 4. August 25, 2001, Lake Youngs Temperature Analysis: (a) Uncorrected T^sub rs^ 5 m MASTER, (b) Corrected T^sub rs^ 5 m MASTER, (c) Uncorrected T^sub rs^ 15 m MASTER, (d) Corrected T^sub rs^ 15 m MASTER, and (e) T^sub k-gage^ From Little Soos Creek (outlet of Lake Youngs).

Changing the viewing geometry from nadir (straight down) to off- nadir elongates the path through which water leaving radiation travels to the sensor. Increased absorption and emission at MASTER'S extreme viewing geometries (up to 40 degrees off-nadir) do not, however, affect the radiant temperature of Lake Youngs (Table 3, Figure 4). Although differences in path length certainly change transmission and path radiance values, there is a rough balance between increased emission and absorption at longer path lengths. MODTRAN simulations suggest this balance exists for water temperatures ranging from 10 to 20C and sensor heights ASL ranging from 2 to 6 km.

Spectral Evaluation of Temperature Retrievals. Multispectral sensors allow an independent check on temperature retrievals. After removal of atmospheric and emissivity effects, recovered temperatures should be similar for all TIR sensor bands (flat "temperature spectra"). Temperature differences between bands can also result from instrument calibration problems, instrument noise, and in-scene spectral variability from nonwater materials. Temperature spectra for a range of corrections, sensors, and water targets are shown in Figure 5.

For the 5 m MASTER data, atmospheric compensation had the least influence on estimated temperature. For these data, large path radiance values lead to a near balance between atmospheric absorption and emission. For Lake Meridian (T is approximately 21.5C, Figure 5a), corrected image temperatures agreed with surface kinetic temperatures. In addition, compensations reduced the standard deviation in temperature across bands from 0.8 to 0.4C. For the Green River (T is approximately 15C, Figure 5b), corrected image temperatures are a degree warmer than surface kinetic temperatures. For this colder river, the shape of the temperature spectra changed minimally for all compensations and spectral flatness was not improved by atmospheric compensation.

For 15 m MASTER data, atmospheric compensations based on MM5 TCW resulted in the flattest spectral shape (Figures 5c and 5d). High temperatures in Band 41 (λ^sub eff^ = 7.86 m) are related to the visible noise and the extreme sensitivity to atmospheric correction (Figure 3b). Lake Meridian MM5TCW corrected temperatures were within 0.2 C of the measured surface kinetic temperature (Figure 5c). On the other hand, the Green River MM5-TCW corrected temperatures were more than 2C higher than instream temperatures (Figure 5d). Use of the midlatitude summer standard atmosphere and the MM5 profiles resulted in higher calculated temperatures and introduced a concave upward shape, indicating overcompensation.

Figure 5e shows temperature spectra for near concurrent August 11, 2001, acquisitions of Lake Meridian temperatures by Landsat ETM+ and ASTER. The flattest spectra for the satellite images of Lake Meridian have only 0.75 cm of TCW, 0.65 cm less than the MM5 TCW estimate. The temperature spectra corrected by the MM5 TCW have elevated temperatures in bands 10 (λ^sub eff^ = 8.29 m) and 11 (λ^sub eff^ = 8.63 m). Surface kinetic temperature measurements agree with Landsat temperatures, but were 1.3C cooler than ASTER temperatures.

Figure 5. Temperature Spectra: (a) Lake Meridian 5 m MASTER, (b) Green River (GR4) 5 m MASTER, (c) Lake Meridian 15 m MASTER, (d) Green River (GR4) 15 m MASTER, and (e) Lake Meridian Landsat ETM+/ ASTER. (Notes for (e): Landsat ETM+ temperatures are in gray and ASTER temperatures are in black. Note different temperature scale but same range for y axis.

Overall Effects of Emissivity and Atmospheric Compensations

Atmospheric and emissivity compensations were further evaluated by examining how they affect mean temperature and standard deviation across bands for selected water targets. Mean temperatures were calculated as the mean of the recommended bands (Table 1). Standard deviation across bands was calculated to evaluate spectral flatness, and thus all bands that were not severely affected by instrument noise were included in the calculations (Bands 42 to 49 for MASTER, Bands 10 to 14 for ASTER). Emissivity compensation increased all uncorrected radiant temperatures by 0.6 to 0.9C. Atmospheric compensation generally increased uncorrected radiant temperatures (max 2.0C), except for the Green River, where temperatures in 5 m MASTER data decreased by 0.2C. The net effect of emissivity and atmospheric compensation was an increase in temperature for all water bodies (Figure 6a). Emissivity compensation did not significantly change the standard deviation in temperature across bands. However, the selected atmospheric compensations reduced the standard deviation in temperatures across bands especially for 15 m MASTER lakes and ASTER lakes (Figure 6b).

Figure 6. Effect of Emissivity and Atmospheric Compensations: (a) Average Change in Trs Resulting From Emissivity and Atmospheric Compensation, and (b) Average Effect of Emissivity and Atmospheric Compensations on Standard Deviation in Temperature Across Sensor Bands. Bars indicate the maximum and minimum corrections applied and resulting standard deviations.

TABLE 4. Percentage of Temperatures Estimated From TIR Images Within 1C of In Situ Temperatures.1

Comparing Image and Concurrent Ground Temperature Measurements:

Summary of Comparisons. Without emissivity and atmospheric compensation, all remotely sensed radiant temperatures (T^sub rs^) were within 3.0C of concurrent in situ temperature measurements (T^sub k-surf^, T^sub k-gage^, and T^sub r-surf^). With no compensations, 89 percent of river gage temperatures (T^sub k- gage^) were within 1C of T^sub rs^, whereas less than 50 percent of surface river temperatures (T^sub k-surf^ and T^sub r-surf^), and in situ lake temperatures (T^sub k-surf^, T^sub k-gage^, and T^sub r- surf^), were within 1C of T^sub rs^ (Table 4).

Applying emissivity and the selected atmospheric compensations improved agreement between T^sub rs^ and concurrent surface temperatures (T^sub r-surf^ and T^sub k-surf^), especially for the MASTER temperatures (Table 4). After compensation, the percentage of T^sub r-surf^ and T^sub k-surf^ within 1C of MASTER estimated temperatures increased to 100 percent (10 measurements). None of the T^sub k-surf^ or T^sub r-surf^ measurements was within 1C of corrected ASTER T^sub rs^, although this result is based on a very small sample size (3). For lake temperatures, emissivity and atmospheric compensation decreased the average difference between Trg and measured T^sub k-surf^ from 2.0C to 1.1C. However, applying compensations did not universally improve agreement between all concurrent stream temperature measurements and T^sub rs^. For streams, compensation decreased the percentage of T^sub k-gage^ within 1C of T^sub rs^ from 89 percent to O percent, though agreement with the surface temperatures improved.

Example: Temporal Variations and Detector Stability. Comparison of ground measurements and Trs reveals the importance of temporal variability when measuring stream temperatures. In Figure 7, Landsat ETM+ T^sub rs^ and ground temperature measurements of the Columbia River on September 17, 2002, are compared (Table 2). Ground temperature measurements include two sets of T^sub r-surf^ measurements (Detector A, Detector B) and one set of T^sub k-surf^ measurements taken from a boat in the center of the Columbia River. For this site, the uncorrected \T^sub rs^ is lower than Detector B, T^sub r-surf^, and T^sub k-surf^. After emissivity and atmospheric compensation, the corrected T^sub rs^ is within 1C of T^sub k-surf^ and the average T^sub r-surf^ from Detector B. Detector A, however, has a persistently lower T^sub r-surf^ than the corrected Landsat ETM+ T^sub rs^. Detectors A and B track some of the same temporal variations in temperature, but Detector A consistently records colder temperatures. Although persistent differences between Detectors A and B are attributed to calibration and detector instability, observed temporal variations could be attributed to environmental factors such as evaporative cooling caused by the wind. This comparison illustrates the importance of having several independent methods for estimating temperature. Direct comparisons of instantaneous image temperatures with the nearest instream temperature measurement may be misleading.

Figure 8 compares 5 m and 15 m MASTER temperatures with T^sub k- gage^ measurements for the Yakima River (Table 2). As expected, both the image and gage temperatures increase through the middle of the afternoon. After emissivity and atmospheric compensation, the image temperatures exceed T^sub k-gage^. The temperature increase from compensation is larger for the 15 m MASTER data than for 5 m MASTER data (see Table 3, Effect of Sensor Height AGL). As a result, the rate of temperature increase for the corrected temperatures (+1.1C) is closer to the temperature trend recorded by the instream gage (+1.3C). This example shows how agreement between image and instream temperatures can be misleading. The uncorrected T^sub rs^ are closer to the T^sub k-gage^, yet they do not capture the observed rate of temperature increase.

Figure 7. Comparison of Landsat ETM+ Temperatures With In Situ Radiant Temperatures (T^sub r-surf^) From Two Detectors (Detector A and Detector B) and In Situ Surface Kinetic Temperatures (T^sub k- surf^) for the Columbia River (Hanford Reach) on September 17, 2002. The standard deviation in Landsat ETM+ temperatures is provided.

Figure 8. Comparison of Gage, 5 m MASTER, and 15 m MASTER Temperatures. The standard deviation in MASTER temperatures is provided.

DISCUSSION

Results from this study provide strategies for processing and interpreting TIR images of water temperature in the land surface environment. This section provides suggestions for analysis and interpretation of water TIR images.

Obtaining accurate surface radiant temperatures requires compensation for emissivity and atmospheric effects. While emissivity compensation increases estimated water temperatures (by approximately 0.8C in this study), the magnitude and sign of the atmospheric compensation depend on many factors including the water temperature, surface height ASL, sensor height ASL, TCW, and the wavelength used.

If nonempirical methods are used, a radiative transfer model is necessary to compute atmospheric transmission and emission. Radiative transfer model predictions are, however, limited by the accuracy of atmospheric water vapor and vertical temperature profiles. After testing a range of strategies, concurrent in situ radiosonde measurements and a high resolution numerical weather model were found to be the most reliable and accurate sources for obtaining these profiles. For all field sites in this study, using the standard mid-latitude summer atmosphere to predict atmospheric effects resulted in overcorrected concaveupward temperature spectra. Uncorrected temperatures were often closer to in situ temperatures and had smaller standard deviation across bands than temperatures corrected using a standard atmosphere.

Independent checks on corrected remotely sensed temperatures demonstrate the success of atmospheric and emissivity compensation. The opportunity to assess spectral flatness is a clear advantage for multispectral over single band TIR data. Although data in bands with large atmospheric effects were not used to calculate temperatures, they were useful for evaluating and constraining the ability of MODTRAN to remove atmospheric absorption and emission. In addition, the spectral flatness criterion helped reveal band specific calibration issues.

Comparison of TIR image temperatures with concurrent and persistent ground truth temperature data is essential, but difficult. Direct comparisons of ground and image temperatures are hampered by a range of factors including the accuracy of T^sub r- surf^ measurements, differences in spatial and temporal averaging, and thermal stratification within the water body.

Remotely sensed water temperatures are snapshot measurements of surface radiant temperature. They are affected by the emissivity of water, atmospheric absorption and emission, and surface effects. Because both image and in situ T^sub r-surf^ detect the water skin temperature, one might expect to use T^sub r-surf^ to validate image temperatures. The T^sub k-surf^ sensors should track broad surface temperature trends but may miss ephemeral surface effects such as wind induced evaporative cooling. Unfortunately, this study found that inexpensive radiant temperature measurements (T^sub r-surf^) do not present a good standard of comparison for image temperatures because of their poor accuracy (2C) and their sensitivity to environmental conditions. More accurate (within 0.5C) T^sub r-surf^ measurements are the best way to validate image temperatures and evaluate variability in surface radiant temperature.

Comparisons between image and field temperatures are further complicated by differences in spatial and temporal averaging. Field measurements record point values, whereas images record temperatures integrated over a much larger area. Most field temperatures were recorded continuously during the time of image acquisition, whereas remotely sensed water temperatures are instantaneous "snapshot" measurements of surface radiant temperature. In this study, bulk water temperature, as recorded by gage and kinetic temperature sensors, changed gradually so that the temperature at the time of image acquisition could be identified to within less than 1C. In contrast, time series of surface radiant temperatures (Figure 7) show significant temporal variability. Although much of this variability could result from random error, the similarities in trends recorded by multiple sensors suggest that surface temperatures are much more variable than bulk temperatures.

Finally, this study reveals that a key issue in comparison of field and image temperatures is thermal stratification. In general, applying corrections improved comparisons between image temperatures and surface temperature measurements but increased differences between instream gage temperature measurements and image temperatures (see Table 4). This discrepancy may result because instream gage temperature measurements were taken at depth. Indeed, a series of point measurements in the field area show that summer surface temperatures were on average 2.2C (1.4C standard deviation) warmer than the temperatures near the streambed where the gages were located. Hence, agreement between image temperatures and ground temperatures does not always imply that accurate temperatures have been recovered from the TIR images. Thermal stratification is not a problem in all stream environments (e.g., shallower streams investigated by Torgersen et al., 2001). Therefore, an assessment of thermal stratification should be made in any field area before TIR images are used to measure water temperatures.

CONCLUSION

Thermal infrared images can be used to estimate the radiant temperature of water bodies that have environmental and economic import. In this study, emissivity compensation increased radiant temperatures by approximately 0.8C, whereas atmospheric compensation had a more variable effect (-0.2 to 2C) depending on sensor height ASL, target temperature, and atmospheric conditions. Overall, the selected compensations improved agreement between image and in situ surface temperatures, but decreased agreement between image and instream gage temperatures. Field observations show that thermal stratification may explain this discrepancy. Agreement between image and in situ temperatures did not always imply that accurate temperatures have been recovered from the TIR images. Unmeasured surface effects, instrument calibration, measurement error, and differences in spatial and temporal scaling could contribute to additional differences between image and in situ surface temperatures. For additional discussion, see Cherkauer et al. (2005). In the end, practical temperature differences between image and in situ surface temperatures were around 1C. This is significantly less than the highest achievable accuracy for TIR sensors (e.g., ASTER NEAT of 0.2 to 0.3C) and the reported accuracy for SST retrievals (as low as 0.2C) (Donlon et al, 2002). Therefore, depending on the application, it is essential to be aware of the types of factors that can affect remotely sensed water temperatures and to understand how the resulting image temperatures relate to the bulk water temperatures needed for locating thermal refugia, water quality monitoring, or temperature modeling.

Kay, Jennifer E., Stephanie K. Kampf, Rebecca N. Handcock, Keith A. Cherkauer, Alan R. Gillespie, and Stephen J. Burges, 2005. Accuracy of lake and Stream Temperatures Estimated From Thermal Infrared Images. Journal of the American Water Resources Association (JAWRA) 41(5):1161-1175.

1 Paper No. 04102 of the Journal of the American Water Resources Association (JAWRA) (Copyright 2005). Discussions are open until April 1, 2006.

ACKNOWLEDGMENTS

This research was supported by the U.S. Environmental Protection Agency's Science to Achieve Results (STAR) program (Grant No. R827675-01-0). However, the research does not necessarily reflect the views of the Agency, and no official endorsement should be inferr\ed. The authors thank Robert Elleman for post-processing MM5 output, Bill Gustafson for technical assistance, Dr. Gail Anderson for assistance with MODTRAN, Dr. Paul Bedrosian for field assistance, and Dr. Steve Warren for helpful insights.

LITERATURE CITED

Banks, W., R. Paylor, and W. Hughes, 1996. Using Thermal Infrared Imagery to Delineate Groundwater Discharge. Groundwater 34:434-444.

Barsi J.A., J.R. Schott, F.D. Palluconi, D.L. Heider S.J. Hook, B.L. Markham, G. Chander, and E.M. O'Donnell, 2003. Landsat TM and ETM+ Thermal Band Calibration. Canadian Journal of Remote Sensing 29 (2): 141-153.

Berk, A., G.P. Anderson, PK. Acharya, J.H. Chetwynd, L.S. Bernstein, E.P. Shettle, M.W. Matthew, and S.M. Adler-Golden, 1999. MODTRAN4 User's Manual. Air Force Research Laboratory, Hanscom Air Force Base, Massachusetts.

Boyd, M., S. Kirk, M. Wiltsey, and B. Kasper, 2002. Upper Klamath Lake Drainage Total Maximum Daily Load (TMDL) and Water Quality Management Plan (WQMP). State of Oregon Department of Environmental Quality. Available at http://www. deq. state. or.us/WQ/TMDLs/ KlamathBasin/UKLDrainage/ UprKlamathTMDL.pdf. Accessed in July 2005.

Cherkauer, Keith A., Stephen J. Burges, Rebecca N. Handcock, Jennifer E. Kay, Stephanie K. Kampf, and Alan R. Gillespie, 2005. Assessing Satellite-Based and Aircraft-Based Thermal Infrared Remote Sensing for Monitoring Pacific Northwest River Temperatures. Journal of the American Water Resources Association (JAWRA) 41(5):1149- 1159.

Coll, C. and V. caselles, 1997. A Split-Window Algorithm for Land Surface Temperature From Advanced Very High Resolution Radiometer Data: Validation and Algorithm Comparison. Journal of Geophysical Researchl02(D14):16697-16713.

Donlon, C.J., PJ. Minnet, C. Gentemann, TJ. Nightingale, LJ. Barton, B. Ward, and M.J. Murray, 2002. Toward Improved Validation of Satellite Sea Surface Skin Temperature Measurements for Climate Research. Journal of Climate 15:353-369.

Ebersole J.L., WJ. Liss, and C.A. Frissell, 2003. Cold Water Patches in Warm Streams: Physicochemical Characteristics and the Influence of Shading. Journal of the American Water Resources Association (JAWRA) 39(2):355-368.

Gao, B.C. and A.F.H. Goetz, 1990. Column Atmospheric Water Vapor and Vegetation Liquid Water Retrievals From Airborne Imaging Spectrometer Data. Journal of Geophysical Research 95:3549-3564.

Gillespie, A.R., 1992. Spectral Mixture Analysis of Multispectral Thermal Infrared Images. Remote Sensing of Environment 42:137-145.

Gillespie, A., S. Rokugawa, T. Matsunaga, J.S. Cothern, S. Hook, and A.B. Kahle, 1998. A Temperature and Emissivity Separation Algorithm for Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) Images. IEEE Transactions on Geoscience and Remote Sensing 36(4):1113-1126.

Goward, S.N., J.G. Masek, D.L. Williams, J.R. Irons, and R.J. Thompson, 2001. The Landsat 7 Mission - Terrestrial Research and Applications for the 21st Century. Remote Sensing of Environment 78:3-12.

Grell, G.A., J. Dudhia, and D.R. Stauffer, 1994. A Description of the Fifth Generation Penn State NCAR Mesoscale Model (MM5). NCAR Tech Note NCAR/TN-398+STR, National Center for Atmospheric Research (NCAR), Boulder, Colorado, 121 pp.

Gu, D., A.R. Gillespie, A.B. Kahle, and RD. Palluconi, 2000. Autonomous Atmospheric Compensation (AAC) of High Resolution Hyperspectral Thermal Infrared Remote Sensing Imagery. IEEE Transactions on Geoscience and Remote Sensing 38(6):2557-2570.

Hook, S., J. Myers, K. Thome, M. Fitzgerald, and A. Kahle, 2001. The MODIS/ASTER Airborne Simulator (MASTER) - A New Instrument for Earth Science Studies. Remote Sensing of Environment 76:93-102.

Hook, S.J., F.J. Praia, R.E. Alley, A. Abtahi, R.C. Richards, S.G. Schladow, and S.O. Palmarsson, 2003. Retrieval of Lake Bulk and Skin Temperatures Using Along-Track Scanning Radiometer (ATSR-2) Data: A case Study Using Lake Tahoe, California. Journal of Atmospheric and Oceanic Technology 20 (4): 534-548.

Hk, TO., E.S. Rutherford, S.J. Brines, D.J. Schwab, and M.J. McCormick, 2004. Relationship Between Surface Water Temperature and Steelhead Distributions in Lake Michigan. North American Journal of Fisheries Management 24 (1):211-221.

Ishiyama, T., K. Tsuchiya, and S. Sugihara, 1995. Influence of Look-Angle on the Water-Surface Temperature Observed With an IR Radiometer. Advances in Space Research 17(l):43-46.

Kaufman, Y.J. and B.-C. Gao, 1992. Remote Sensing of Water Vapor in the Near IR From EOS/MODIS. IEEE Trans. Geosci. Remote Sensing 30:871-884.

Kidder, S.Q. and TH. Yonder Haar, 1995. Satellite Meteorology: An Introduction. Academic Press, London, United Kingdom.

Kumar, A., P.J. Minnett, G. Podesta, and R.H. Evans, 2003. Error Characteristics of the Atmospheric Correction Algorithms Used in Retrieval of Sea Surface Temperatures From Infrared Satellite Measurements: Global and Regional Aspects. Journal of the Atmospheric Sciences 60:575-585.

Marine K.R. and J.J. Cech, 2004. Effects of High Water Temperature on Growth, Smoltification, and Predator Avoidance in Juvenile Sacramento River Chinook Salmon. North American Journal of Fisheries Management 24(1):198-210.

NASA (National Aeronautics and Space Administration), 2004. Land Processes Distributed Active Archive Center: Earth Observing System Data Gateway. Available at http://edcim-swww.cr.usgs.gov/pub/ imswelcome/. Accessed in August 2005.

NASA-JPL (National Aeronautics and Space Administration-Jet Propulsion Laboratory), 2004a. MASTER Level-IB: Flight No. 01-005- 03. California Institute of Technology. Available at http:// masterweb.jpl.nasa.gov/data/MissionView.htm7Missionid= 100249. Accessed in August 2005.

NASA-JPL, (National Aeronautics and Space Administration-Jet Propulsion Laboratory), 2004b. MASTER Level-IB: Flight #01005-04. California Institute of Technology. Available at http:// masterweb.jpl.nasa.gov/data/MissionView.htm7Missionid=100250. Accessed in August 2005.

NASA-JPL (National Aeronautics and Space Administration-Jet Propulsion Laboratory), 2004c. MASTER Level-IB: Flight No. 01-005- 07. California Institute of Technology. Available at http:// masterweb.jpl.nasa.gov/data/MissionView.htm7Missionid=100253. Accessed in August 2005.

NASA-JPL (National Aeronautics and Space Administration-Jet Propulsion Laboratory), 2004d. MASTER Level-IB: Flight No. 01-005- 08. California Institute of Technology. Available at http:// masterweb.jpl.nasa.gov/data/MissionView.htm7Missionid=100254. Accessed in August 2005.

Poole, G.C. and C.H. Berman, 2001. An Ecological Perspective on In-stream Temperature: Natural Heat Dynamics and Mechanisms of Human- Caused Thermal Degradation. Environmental Management 27(6):787-802.

Salisbury, J.W. and D.M. D'Aria, 1992. Emissivity of Terrestrial Materials in the 8-14 mm Atmospheric Window. Remote Sensing of Environment 42:83-106.

Schott, J.R., J.A. Barsi, N.G. Raqueno, and D.D. Alwis, 2001. Calibration of Landsat Thermal Data and Application to Water Resource Studies. Remote Sensing of Environment 78:108-117.

Thome K., F. Palluconi, T. Takashima, and K. Masuda, 1998. Atmospheric Compensation of ASTER. IEEE Transactions on Geoscience and Remote Sensing 36(4):1199-1211.

Torgersen, C.E., R.N. Faux, B.A. Mclntosh, N.J. Poage, and D.J. Norton, 2001. Airborne Thermal Remote Sensing for Water Temperature Assessment in Rivers and Streams. Remote Sensing of Environment 76:386-398.

Torgersen, C.E., D.M. Price, H.W. Li, and B.A. Mclntosh. 1999. MuItiscale Thermal Refugia and Stream Habitat Associations of Chinook Salmon in Northeastern Oregon. Ecological Applications 9:301- 319.

USGS (U.S. Geological Survey), 2005. National Center for Earth Resources Observation and Science (EROS): Earth Explorer. Available at http://edcsnsl7.cr.usgs.gov/EarthExplorer/. Accessed in August 2005.

Washington State, 2003. Washington Administrative Code, Title 173, Chapter 201A: Water Quality Standard for Surface Waters of the State of Washington. Office of the Code Revisor, Olympia, Washington. Available at http://www.leg.wa.gov/WAC/index. cfm?fuseaction=chapterdigest&chapter=1730 -201A. Accessed in July 2005.

Yamaguchi, Y., A.B. Kahle, H. Tsu, T. Kawakami, and M. Pniel, 1998. Overview of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER). IEEE Transactions on Geoscience and Remote Sensing 36(4):1062-1071.

Young, S.J., B.R. Johnson, and J.A. Hackwell, 2002. An In-scene Method for Atmospheric Compensation of Thermal Hyperspectral Data. Journal of Geophysical Research 107(D24):4774.

Jennifer E. Kay, Stephanie K. Kampf, Rebecca N. Handcock, Keith A. Cherkauer, Alan R. Gillespie, and Stephen J. Burges2

2 Respectively, Ph.D. Candidate (Kay), Affiliate Assistant Professor (Handcock), and Professor (Gillespie), Department of Earth and Space Sciences, University of Washington, Box 351310, Seattle, Washington 98195; Ph.D. Student (Kampf) and Professor (Burges), Department of Civil and Environmental Engineering, University of Washington, Box 352700, Seattle, Washington 98195-2700; and Assistant Professor (Cherkauer), Department of Agricultural and Biological Engineering, Purdue University, 225 South University Street, West Lafayette, Indiana 47907-2093 (E-Mail/Kay: jenkay@u.washington.edu).

APPENDIX

Copyright American Water Resources Association Oct 2005

Source: Journal of the American Water Resources Association

More News in this Category