Accuracy assessment of land surface temperature ...

Environ Monit Assess DOI 10.1007/s10661-013-3565-9

Accuracy assessment of land surface temperature retrievals from Landsat 7 ETM + in the Dry Valleys of Antarctica using iButton temperature loggers and weather station data Lars Brabyn & Peyman Zawar-Reza & Glen Stichbury & Craig Cary & Bryan Storey & Daniel C. Laughlin & Marwan Katurji

Received: 23 July 2013 / Accepted: 19 November 2013 # Springer Science+Business Media Dordrecht 2013

Abstract The McMurdo Dry Valleys of Antarctica are the largest snow/ice-free regions on this vast continent, comprising 1 % of the land mass. Due to harsh environmental conditions, the valleys are bereft of any vegetation. Land surface temperature is a key determinate of microclimate and a driver for sensible and latent heat fluxes of the surface. The Dry Valleys have been the focus of ecological studies as they arguably provide the simplest trophic structure suitable for modelling. In this paper, we employ a validation method for land surface temperatures obtained from Landsat 7 ETM + imagery

L. Brabyn (*) : G. Stichbury Geography, Tourism and Environmental Planning, University of Waikato, Hamilton, New Zealand e-mail: [email protected] P. Zawar-Reza Centre for Atmospheric Research, University of Canterbury, Christchurch, New Zealand C. Cary : D. C. Laughlin Department of Biological Sciences, University of Waikato, Hamilton, New Zealand B. Storey Gateway Antarctica, University of Canterbury, Christchurch, New Zealand M. Katurji Atmospheric Modelling and Dynamics, Michigan State University, East Lansing, MI, USA

and compared with in situ land surface temperature data collected from four transects totalling 45 iButtons. A single meteorological station was used to obtain a better understanding of daily and seasonal cycles in land surface temperatures. Results show a good agreement between the iButton and the Landsat 7 ETM + product for clear sky cases. We conclude that Landsat 7 ETM + derived land surface temperatures can be used at broad spatial scales for ecological and meteorological research. Keywords Landsat . Land surface temperature . Dry Valleys . Antarctica

Introduction Surface temperature at a hierarchy of scales is an important parameter for biological, hydrological and climactic processes in the Antarctic continent. At a planetary level, Antarctica is a heat sink and plays a significant role in the global atmospheric general circulation patterns and forcing ocean current systems (Van den Broeke 2004). A recent review of climate change and the environment by the Scientific Committee on Antarctic Research (SCAR; http://www.scar.org/) came to an alarming conclusion (Turner et al. 2009) that the stratospheric ozone depletion over the continent has probably masked the potential warming effects of global climate change. Therefore with the eventual predicted recovery of the ozone hole in 50 years,

Environ Monit Assess

surface temperatures should start showing an increasing trend. Associated with warming will be changes to the sea ice extent, and its effect on algae production, global ocean currents and marine food chains (Turner et al. 2009). Climate change is predicted to impact first and most severely the higher latitudes (Callaghan et al. 1992; Kennedy 1995; Vincent 1997; Walker 1997), giving importance for establishing a cost-effective, high-resolution (temporally and spatially) climate monitoring network across Antarctica. Antarctica is the driest, windiest and coldest continent on earth due to its geography, topography and a mean elevation of approximately 2,000 m. Most of the continent is ice and snow covered, with only about 0.3 % bare ground. The McMurdo Dry Valleys of Victoria Land (hereafter referred to as the Dry Valleys) forms the largest contiguous bare land (Hemmings 2001). As the namesake implies, the Dry Valleys are regions where water is scare. This is because precipitation is low and is frozen and unavailable most of the year to biology, except when eventually surface temperatures climb above the freezing point during the summer months from November to February (Peck et al. 2006). For this reason temperature is considered to be an important co-determinant of habitat, along with the availability of snow or glaciers in the vicinity to provide melt water. To model the distribution of biology in Antarctica requires a comprehensive understanding of the temporal and spatial variations in temperature. Satellite imagery of the thermal infra-red band is becoming increasingly available and can be used to calculate land surface temperature (Suga et al. 2003). Both MODIS and Landsat thermal infra-red images have been collected globally by NOAA and NASA, respectively. These images have been archived and provide a comprehensive depository of data that can be converted to temperature. MODIS captures infra-red at 1-km spatial resolution, while Landsat 7 ETM + is 60 m. MODIS has a daily return rate, while Landsat has a 16day return rate. As both satellites are polar orbiting, images overlap considerably over Antarctica, therefore the return rate is effectively more frequent. There has been a considerable amount of data compilation of temperature surface data of Antarctica using MODIS and the Advanced Very High Resolution Radiometer (AVHRR), and these have been used to show spatial and temporal trends (Shuman and Comiso, 2002; Schneider et al. 2004; Turner et al. 2005).

Remotely sensed data still need to be validated against in situ measurements, yet high-resolution spatial data is rarely available, especially in a remote region like the Dry Valleys. Inexpensive iButton temperature sens o r s a l l e v i a t e t h i s s i t u a t i o n ( h t t p : / / w w w. maximintegrated.com/products/ibutton/). The iButtons are small (button-sized) data loggers that were originally designed for monitoring ambient temperature during transport and storage of perishable items, such as food. Increasingly they are being used for regional environmental studies such as modelling snow melt (Lundquist and Lott, 2008). They can record and store data every 4 h for up to a year, making them ideal for long-term measurements. They have an operating temperature range of −40 to +85 (°C), and being inexpensive (each costs less than US$50), means that a significant number can be deployed across a landscape to provide a spatially comprehensive in situ dataset of land surface temperature. Landsat-derived temperatures have been validated in many different environments but not in Antarctica, especially against a spatially high-resolution database that can be provided with iButton loggers. Suga et al. (2003) conducted a validation study of Landsat-7ETM + temperatures in the Hiroshima city and bay area in the western part of Japan. They used portable thermometers calibrated with reference to a standardized thermistor to ground truth the Landsat7 ETM + temperatures. Five sites covering a range of landuses were measured at four different times (total of 20 measurements). Correlation coefficients were from 0.9821 to 0.9994, and the difference of Landsat 7 ETM + estimation and truth observation were from 0 to 1.5 °C. Hale et al. (2011) examined a range of statistical techniques for validating satellitederived temperatures with in situ data. They examined MODIS and Aster images and concluded that the simple regression performed just as well as complex statistics. There appears to be more scientific validation of MODIS temperature readings than other satellites (Wan et al. 2002; Wan et al. 2004; Coll et al. 2005). There also is a substantial research effort to calibrate satellite thermal data for water temperature monitoring (Lathrop and Lillesand, 1987; Schott et al., 2012). This paper researches the accuracy of Landsat 7 ETM + for measuring land surface temperature when circulating 705 km above the ground by comparing this with iButton temperature measured on the ground at approximately the same time. Apart from variation in aspect,


slope and geology, the Dry Valleys in Antarctica is a relatively homogeneous and simple, single-tiered environment compared to warmer climates that have multiple tiers of mixed landcovers ranging from forests, pastures and urban areas. If the individual pixel readings of Landsat 7 ETM + are accurate, then this effectively provides a land surface temperature reading spaced 60 by 60 m across the landscape. Field work and downloading of data temperature loggers will be unnecessary and a land surface temperature can be generated for a region relatively cheaply and used in GIS modelling of habitats, hydrology and climate systems.

Physical setting As mentioned briefly above, the Dry Valleys is the continent’s largest continuous expanse of ice-free ground—approximately 6,000 km2 (Hemmings 2001) and are an important research site because of the uniqueness and relative simplicity of the terrestrial life. The ecology of the Dry Valleys is extremely unique and is an important site for the Long-Term Ecological Research (LTER) programme. The Dry Valleys are glacially excavated, and are thought to be ice-free due to uplift of the Transantarctic Mountains cutting off the flow of glaciers from the East Antarctic Ice Sheet. The Dry Valleys remain ice-free through ablation of snow and ice which exceeds any accumulation (Fitzsimons et al. 2001; Fountain et al. 1999; Friedmann 1982; Hopkins et al. 2006; Horowitz et al. 1972). The landcover of the Dry Valleys consists of alpine, piedmont and terminal glaciers, lakes mostly under permanent ice-cover, bare soils and stream channels (Gooseff et al. 2003). During the summer (November to February) there is continual daylight and during the winter months there is continual darkness with valley floor temperatures falling as low as −40 °C (Fountain et al. 1999). The mean annual air temperature is −20 to −25 °C (Doran et al. 2002); however, during the summer months the ground temperature may reach higher than 15 °C for short periods during the day (Horowitz et al. 1972) but overall the summer temperatures rarely rise more than 0 °C. The cold temperatures vastly decrease water vapour content in the atmosphere (Horowitz et al. 1972), and precipitation is limited to around 10 cm per year in the form of snow (Fountain et al. 1999), which often sublimates rather than melts (Horowitz et al.

1972). The low availability liquid water therefore qualifies the Dry Valleys as a cold desert (Peck et al. 2006). This aridity is enhanced in the Dry Valleys by the dry katabatic (down slope) winds off the Antarctic plateau (Friedmann 1982). Surface temperature plays a significant role in generating local to continental scale winds (Bromwich et al. 2007). Information on the temporal and spatial variations in temperature can be used to model and forecast katabatic winds. Remote sensing has had an important role in understanding Antarctica. Prior to the availability of thermal satellite images, temperature recording was limited to Automatic Weather Stations (AWS), requiring heavy logistics support and annual maintenance. Logistical and financial constraints has meant that single point temperature data are recorded from a limited number of locations in the Dry Valleys using AWS. There is a great need for having a comprehensive spatial and temporal dataset for surface temperature, and remote sensing using satellites thermal images appear to offer a costeffective solution. This study was conducted in the Miers Valley, which is one of four small valleys that comprise a region called the Denton Hills (see Fig. 1). The Denton Hills are located at the southern end of the Dry Valley system, which is located in the Ross Sea region of Antarctica. Miers Valley is oriented east–west; two glaciers—Adams and the Miers—dominate the westernmost section while the eastern side is open to the Ross Ice Shelf. The valleys in the Denton Hills (Hidden, Miers, Marshall and Garwood) are smaller than the large Dry Valleys further north yet the elevation ranges from sea level to approximately 1,300 m.

Methodology Collection of iButton and Automatic Weather Station temperature data In January 2009, 45 iButtons (model DS1921G) were placed 2 cm below the surface on four transects across the middle of the Miers Valley (as shown in Fig. 2). The transects were north–south oriented covering a range of elevations on both north and south aspects of the valley sidewalls. The horizontal distance between the most northern and southern locations is 4.4 km, while the west to east spread is approximately 3 km. The iButtons


Fig. 1 Location of Denton Hills Study Area


were located at 100-m elevation intervals, so as the ground got steeper the horizontal distance between them decreased. The 2 cm depth ensures that there is no direct heating by solar insolation. The locations were recorded using GPS (accuracy ±6 m) and marked by an orange stick for easier retrieval. The iButtons were recovered in the next Antarctic field season in December 2009, providing almost an entire annual dataset. Each iButton record has a unique identifier and the loggers are initialized and downloaded using two devices (DS1402-RP3+ and DS9490R). The logging interval was set at 4 h to ensure available memory was not overloaded. Downloaded files were imported in ArcGIS software using the ‘ADD XY DATA’ function and stored as a point shapefile. For long-term ecological monitoring purposes, an Automatic Weather Station operates continuously in the Miers Valley. The AWS has been operational since 2008 and is serviced every field season in summer. The location of the AWS was 220 m from the closest iButton (see Fig. 2) and provides a higher temporal resolution dataset. The AWS samples a variety of meteorologically important parameters such as wind velocity and relative humidity, but the one relevant for this research is the soil temperature measured at the depth of 2 cm, which is recorded every 15 min. Soil temperature is measured with a HOBO S-TMB-M002 sensor with accuracies of ±0.2 °C suitable for operation in −40 to 100 °C environment. The AWS temperature was used to understand how the temperature changes over time (temperature gradient) so that the time lag between the Landsat and Fig. 2 iButton and automatic weather station locations in the Miers Valley

iButton recordings can be compensated, as described in Section 3.3. Landsat 7 ETM + temperature calculations Ground temperature data can be calculated from Landsat 7 ETM + using band 6 (60 m2 resolution), which captures the thermal infra red spectrum (10.4– 12.5 μm). Cloud cover interferes with the ground reflectance, therefore only images from relatively clear days can be used. Images during the dark winter months were not archived; therefore, only images during the summer months were used. Each image was downloaded and manually checked for cloud cover for the study period. As the Landsat 7 satellite is polar orbiting, images are captured more frequently than the usual 16 days for lower latitudes closer to the equator. Table 1 shows the date and time of the 19 images that were cloud free and captured during the study period. All images were captured between 19:30 and 20:30 GMT, which is 8.30 am to 9.30 am in the morning using local time (New Zealand summer time—GMT + 13 h). The Landsat 7 ETM + thermal infra red sensor detects the absolute radiance and stores these values between 0 and 255 for compression reasons—known as digital numbers (DN). The thermal infra red sensor only detects radiance values above the atmosphere (top of the atmosphere—TOP). To produce a temperature surface from a Landsat image we used the “minimum/maximum spectral radiance scaling factor technique” as specified by the Landsat Users Handbook (http://

Environ Monit Assess Table 1 Capture date and time of Landsat 7 ETM + images used in the study Image date

Image time (GMT)

Image date

Image time (GMT)

22/01/2009

20:04

7/03/2009

20:29

24/01/2009

19:52

11/03/2009

20:04

25/01/2009

20:34

4/11/2009

20:17

29/01/2009

20:10

6/11/2009

20:04

3/02/2009

20:28

8/11/2009

19:52

14/02/2009

20:10

15/11/2009

19:58

16/02/2009

19:58

18/11/2009

20:29

19/02/2009

20:29

24/11/2009

19:53

23/02/2009

20:04

29/11/2009

20:11

25/02/2009

19:52

landsathandbook.gsfc.nasa.gov/), which required two steps: (1) conversion of DN to top of the atmosphere radiance (LTOA), and (2) conversion of atmospherically corrected radiance to temperature. To convert the Landsat 7 ETM + DN back to the top of the atmosphere radiance (LTOA) in watts per steradian per square meter (W sr−1 m−2), the following equation was used: LTOA ¼

Lmax ‐Lmmin ðDN‐ QCALmin Þ þ Lmin QCALmax ‐ QCALmin ð1Þ

Where LTOA=spectral radiance at the sensor aperture, QCAL=255, QCALmin =1, Lmax and Lmin =TOA radiances scaled to QCALmax and QCALmin DN = digital numbers values for band 6. LTOA can then be atmospherically corrected using the equation below. Lλ ¼

LTOA −Lu 1−ε0 − Ld τεo ε0

ð2Þ

Where Lλ = the atmospherically corrected radiance, LTOA = spectral radiance at the sensor aperture (TOA), Lu is atmospheric path radiance or upwelling, Ld is sky radiance or downwelling, τ is the atmospheric transmissivity and εo is surface emissivity. Apart from surface emissivity, these variables are specific to each image and were obtained from the following NASA website: http:// atmcorr.gsfc.nasa.gov/atm_corr.html.

Surface emissivity is the relative ability of a surface to emit energy by radiation. It is the ratio of energy radiated by a particular surface to energy radiated by a black body at the same temperature. Typically emissivity for granite may be 0.45 and sandy gravel 0.28 at 3 8 ° C ( h t t p : / / w w w. m o n a r c h s e r v e r. c o m / TableofEmissivity.pdf), however this will vary considerably with temperature. The iButtons were located on a mainly gravel granite composite and a range of emissivity values were experimented with. Even high emissivity values excessively over estimated temperatures compared to the iButton data. In the absence of emissivity information for the Antarctic environment, a value of 1 was uniformly applied. This means that the model is simplified to atmospheric transmissivity and atmospheric path radiance to calculate the atmospheric correction, which in any case is very limited in its effect in the study area due to the atmosphere conditions over the study area. Radiance to temperature conversions were made using the Planck equation or the Landsat 7 ETM + specific estimate of the Planck curve: T¼

K2 K1 þ1 In Lλ

ð3Þ

Where T is the temperature in Kelvin: Lλ m spectral radiance in W/m2×sr·μm; and K1 and K2 are calibration constants 666.09 and 1,282.71, respectively. Kelvin was then converted to Celsius by adding 273.15°. All data processing was performed using the ArcInfo workstation platform.

Spatial integration of Landsat and iButton data The Landsat 7 ETM + produces geographically referenced surfaces (raster layers) of temperature. These Landsat temperature surfaces can be spatially joined to the GIS point shapefile of the iButtons, since both datasets are geographically referenced. The ArcGIS tool called “Add Surface Information” provides this function. The result of this spatial join is a table of both iButton and Landsat temperatures, which can then be used for further analysis. The spatial resolution of the Landsat thermal infra-red band is 60 m and the GPS accuracy is 6 m, therefore the Landsat pixel will be a generalisation of the surrounding area of the iButton location.


Landsat 7 ETM + has a sensor fault that causes random lines of pixel to drop out occasionally. Sometimes the dropouts corresponded with the locations of some iButtons, consequently there were only 725 comparison records available rather than the 855 expected (19 Landsat records × 45 iButtons). Landsat 5 TM could have been used to replace these pixel drop outs but having 725 comparison records was considered sufficient for this study. The iButton and Landsat ETM + temperatures were then compared using a simple linear correlation. The iButton loggers were set to store data every 4 h, therefore at some locations there can be as much as 2 h difference between the satellite overpass and the in situ measurements. To evaluate the significance of this mismatch in timing, a correction was applied to the iButton data. As is pointed out in the coming sections, in the morning when the overpass occurs, temperatures can be rising rapidly. The gradient of temperature increase is calculated for each morning from the weather station data—described below—and is used to correct for the time difference. ΔT AWS T correction ¼ t overpass − t iButton ð4Þ ΔT AWS Where the first term on the right hand side of the equation calculates the rate of change in ground temperature at the time of the Landsat overpass (ΔT is change in temperature and Δt is change in time), and the second term calculates the time difference between the overpass and the iButton logged temperature. The Tcorrection is applied to the iButton temperatures for further analysis as explained below.

Validation and discussion Effects of the time lag The AWS 15 min ground temperature recordings show that the gradient ranges between 0.6 to 3 °C/h (0.01 to 0.05 °C/min). Figure 3 illustrates the time-series of surface temperature measured by the AWS at 15-min intervals and the LANDSAT and iButton measurements at the time of satellite overpass. The local time of the overpass corresponds to the morning period when the sun is rising from its minimum elevation (at this latitude the sun does not set for the period considered). The

southern valley sidewall casts a shadow between midnight to 4 a.m. (+12GMT), but at the time of overpass both the iButton network and the AWS have been exposed to the sun for at least 2 h. Given that the maximum time difference between an overpass and AWS logged temperature record can be 8 min, the mismatch in time can only account for 0.08 to 0.4 °C of the difference between the AWS and the Landsat 7 derived temperatures. In general, the summer months (November to January) have higher surface temperatures as compared to the autumn months of February and March. Since the soil is extremely dry, there is as much as 20 °C difference between the daily minimum and maximum temperatures. Even in the summer days, when sun reaches low elevation and shadow is cast by the sidewalls on the AWS, the surface temperature can drop below 0 °C. Correlation between iButton and Landsat temperatures A significant positive linear correlation between iButton temperature 2 cm below the soil surface and Landsat 7 temperature as illustrated by Fig. 4a and Table 2. The majority of observations that fell outside the 95 % prediction intervals were obtained in November from mid to high elevation sample locations. This can be explained by the fact that the subsurface soil temperatures from those higher elevations early in the summer are colder than the air temperature above the soil surface. When the iButton temperatures were corrected for the time lag, using the temperature gradient calculated from the AWS data, there was a slight decrease in the r value (see Fig. 4b). This result was not expected and indicates that the AWS-derived temperature gradient over compensated for the time lag between the Landsat and iButton recordings. The discrepancy in the correlation analysis could be partly explained by incorporating physiographic variables into a linear model (Table 2). The statistical model to explain variation in Landsat 7 temperature with only iButton temperature was clearly outperformed by the model that incorporated additional physiographic covariates (elevation, slope and aspect). The Akaike Information Criterion (AIC) which is a metric for the goodness of fit of the statistical model shows a difference of 115 between the two models, indicating a large improvement in model fit with addition of new predictors. However, the additional covariates only explained an additional 2 % in the total variation in Landsat 7

Environ Monit Assess Fig. 3 Time-series of AWS and Landsat 7 ETM + temperatures

.

Corresponding time to Landsat recording

//

temperature (Table 2). This suggests that site-to-site variation explains only a minor component of the discrepancy between Landsat 7 and iButton temperature measurements. Albedo was not a significant term in the model, likely due to the low variance in albedo that occurred in the sample.

Conclusion The in situ iButton and AWS ground temperature records collected in the Dry Valleys of Antarctica strongly correlate with temperature remotely sensed from

Fig. 4 a, b Scatter plot of Landsat 7 versus iButton temperatures

705 km in space using the thermal band of Landsat 7 ETM+. We therefore conclude that calculating ground temperature of a simple, single-tiered environment using Landsat 7 ETM + is a reliable method. Our research is consistent with Suga et al. (2003) study and therefore it is likely that temperatures derived from Landsat 7 ETM + will be reliable for most desert environments throughout the world that are relatively bare of vegetation. Schott et al. (2012) conducted a thermal infrared radiometric calibration of the entire Landsat 4, 5 and7 archive (1982–2010) and have demonstrated that these images have an error of ±0.6 K for a temperature range of 275–305 K. As their research focused on calibrating

Environ Monit Assess Table 2 Model selection results comparing Landsat 7 temperature as a function of iButtons alone (model 1) and combined with physiographic variables (model 2) Model 1

Coefficient P value

Landsat 7 temperature Intercept −0.71

0.85 0.0002

iButton temperature

1.03