AN ACCURACY ASSESSMENT OF AATSR LST DATA USING EMPIRICAL AND THEORETICAL METHODS Elizabeth Noyes, Gary Corlett, John Remedios, Xin Kong, and David Llewellyn-Jones Department of Physics and Astronomy, University of Leicester, Leicester, LE2 1AJ, UK ABSTRACT The Advanced Along-Track Scanning Radiometer (AATSR), onboard ESA s Envisat Satellite, is a precision radiometer that provides data from which global observations of land surface temperature (LST) can be derived. These LST data are retrieved using a nadir-only split window algorithm, with a target accuracy of 2.5 K during the day and 1.0 K at night. The algorithm will also be applied operationally to data recorded by the AATSR s predecessors, ATSR-1 (ERS-1) and ATSR-2 (ERS-2), leading to a > 15-year record of global LST data. We present the results of comparisons between operational AATSR LST retrievals and collocated in situ data that have been recorded continuously at three sites over at least one year. The exercise provides an excellent opportunity to assess the long-term accuracy of the satellite LST data. The results demonstrate that, although the overall agreement between the AATSR and in situ data is close to the target accuracy of the retrieval scheme, there are some notable seasonal biases in the data. These seasonal biases are also apparent in AATSR LST retrievals that have been simulated using a radiative transfer model with atmospheric and surface parameters that are appropriate for each of the sites, over one year. Examination of the variation of simulated and observed bias with LST, water vapour and atmospheric temperature indicates that the primary source of bias is most likely to be the current operational algorithm s high sensitivity to these parameters. Key words: Envisat, AATSR, LST, Validation, Radiative Transfer Model. 1. INTRODUCTION Land surface temperature (LST) is a key variable that drives the exchange of long-wave radiation between the Earth s surface and the atmosphere, and is therefore a parameter of great interest for climate-related studies. Several researchers have shown that it is possible to derive estimates of LST from satellite observations that are made in the infrared region of the electromagnetic spectrum [1, 2, e.g]. This has lead to the emergence of a number of operational satellite LST products, such as those from the Moderate Resolution Imaging Spectroradiometer (MODIS), the Advanced Along-Track Scanning Radiometer (AATSR) and the Spinning Enhanced Visible and Infra-red Imager (SEVIRI). In this paper, we report on the results of comparisons between operational AATSR LST retrievals and collocated in situ data that have been recorded continuously at three sites over at least one year. Investigations into sources of bias in these LST data, using both empirical and theoretical data sets, are also presented. 1.1. LST from the AATSR The current AATSR LST algorithm is a nadir-only split window algorithm that is described in full by [3]. In brief, LST is derived according to Eq. 1: LST = a f,i,pw +b f,i (T 11 T 12 ) n +(b f,i +c f,i )T 12 (1) where a f,i,pw, b f,i and c f,i are coefficients that depend on land type class (or biome) i, fractional vegetation cover f and precipitable water pw, and T 11 and T 12 are the AATSR brightness temperatures (BT) in the 11 and 12 µm channels, respectively. The purpose of the coefficients is to correct for the effects of the atmosphere and surface emissivity, in order to retrieve an effective LST over the field of view (FOV) of the AATSR observation. At the time of writing, the AATSR product uses 14 different biomes, including one lake class. The auxiliary data i, f and pw depend on geographical location and are tabulated at 0.5 degrees resolution. f and pw are also dependent on season. The n term is a small tuning parameter that is a function of zenith angle. Proc. Envisat Symposium 2007, Montreux, Switzerland 23 27 April 2007 (ESA SP-636, July 2007)
2. COMPARISON BETWEEN AATSR AND IN SITU LSTS 2.1. Test Sites Comparisons between AATSR and collocated in situ LSTs have been performed at three, well-established test sites located in Niamey (Niger), Oklahoma (USA) and Cardington (UK). In each case, the in situ LSTs have been estimated from ground-based observations made by radiometers operating in the 8-14 µm waveband on a nearcontinuous basis for more than one year. The geographical locations of these sites are given in Tab. 1; further details of the instrumentation at each site are given in Sections 2.1.1 and 2.1.2. 2.1.1. Niamey, Niger (Biome 7) and Oklahoma, USA (Biome 12) Field Sites Both the Niamey and Oklahoma field sites are instrumented as part of the Atmospheric Radiation Measurement (ARM) programme. Niamey is the current ARM mobile facility, and has been in operation since April 2005. The Oklahoma site is the most extensive ARM site and was established in July 1993. Both sites are equipped with a range of instrumentation that includes i) both upward and downward-looking 9-12 µm Heitronics KT19.85 Infrared Radiation Pyrometers for observing the radiometric temperature of the sky and surface, respectively ii) upward looking Radiometrics WVR-1100 microwave radiometers from which column water vapour and liquid are derived and iii) Vaisala ceilometers (Model CT25K) for detecting cloud base heights. The sites also house a range of standard meteorological equipment, such as humidity sensors, thermometers, etc. The reported accuracy of the pyrometers is 0.5 K; further specifics can be found in the ARM Infrared Thermometer Handbook [4]. The downward-looking radiometers used in this study are mounted at a height of 10 m, where the data are averaged over one minute. For the Niamey site, the in situ radiometric data available for this study extend from 24 November 2005 until 19 November 2006. For the Oklahoma site, the analysis period extends for almost three years, from 18 December 2003 until 19 November 2006. The data sets recorded by the ARM instruments are near continuous. 2.1.2. Cardington, UK (Biome 12) The Cardington field site is an operational Met Office site located in open farmland within the UK. The site is equipped with a wide range of instrumentation that includes i) a downward-looking Heitronics KT15.82D radiometer that operates in the 8-14 µm bandwidth mounted at 2 m height for surface temperature observations, ii) upward and downward Kipp & Zonen CG4 pyrgeometer for up and downwelling radiation measurements in the 4.5-40 µm bandwidth and iii) an upwardlooking WVR-110 microwave radiometer from which column water vapour and liquid are derived. The site also houses a full range of standard meteorological equipment, such as humidity sensors, thermometers, etc, and a ceilometer for cloud observations. Unfortunately, the data from the ceilometer were not available for this study. The estimated accuracy of the thermal infrared radiometer is 0.5 K. The data used in this study, which are also near-continuous, are averaged into 10 minute bins and extend from November 2004 until July 2006. 2.2. Comparison Methodology All the comparison results reported here have been produced using operational 1 km AATSR LST data. Near real time (NRT) orbit data have been used for comparisons after 10 March 2004. Prior to this date, which is when the AATSR LST product became operational, reprocessed level O orbit data have been used for the analysis. The general methodology adopted for comparing the AATSR with corresponding in situ LST data is as follows: 1. The in situ radiometric temperatures are corrected for surface emissivity effects and an estimate of point kinetic skin temperature is derived. Approximate in situ emissivities for the infrared radiometers have been derived from the Aster Spectral Library [5]. 2. For each AATSR overpass, the nominal 1 km pixel containing the location of the in situ observations is extracted. 3. The AATSR LST for this pixel and the in situ observation that is temporally closest to the exact time of the AATSR overpass are recorded as a matchup. In practise, the maximum temporal offset between the time of the AATSR overpass and the in situ observation is less than 5 minutes. 4. Any matchups flagged cloudy by the operational AATSR cloud screening algorithms are rejected. Additional cloud screening is performed on the remaining matchups and any remaining cloudcontaminated overpasses are rejected. In situ ceilometer data have been used for this purpose at both the ARM sites. In situ microwave liquid water and downwelling long-wave radiation observations have been used at the Cardington test site. 5. The mean and standard deviation of the quantity AATSR minus in situ LST difference is calculated for all cloud-free matchups falling within three standard deviations (3σ) of the overall (cloud-free) mean. These statistics are produced separately for day and night. The purpose of this 3σ screening
Table 1. Results of comparisons between AATSR and collocated in situ LST data. Column Biome refers to the AATSR biome classification of the site. Column Dates refers to the range of dates over which the comparisons have been performed. Biome Site Latitude Longitude T range N Data Bias ± StDev (K) Dates (K) Day Night Day Night 7 Niger 13.477 002.176 302-325 9 3-0.8 ± 2.4-0.3 ± 1.0 11/05-11/06 12 Oklahoma 36.600-097.500 265-321 47 81-1.4 ± 2.6-0.1 ± 1.8 12/03-11/06 12 Cardington 52.100-000.425 271-316 16 8-0.8 ± 2.6 1.3 ± 1.1 11/04-07/06 is to remove any extreme outliers that may bias the statistics. These outliers may result from undetected cloud or instrument malfunction, or some other anomaly that does not represent typical conditions. 2.3. Results The overall results of the comparisons are presented in Tab. 1. The results for night time overpasses are more consistent than those corresponding to the daytime at each test site, where the standard deviation of the biases ranges between 1.0 and 1.8 K, compared with 2.4-2.6 K during the day. The mean bias is negative in all cases with the exception of the night time Cardington results, which is +1.3 K. However, it should be noted that the number of matchups quite low (less than ten) for these results, and for both the night and day ARM-Niamey results, so therefore may not be truly indicative of the performance of the satellite LST retrievals for these cases. In general, the results are very encouraging and suggest that the operational AATSR LST retrievals are performing quite well at these sites. The great advantage of using long-term continuous data sets for assessing the accuracy of geophysical satellite data is that any seasonal variation in bias can be established. While still useful, dedicated validation campaigns usually provide only a snapshot of their performance, which may be misleading. Fig. 1 shows the variation AATSR minus in situ LST difference for cloud-free matchups with time over the ARM-Olahoma test site. A clear seasonal cycle is observed in these data, particularly for night time matchups, where the AATSR LSTs are found to be warmer in the summer and colder in the winter. The existence of this apparent seasonal cycle in AATSR LSTs over this site has already been reported by [6]. The results presented here are an update of these results, where the data set has been extended by one year and a more rigorous cloud screening has been employed. An equivalent analysis of the matchups for the Cardington and Niamey test sites suggests a similar seasonal bias at these locations (not shown). However, the cycles are not as clear due to the smaller number of matchups and the presence of some unexpected outliers (e.g. some cold-biased matchups that occur during the UK summer heatwave in 2006 at the Cardington site). Figure 1. Comparison between AATSR and in situ LST data over the ARM-Oklahoma field site in the USA (biome 12) for cloud-free overpasses between December 2003 and November 2006 (inclusive). The y-axis shows the difference between the AATSR and collocated in situ LSTs. The red, dashed lines represent the day-time 2.5 K target accuracy of the AATSR LST product, and the blue, dashed lines the night time 1.0 K target accuracy. 3. INVESTIGATION INTO SOURCES OF BIAS IN AATSR LST RETRIEVALS The results presented in Section 2 demonstrate that, although the AATSR LST retrievals and in situ LSTs are generally within reasonable agreement, there are some notable biases. In particular, the AATSR LST retrievals appear to have a tendency to be warm-biased in the summer and cold-biased in the winter. One major limitation of this kind of satellite-in situ data comparison experiment is that we are assuming that the in situ observation is accurate and truly representative of the kinetic LST over the area observed by the satellite sensor. In practise, this assumption will never truly be met because LST is so spatially and temporally variable. In order to verify the presence and nature of any seasonal bias in the retrievals, two sets of diagnostic experiments have been performed. Firstly, the apparent variation in bias has been examined with respect to LST, water vapour and atmospheric temperature using supplementary in situ data sets. Secondly, a line-by-line radiative transfer model has been used to simulate any bias in the AATSR LST retrieval scheme for a suite of atmospheric
and surface conditions. 3.1. Empirical Investigations Fig. 2 shows the variation in observed AATSR LST bias verses (a) in situ LST and (b) in situ microwave water vapour retrievals for the Oklahoma site. The degree of scatter in both sets of results is quite high and we obtain values for r 2 of 0.23 and 0.29 for day and night, respectively, for LST, and 0.22 and 0.34 for water vapour. Despite these low correlation coefficients, there does appear to be a relationship where apparent bias increases with both increasing water vapour and LST. In the case of the latter, this relationship is quite linear for matchups where the water vapour is less than 3 cm. Although less conclusive, given the number of matchups, a positive correlation between apparent bias and water vapour loading/lst is also indicated by the equivalent results for the Niamey and Cardington sites (not shown). Examination of European Centre for Medium-Range Weather Forecasts (ECMWF [7]) water vapour profiles corresponding to some of the Oklahoma matchups also supports a bias dependency in the AATSR LST algorithm on water vapour (Fig. 3). In addition, the ECMWF profiles also suggest that the accuracy of the algorithm is dependent on atmospheric temperature, where higher temperatures correspond to a warmer bias. This is perhaps not surprising given that LST, atmospheric temperature and water vapour are closely coupled. 3.2. Theoretical Results The empirical investigations into sources of bias in the AATSR LST algorithm discussed in Section 3.1 lead to the conclusion that the accuracy of the AATSR LST algorithm is dependent on water vapour, LST and atmospheric temperature. In this section, this conclusion is confirmed through the results of radiative transfer simulations of bias. We have employed the Oxford Reference Forward Model (RFM - see http://www.atm.ox.ac.uk/rfm/) to simulate one year of AATSR brightness temperatures (BTs) for actual cloud-free overpasses over 13 test sites, which include the ARM sites described above. The simulations utilise ECMWF atmospheric profiles and skin temperatures, and the actual AATSR channel response functions. After applying the appropriate AATSR LST retrievals coefficients to the simulated BTs, the difference between the resulting (simulated) AATSR LST and the original ECMWF skin temperature input into the model is the estimated retrieval bias under those conditions. As an emissivity corresponding to a surface where the fraction of green vegetation is constant at 0.5 is used in the model and in the (simulated) AATSR LST retrievals, any variation in simulated bias is solely due to variation in atmospheric conditions and/or true LST. Fig. 4 shows examples of the simulated AATSR LST bias over two sites, one of which corresponds to the Oklahoma Figure 5. Variation of simulated bias with ECMWF LST, water vapour and atmospheric temperature for the ARM- Oklahoma test site for one year, where red corresponds to day and blue to night time overpasses. test site. A similar seasonal bias to that apparent in the in situ comparison results is also observed in these simulations, and for all but two of the other 13 case studies considered in this experiment. Correlation of the simulated bias with the water vapour and LST variables used in the simulations yields that the bias is strongly dependent on these parameters. In general, the bias increases with increasing LST and water vapour (Fig. 5). The relationship is typically non-linear in each case; for water vapour, the relationship is quite linear up to approximately 3 cm precipitable water, which is is consistent with the empirical results shown in Fig. 2. The mean simulated bias and the amplitude of the seasonal cycle varies strongly with geographical location and the LST retrieval coefficients (which depend on land surface type and vegetation fraction). This is true even for sites that are classified as the same biome (e.g. Oklahoma and Cardington). In addition, the simulated biases are usually somewhat smaller than the observed biases (e.g. the examples for Oklahoma described in this paper). The exact reason for this is still under investigation, but is most likely to be due to differences between the conditions presented by
Figure 2. Relationship between observed apparent AATSR LST bias and (a) in situ LST (as observed by the site thermal infrared radiometer), (b) in situ water vapour (as observed by the site microwave radiometer). The red, dashed lines represent the day-time 2.5 K target accuracy of the AATSR LST product, and the blue, dashed lines the night time 1.0 K target accuracy. Figure 3. Examples of ECMWF atmospheric profiles for matchups over the ARM-Oklahoma field site between January and August 2005, for (a) water vapour and (b) atmospheric temperature. In each case the solid lines correspond to day time matchups and the dotted lines to night time matchups, where the lines have been coloured according to the observed AATSR minus in situ LST bias. Figure 4. Variation of simulated bias over (a) the ARM-Oklahoma test site and (b) the biome 2 test site at -097.500, 36.600 for one year, where red corresponds to day and blue to night time overpasses. The dashed lines indicate the target accuracy of the AATSR LST product.
the ECMWF data sets and the corresponding actual conditions on the ground. The ECMWF data are best-guess model data produced for a 2.5 degree grid, six times daily, such that an exact representation of, for example the Oklahoma test site, at the time of an AATSR overpass is not possible. In addition, the ECMWF data may not capture extreme weather conditions that may occur. Another source of the observed-simulated bias differences is undoubtedly due to the fact that the in situ data that have been compared with the observed AATSR LST data are not truly representative of the conditions on the ground within the AATSR FOV. This is particularly apparent during the day, when solar-heating is present and LST anisotropy is highest. The simulation results suggest that the seasonal bias in the algorithm should be quite similar for day and night. The fact that this is not seen in the observational results (day cycle amplitude is higher than night amplitude) suggest that LST anisotropy could also be a significant factor that is affecting the empirical comparison results. 4. CONCLUSIONS In general, AATSR LSTs show reasonable agreement with collocated in situ LSTs, with an apparent accuracy that is within or close to the target accuracy of the AATSR LST retrieval scheme. The agreement is significantly better at night, which can probably be attributed to higher LST anisotropy during the day, where the in situ LSTs are less representative over the area of the satellite pixel. The nature of the AATSR-in situ bias appears to be seasonal, where AATSR is warm during the summer months, and cold during the winter. Correlation of the observed bias with in situ observations of total column water vapour and LST for suggests that the accuracy of the LST retrieval algorithm is quite sensitive to these parameters. An overall increase in bias is observed with both increasing LST and water vapour loading. In addition, a dependency on atmospheric temperature is found, when the observed bias is examined with respect to corresponding ECMWF profiles. The presence of a seasonal bias has been confirmed through performing simulations of AATSR LST bias using a line-by-line radiative transfer model for a number of sites, including some where in situ LST data have been collected. The atmospheric/lst sensitivity of the AATSR LST retrieval scheme is also confirmed by correlating these parameters used in the model simulations with the simulated AATSR LST bias. In both the simulation and empirical results, the relationship between bias and water vapour loading is found to be quite linear up to approximately 3 cm. supplied by ESA and NEODC. We also wish to acknowledge the support of the UK Department of Environment, Food and Rural Affairs (Defra) who are the principle funding agency for the AATSR programme, with a contribution from the Commonwealth of Australia, in support of their programme of climate prediction and research. The in situ radiometric data acquired over the Oklahoma ARM site were supplied by the Atmospheric Radiation Measurement (ARM) Program. We are grateful to the British Atmospheric Data Centre (BADC) which provided us with access to the Met Office Cardington dataset and atmospheric data from the European Centre for Medium Range Weather Forecasts (ECMWF). REFERENCES [1] F. Becker and Z. Li. Towards a local split window method over land surfaces. International Journal of Remote Sensing, 11(3):369 393, 1990. [2] Z. M. Wan and J. Dozier. A generalized splitwindow algorithm for retrieving land-surface temperature from space. IEEE Transactions on geoscience and remote sensing, 34(4):892 905, 1996. [3] A.J. Prata. Land surface temperature measurement from space: AATSR algorithm theoretical basis document. Contract Report to ESA, CSIRO Atmospheric Research, Aspendale, Victoria, Australia, 2002. [4] Atmospheric Radiation Measurement Climate Research Facility. Infrared Thermometer Handbook, 2005. [5] ASTER Spectral Library. Reproduced from the ASTER Spectral Library through the courtesy of the Jet Propulsion Laboratory, California Institute of Technology, Pasadena California. Coypright c 1999. California Institute of Technology. All Rights Reserved., 1999. [6] E. J. Noyes, S. A. Good, G. K. Corlett, X. Kong, J. J. Remedios, and D. T. Llewellyn-Jones. AATSR LST Product Validation. In Proceedings of MAVT Workshop, Frascati, ESA (SP-615), March 2006. [7] ECMWF. The description of the ECMWF/WCRP level III-A global atmospheric data archive, 1995. ACKNOWLEDGMENTS This work was carried out under European Space Agency (ESA) contract 19054/05/NL/FF. The AATSR data were