The International Land-Surface Temperature Workshop was held on September 17-19, 1996, at the University of California at Santa Barbara. Jeff Dozier, Dean of the UCSB School of Environmental Science & Management, welcomed the participants. Twenty-five participants from the USA, France, Australia, and Japan attended the workshop. Twenty presentations were followed by two discussion sessions. It was a successful and productive workshop. The important findings of the workshop are outlined below together with the recommendations for further actions.
The workshop was part of a continuing effort to maintain contact among members of the EOS community that are concerned with the improvement of land-surface temperature (LST) algorithms, the definition of procedures for validation of LST, and the identification of the sources and the magnitude of measurement uncertainties. The specific goals of the workshop were to clarify the present state of the art in LST estimation from spaceborne sensors and to identify future directions, including issues requiring further research effort. A subsidiary goal was to establish a closer relationship between LST algorithm designers and the LST user community.
The importance of accurately determining LST to support an improved understanding of land-surface processes, including land-surface forcing, and the correlation of LST with the enhanced greenhouse effect, were some of the issues identified by Z. Wan in an overview paper titled "Challenges and opportunities for LST." The prospect of suitable data sets for LST research is soon to be enhanced by the impressive range of on-orbit sensors to be launched over the next few years. To advance the science, algorithm developers need to improve validation programs, collaborate more in the development and refinement of LST and land-surface emissivity (LSE) algorithms, undertake comprehensive and coordinated field campaigns, and forge closer relationships with General Circulation Model (GCM) scientists.
The technical aspects of the MODIS instrument design, and key role that it plays in the provision of accurately calibrated shortwave (SWIR) and longwave (LWIR) infrared radiances for LST research, were reviewed by C. Schueler of SBRS (Hughes) in a paper, "Technologies for temperature sensing from space." The specific algorithm proposed for application to MODIS to derive LST was presented in a paper, "MODIS generalized split-window LST algorithm," by Z. Wan and J. Dozier, who outlined the theoretical basis of the approach, the sensitivity and error analysis, and the results from validation campaigns conducted at Railroad Valley Playa, Nevada, with the MODIS Airborne Simulator (MAS). The algorithm assumed that the band emissivities for the surface under investigation were well characterized. According to simulations in wide ranges of atmospheric and surface conditions, the rms errors in retrieved LST were typically 0.7 K. A follow-on paper, "MODIS day/night LST algorithm for retrieving land-surface temperature and emissivity," by Z. Wan and Z-L. Li, proposed a MODIS day/night algorithm that has the ability to reduce the atmospheric effects caused by the uncertainties in atmospheric temperature and water vapor profiles in the process of simultaneous retrieval of surface temperature and band emissivities. Validation data over Railroad Valley Playa, Nevada, showed retrievals from MAS had an accuracy of 1K, but there is a significant difference between the retrieved emissivities and those measured from samples in the laboratory. A paper titled, "Thermal infrared surface radiance and its validation," was presented by F. Palluconi and addressed the role of ASTER in surface radiance measurement. The approach adopted applied radiative transfer methods to determine the radiance at the satellite. A sensitivity study concerning the impact of atmospheric temperature, water vapor, ozone, and visibility on the radiance was presented. Also described was a validation program which was conducted over instrumented lakes in California and Nevada.
A. Gillespie, T. Matsunaga, S. Rokugawa, and S. Hook, in a paper titled, "Temperature and emissivity separation from Advanced Spaceborne Thermal Emission and Reflectance Radiometer (ASTER) images," provided a description of a temperature and emissivity separation algorithm (TES) ultimately designed for application to ASTER. The approach involved an iterative scheme to remove the effect of downwelling sky irradiance reflected by the surface. From the validation data presented, it appeared that the scheme worked well over a variety of land-surface-cover types. Further, the derived spectral variation in emissivity compared well with in situ data acquired at the Railroad Valley Playa, Nevada, test site.
D. Ellement, M. Lynch, B. White, and I. Tapley presented "Land surface temperature estimation with AVHRR and numerical models applied to Western Australian field sites." With preset emissivities, split-window LST algorithms were developed that are accurate to about 1 K over several instrumented test sites. A model of the diurnal LST cycle was being evaluated and applied to a remote region in the north of Western Australia. T. Schmugge and C. Coll's paper, "Application of the TES algorithm to TIMS data from HAPEX-Sahel," described the application of the TES scheme to Thermal Infrared Multispectral Scanner (TIMS) imagery. The emissivity normalization derivation performed well but required a reasonably good first guess; for the emissivity min-max difference approach, the performance was comparable, but there was much less sensitivity to the first guess. Some difficulty was encountered for application to gray bodies.
A paper by M. Moriyama, "Error analysis of ASTER T/E Separation," described an implicit scheme which employed the covariance of the observations to estimate uncertainty in the retrieved surface variables. "Simultaneous determination of atmospheric correction parameters, LST and spectral emissivity from TIR multispectral data over land," by H. Tonooka used a multi-pixel method, based on scene modeling, to estimate sky radiance and the surface parameters. The scheme was applied to TIMS data, and the performance and limitations were discussed.
The important information content in high-spectral-resolution infrared radiometry was the key point of the paper, "Land surface temperature and emissivity estimation with high spectral/high spatial resolution sensors," by H. E. Revercomb, M. J. Lynch , L. E. Gumley, K. I. Strabala, and P. F. W. van Delst. High-spectral-resolution radiometry allowed the sampling of spectral regions in between atmospheric emission lines, where the atmosphere is highly transparent, and the downwelling thermal radiance is negligibly small. This approach permits a separation of the surface temperature and the spectral emissivity. "Validating remotely sensed land surface temperatures for surface radiation studies," by A. J. Prata, R. P. Cechet, I. F. Grant, and G. F. Rutter, outlined an impressive program that continued the development of a network of ground-truthing stations spanning the Australian continent designed to support validation and modeling studies using satellite data. The comprehensive data sets being acquired at existing sites were described. Finally, the additional information gained using the Along-Track Scanning Radiometer (ATSR) to evaluate LST was discussed and illustrated with examples using ATSR imagery over an Australian field site. S. N. Goward, R. O. Dubayah, K. P. Czajkowski, A. Waltz, and S. Liang, in a paper, "Validation of the split-window land surface temperature algorithms," outlined activities in the AVHRR Pathfinder program and the studies that they were undertaking in global primary production and modeling the surface energy budget. They compared the results of an analysis of the performance of 12 split-window algorithms applied to data sets from BOREAS, FIFE, and HAPEX-Sahel and undertook an estimation of the sources of error in the resulting LST products. The paper concluded with a discussion of the role of spatial scaling of data sets when statistics derived from a sensor of one spatial scale are compared with those derived from a sensor operating at a different spatial sampling scale (e.g., AVHRR and Landsat TM).
"MODIS and MAS LST field campaigns," by W. Snyder, Z. Wan, Y. Zhang, and Y. Feng, described field work conducted at Railroad Valley Playa on June 4, 1996, and outlined activities planned for a further BOREAS experiment later in 1996. These field campaigns were part of preparations for validation underflights of MODIS with MAS. Their error analysis showed that contributions of 0.3 K, 1.0 K, and 0.3 K were assignable to temporal, spatial, and calibration sources, respectively, giving an accumulated error of 1.09 K. The analysis of the error budget for vicarious calibration of TIR sensors was addressed in a paper, "Selecting appropriate sites for calibration of TIR sensors," by Z . Wan. Associated modeling studies, which assumed realistic uncertainties in the knowledge of the atmospheric state (3 K in temperature, 30% in water vapor profiles, 10% in water vapor absorption coefficients), were presented. For successful vicarious calibration, the sources of uncertainties (radiative transfer ~0.2%, surface emissivity ~0.003, measured LST ~0.88%, calculated radiances at the top of the atmosphere ~0.37%, ~0.71% and ~0.65%, for MODIS bands 29, 31, and 32 respectively) were expected to produce radiance rms errors of 1.01%, 0.79%, and 0.74% in these three MODIS TIR bands. A dry region in midwestern Tibet, and possibly in Bolivia, is the area where it is expected that these vicarious calibration accuracy requirements could be achieved.
The sole paper on the role of microwave radiometry, "Surface temperature estimation over land using satellite microwave radiometry," was presented by E. G. Njoku. After reviewing the key issues of concern in surface sensing (including surface soil moisture, soil/vegetation temperature, surface reflectance, vegetation canopy opacity and fractional cover, atmospheric opacity and mean temperature, and polarization), the performance of regression and non-linear iterative retrieval methods for temperature were presented. For a large simulated data set, with multichannel measurements and homogeneous conditions, these two methods can retrieve surface temperatures with RMS errors of 2.1 K and 0.4 K, respectively, for assumed radiometric noise of 0.2 K. However, the effects of modeling error and sub-pixel heterogeneity can be expected to increase the retrieval error significantly.
R. Dickinson, M. Jin, and X. Zeng, in a paper titled, "A dataset of land surface temperature diurnal cycle from MODIS data and CCM/BATS," and a related presentation by X. Zeng and R. E. Dickinson titled, "How to use skin temperature in land surface modeling--the consideration of surface sublayer," described the coupling of satellite skin temperature with the NCAR Community Circulation Model (CCM), coupled with the Biosphere-Atmosphere Transfer Scheme (BATS) over various land-cover classes in the model. The performance of the model-estimated skin temperature for the FIFE data set (July 1987) was presented. The measurement error in skin temperature ratioed to the skin - air temperature difference was identified as a key requirement for accurate model performance, including flux estimation.
"TIR BRDF measurements and modeling," by W. Snyder and Z. Wan, outlined laboratory facilities suitable for making measurements on samples collected in the field. The importance of translating the laboratory measurements on components to MODIS scene parameters was illustrated with a discussion of BRDF kernels and emissivity anisotropy as a function of zenith angle. A related paper by Y. Feng, Y. Zhang, and Z. Wan, "Measurement of the thermal infrared spectral emissivity of foliage," described improvements made to laboratory instrumentation and measurement procedures. The data were recorded over the 3-14 um range, but the band-averaged emissivities of vegetation canopies, for MODIS bands 31 and 32, showed that expected scene emissivities will vary over a very narrow range.
The meeting concluded with two review sessions chaired by S. N. Goward and F. Becker. The conclusions and recommendations are summarized below.
The prime task of the LST algorithms is to accurately correct both the atmospheric and emissivity effects in the TIR data for recovering LST. For land covers with high and stable emissivities, such as lake surfaces, snow, ice, and vegetation, split-window LST methods can be used to retrieve LST with surface emissivities estimated from ancillary information or prior knowledge. The coefficients of the split-window algorithm are derived from model simulations or field measurements conducted under certain atmospheric conditions. In high-humidity conditions the accuracy of split-window methods can be improved by adjusting these coefficients based on viewing angle and external assessment of the ranges of the atmospheric water vapor and temperature from satellite sounding, radiosonde, or meteorological analysis. Vegetation has a high value of, and little angular variation in, its emissivity in the split-window range (10-13 um) because the component emissivity of vegetation is already high and is increased further by its structural properties. In semi-arid and arid regions, the surface emissivity varies over a wide range. This can result in a significant error in LST retrieved by the split-window method.
Several multi-band and/or multi-temporal methods have been developed for retrieving surface temperature and emissivity simultaneously. These methods utilize the special capabilities of specific sensors in remote sensing of the surface TIR status. For example, ASTER has five bands in the 8-12 um range, MODIS has three bands in the 3.5-4.2 um range and four bands in the 8-13.5 um range, and the High resolution Interferometer Sounder (HIS) and Atmospheric Emitted Radiance Interferometer (AERI) provide high-spectral-resolution interferometric data in the 590-2750 cm-1 range. The benefit of high spectral resolution of the latter sensors does permit a separation of the reflected atmospheric downwelling radiance from the surface-emitted radiance because the atmospheric emission line structure is resolved. This will be a benefit for validation, but it will be some time before this capability exists on orbit. With the advances in TIR sensor technology and in LST algorithms, and with the synergism between LST products generated from data of different satellite sensors with mixed characteristics in spatial, temporal, and spectral resolutions, it is possible to provide LST products for global and regional studies.
It is essential to make comprehensive error and sensitivity analyses of LST algorithms over wide ranges of atmospheric and surface conditions. A common source of error occurs when the resampling or mapping is made to obtain LST values at geolocated grids from the LST field that is retrieved from airborne or satellite data by whatever LST algorithms. In most applications, LST values are required at geolocated grids for temporal analysis and for uses combined with other data. The size of this error depends on the gradient in the retrieved LST field and it may be significantly large near boundary areas. In such areas, mis-registration of day and night data would increase the error of the MODIS day/night LST algorithm. Numerical simulations of the mis-registration in areas where pixels are mixed with two components with different emissivities and at different temperatures show that the MODIS day/night LST algorithm still works well (the RMS error in retrieved LST values over wide ranges of conditions is smaller than or near 1K) as long as the uncertainty in registration does not exceed 20 percent. Therefore, it is proposed to use the MODIS day/night LST algorithm to retrieve surface temperatures and band emissivities initially at 5-km resolution (the resolution used in the MODIS product of atmospheric temperature and water vapor profiles is five by five 1-km pixels).
A clearer understanding of the applications of LST is needed. For instance, the LST accuracy needed in climate models is not a constant but a function of the surface-to-air temperature difference. More study is required regarding the relation between the LST retrieved from TIR data and the LST and the lower boundary fluxes in climate models. Also, spatial scaling plays an important role in global climate modeling. Study is ongoing as to how LST scales and on climate modeling. Further, polar satellite LST provides `snapshots' during the diurnal cycle that must be incorporated into climate modeling.
There is a need for more-conclusive in situ validation and accurate field measurement data that address sampling and instrumentation issues properly. Sampling a dynamic and spatial-varying, view-angle -dependent temperature field is often a dominant source of error. We need to consider combined use of radiometric and kinetic surface sensors and their placement. It is obvious that accurate field validation of LST can be made only over large flat uniform test sites and that comprehensive numerical simulations are needed to validate the inherent capability of LST algorithms in dealing with pixels mixed by components with different emissivities and at different temperatures. Significant improvement would result from the use of airborne sounders that are nadir-looking coincident with a scanner. There is also a need for long-term sites to establish accuracy under varying conditions and to provide data to a larger community.
More attention is needed for cirrus clouds and aerosols. For instance, the capability of cirrus detection at night may be questionable. Although aerosols play only a small role under normal conditions, there are certain areas with regularly high aerosol values that will bias the retrieved LST. The aerosol parameters in most atmospheric transmittance models may not be satisfactory and may require improvement via controlled field experiments.
There is continued potential for new and better LST algorithms. These algorithms will motivate and follow the development of cheaper and better instruments--for instance, high-spectral- and high-temporal-resolution sensors. For self-contained algorithms, an increase in the number of bands will allow better estimates of the atmospheric characteristics and possibly reduce the sensitivity to land-surface spectral emissivity if the signal-to-noise ratio of observation data is large enough.
Improvements for external methods will consist of the incorporation of assimilation data over time. Also, it is expected that sounder data will become more common and more accurate. This will provide the atmospheric profiles of temperature and water vapor needed for LST recovery. Passive microwave instruments provide a valuable, independent assessment to incorporate into LST algorithms. But it is important to understand the physical difference between the surface "skin" temperature measured by TIR sensors and microwave-measured temperature in real applications. The accuracy of LST estimated from microwave data is limited by the uncertainties in surface emissivity, which is affected by surface moisture variations. This would be improved with longer wavelength channels in future instruments.
The accuracy and role of geostationary sensors for providing higher temporal sampling of land temperature should be investigated. Such sensors offer a higher probability of achieving cloud-free conditions for a given location and also would provide data sets at a time more appropriate for model assimilation.