Overview..
The use of satellite remote sensing to monitor changes in biospheric processes, such as phenology, primary production, and evapotranspiration, requires frequent repeat observations. The only currently operating remote sensing platform capable of providing the temporal frequency required for monitoring such processes is the Advanced Very High Resolution Radiometer (AVHRR) instrument onboard the NOAA series of meteorological satellites. Unfortunately, AVHRR data may be at a coarser spatial resolution than some applications require. Thus it is important to be able to utilize techniques that improve the comparability of data acquired from different sensors, allowing them to be used interchangeably or to augment spatial and temporal observation quality.
One means towards this end utilizes transformation of at-sensor radiances to spectral vegetation indices (SVIs), which are generally more sensitive to vegetation properties (leaf area and light interception) than to abiotic factors (clouds, aerosols, soil and background reflectance). There have been many papers reporting the performance of SVI transformations since the early work of the late-1970's. Much research suggests that SVI transformations are not a complete solution, as they are frequently still affected by factors other than the vegetation properties that are sought. Similar problems exist for techniques designed to estimate radiometric surface temperature (Ts).
Comparability issues between sensors extends to the relationship between SVIs and Ts, which has been utilized for a number of applications, including the estimation of evapotranspiration, energy balance components (latent and sensible heat flux), surface moisture status, and more recently, air temperature and landcover classification. It is thus important to be able to characterize and to understand sources of variation in SVI-Ts relationships at a variety of scales, both temporal and spatial. Several studies have demonstrated sensitivity to variations in fractional vegetation cover, soil background temperature, thermal inertia, and topography. Some results suggests that extension of processes inferred from SVI-Ts relationships to regional scales may require use of ancillary information or augmentation with simulation models.
Examples from a Grassland Site..
Using a data set collected in a central United States grassland that was the site of the First International Satellite Land Surface Climatology Project (ISLSCP) Field Experiment (FIFE) demonstrated that surface reflectance and Ts retrieved from radiometrically calibrated and atmospherically corrected radiances measured by high resolution satellite sensors, such as the Landsat Thematic Mapper (TM) and System Pour l'Obervation de la Terre (SPOT) Haute Resolution Visible (HRV), were comparable to observations from aircraft and near-surface instruments (see an abstract of one such paper).
The objectives of this study were to extend these analyses to an investigation of the comparability and temporal dynamics of surface reflectance, the normalized difference vegetation index (NDVI), and Ts derived from a number of multi-temporal, multi-resolution satellite and aircraft sensors through several growing seasons, and to examine variation in the relationship between NDVI and Ts with respect to the spatial and temporal properties of the observations.
The Imagery and Correction Techniques
The data used in this analysis were collected during field campaigns conducted as part of FIFE between 1987 and 1989 in central Kansas. The 225 km2 study area is a complex mixture of C3-C4 grasslands (~85% of the site), shrublands, riparian forest corridors, agriculture, and non-vegetated areas. Overlying the land cover types were a range of management practices that include grazing, burning and crop rotation. Figure 1 is a composite of three Landsat Thematic Mapper NDVI images of the study area acquired in August of 1987, 1988, and 1989. Each NDVI image was calculated from atmospherically corrected land surface reflectance in the visible (Vis) and infrared (IR) wavelengths, where NDVI = (IR-Vis)/(IR+Vis). Roads are black (low NDVI in each scene) whereas riparian forest corridors and some agricultural areas are white (high NDVI in each scene). A great deal of landscape heterogeneity and inter-annual variability is visible in the composite image, ranging from variations in fire history and topography, most noticeable in the upper left quadrant of the image (the Konza prairie long term ecological research station), to variations in grazing history and agricultural practices throughout the remainder of the image. Inter-annual variability was introduced by a wetter-than-normal year (1987) followed by a drought year (1988) and a recovery period (1989). Variability captured in this image is discussed further below.
Atmospheric attenuation of different sensor bands varies with their relative placement in different portions of the wavelength spectrum. The figure below illustrates this graphically using an on-site atmospheric sounding profile in a simulation from the LOWTRAN-7 atmospheric radiance model. Typical reflectance spectra of vegetation is also shown, illustrating the between-channel (visible, infrared) differences that provide the discriminant power of the NDVI. It is clear from Figure 2 that differences in sensor wavelength bands may result in different values of the NDVI, not only as a result of different atmospheric effects but also from variations in leaf reflectance, predominantly in the visible wavelengths. Despite these differences among sensors, the current study suggests that this extends to other sensors for a range of conditions (discussed below).
In the atmospheric continuum, oxygen and water vapor absorption bands in AVHRR channel 2 (0.71-0.98 mm) are larger than in other commonly used satellite sensors. However, relative attenuation in the AVHRR thermal channels (Figure 2b) can be used to compensate for this effect by estimating the total column abundance of precipitable water (PW) resulting from differential attenuation of the at-sensor signals in the 10.3-11.3 mm region (channel-4) and the 11.4-12.3 mm region (channel-5). The combined effect of decreased atmospheric transmission and increased thermal emission with increasing optical path length results in a near-linear relationship between the radiometric temperature difference of the thermal channels and PW.
Multi-Sensor, Multi-Date Comparisons of Surface Reflectance and NDVI
Variations in AVHRR-LAC surface reflectance with changes in viewing geometry are shown over a growing season period in the next plot. The large variation in surface reflectance is typical of multi-angle reflectance data sets, in which the directional reflectance properties of the surface relative to the sun and sensor result in a large increase in surface reflectance off-nadir, particularly in the backscatter direction. Forward scatter reflectances were notably lower than those in the backscatter direction, which is typical for small view zenith angles but not necessarily for larger viewing angles (i.e., > 30 degrees). This characteristic of the data was due to the orbital geometry of the NOAA satellite, in which acquisitions viewing in the forward direction were acquired earlier in the day when solar elevation angles were low. The lowest reflectances in both the visible and infrared were in the range of 10 to 20 degrees in the forward direction.
Surface reflectance for those LAC scenes selected by maximum value compositing are represented by solid circles (visible) and triangles (infrared).
Compositing tended to select acquisitions in the forward scatter direction as these had the largest NDVI values within each 10 day period. This result was consistent with previous characterization of maximum NDVI values. Thus the technique reduced variations in NDVI that were a result of viewing conditions (i.e., sun-sensor geometry) rather than surface properties, as it was initially designed to do.
The differences in NDVI among sensors, over a time period when all were available (i.e., day 130 - 300 in 1987), were significant when calculated from radiance and from exoatmospheric reflectance, but not when calculated from atmospherically corrected reflectances. Thus the corrections inherent in the calculation of exoatmospheric reflectance (i.e., differences among in-band solar spectral irradiance, solar angle, and variability in Earth-Sun distance) significantly improved the comparability of NDVI, but the atmospheric corrections improved the comparability even more and were essential for the multi-sensor analyses that followed.
NDVI calculated from surface reflectance for the LAC composite selections are shown together with the SPOT-HRV and Landsat-TM NDVI in the next plot.
There are several items of note in this figure.
Multi-Sensor, Multi-Date Comparisons of Ts
Results suggests that, as with the multi-sensor NDVI, radiometric Ts can be retrieved from remotely sensed data with a good degree of confidence, providing the data are properly calibrated and atmospherically corrected. However, due to a history of inconsistent results, further analyses of remotely sensed thermal data sets are needed (see an abstract of a related paper).
Use of the Relationship between Surface Temperature and Vegetation Indices
The relationship between retrieved surface temperature and the NDVI vegetation index, called the TvX slope, is shown in the next figure for several dates that characterize most of the variability through the growing season. Differences between the TvX slopes were related to site-averaged soil moisture, and partitioning of energy fluxes into sensible and latent heat (the evaporative fraction, EF), with the steepest slope (-41.1 on July 19) associated with low soil moisture (19.8%) and relatively low EF (0.66) and the shallowest slope (-17.0 on August 15) associated with a high soil moisture (34.4%) and high EF (0.75). On those days with coincident site-averaged surface measurements the TvX slope predicted 65% of the variability in soil moisture, but just 41% of the variability in EF. The different results with soil moisture and EF were likely a result of the small number of days in 1987 on which soil moisture fell below the 20% threshold that results in strong plant physiological control. For example, while the association of TvX, soil moisture, and EF was distinct on days when physiological control was induced (such as July 19), the majority of other days were intermediate in all variables (TvX slope, soil moisture and EF). As a result, all the variables were characterized by a smaller range than would be encountered in a more typical (less wet) year. This view was supported by the results for other years from different instruments (i.e., Landsat TM and the aircraft sensors), which all had steeper TvX slopes associated with reduced LAI and soil moisture.
Summary and Conclusions
An extensive remotely sensed data set, recently made available to the scientific community, was used to examine the comparability of multi-temporal, multi-resolution imagery acquired from a suite of satellite and aircraft sensors. Differences among calibrated, atmospherically corrected, and site-averaged multi-sensor NDVI were not statistically significant between sensors, despite different radiometric and observational acquisition characteristics. When extended to the utilization of sensors in an integrated manner, short-term variability in vegetation properties (e.g., green LAI) was captured. Differences in surface radiometric temperatures among sensors were also insignificant and retrieved values were in good agreement with near-surface measurements (within 2.8 degrees C).
The relationship between Ts and NDVI (the "TvX" slope) characterized by the various sensors was largely determined by seasonal changes in vegetation cover and soil moisture, and to a lesser extent, time of day. There was greater sensitivity of Ts to variation in net radiation and available soil moisture than there was in NDVI, which reflects a lag time in vegetation response to site conditions. Variations in the TvX slope were also related to vegetation physiological control on the partitioning of net radiation into latent and sensible heat fluxes, with steeper slopes during periods of reduced soil moisture and vegetation amount. Observed comparability among sensors suggests that more frequent temporal characterization of changes in the slope of the TvX relationship, both diurnally and through the growing season, may be possible with the use of multi-sensor data.
These results suggest high resolution (230m) remotely sensed data can be augmented with inexpensive, frequent temporal observations (e.g., AVHRR acquisitions) to monitor biophysical variables, provided the data are radiometrically calibrated, atmospherically corrected, and characterized with respect to viewing geometry. Further exploration of coordinated multi-sensor data sets is needed in areas of even greater spatial heterogeneity.
The research discussed here has been published in the following publications:
S.J.Goetz, 2002, Recent advances in remote sensing of biophysical variables: an overview of the special issue, Remote Sensing of Environment, 79(ER2-3): 145-146.
S.J.Goetz, 1997, Multi-sensor analysis of NDVI, surface temperature and biophysical variables at a mixed grassland site, International Journal of Remote Sensing, 18(1):71-94.
Other papers that may be of interest include:
Goetz, S. J. and S. D. Prince. 1996. Remote sensing of net primary production in boreal forest stands, Agricultural and Forest Meteorology 78 (3): 149-179.
Goetz, S. J. and S. D. Prince. 1998. Variability in carbon exchange and light utilization among boreal forest stands: implications for remote sensing of net primary production,, Canadian Journal of Forest Research 28(3) : 375-389.
Goetz, S. J. and S. D. Prince. 1998. Modeling terrestrial carbon exchange and storage: evidence and implications of functional convergence in light use efficiency,, Advances in Ecological Research 28 : 57-92.