Land use applications are expected to be among the major drivers for the selection of radar parameters - wavelengths, polarisations, and spatial resolutions - for future satellite SAR missions. Data from SIR-C / X-SAR are important inputs for defining such requirements and have given us valuable insight into the particular needs of agricultural mapping and the retrieval of forest biomass.
Our work is to be seen in the context of other studies we have been conducting, including investigations with airborne instruments and other satellite SARs. It is of particular importance to understand how polarimetric and / or multifrequency data may be used as part of a multi-temporal strategy with a permanently orbiting satellite, such as ESA's Envisat which is currently under construction, or NASA's LightSAR concept.
Our principal agricultural site is at Feltwell in the southeast of England. Feltwell was previously the target for multi-frequency AIRSAR experiments in the summers of 1989 and 1991, as well as multi-temporal campaigns with ESA's C-band ERS-1 satellite SAR in 1992 & 1993. Forestry studies have previously centred upon the Thetford Forest site, adjacent to Feltwell. Thetford was again imaged as part of the SIR-C / X-SAR missions, but studies have been expanded to include natural and regenerating tropical forest sites at Tapajos and Manaus in Brazil.
In Section 2, we describe the agricultural experiment, and the key results obtained from SIR-C/X-SAR. We consider the possible importance of the results in the wider context of the increasingly operationalised application of agricultural mapping by radar. In Section 3, we describe the forestry studies and place SIR-C results alongside those from JERS-1 in order to reinforce the growing consensus on appropriate radar designs for forest applications.
The Feltwell site was successfully imaged by a series of data-takes from the SIR-C/X-SAR missions in both April and October 1994. The site is a flat and intensively managed arable region spanning the edge of the drained peat fenlands of eastern England. Major crops include grains, potatoes, sugar beet and carrots, with a diverse selection of minor vegetable crops and grass. Ground activities for both missions were aimed at generating crop maps, and providing detailed characterisation of the soil and plant conditions in a small sub-set of the imaged fields. Plant structural data and soil roughnesses were collected in ten fields, and soil moisture data in those and a further 20 fields. The measurements were specifically intended to drive modern physical microwave scattering models, including the second-order radiative transfer model RT2 developed by members of the team (Saich et al 1995).
An apparent drawback to the experiments was their timing relative to the main growing season in England. For the April mission, most fields were seedbeds in various stages of preparation, with the principal exception of the winter grains and grass fields. The October mission took place after grain harvest, and during harvesting of sugar beet. It will be seen, however, that the early-season April mission has provided one of the key results.
Radar signatures have been analysed on a field-averaged basis using manual designation of field boundaries in a vector database. Discrimination of specific crops using those radar signatures has been assessed alongside results derived previously for this site in the 1991 AIRSAR deployment. Results have been supported by radiative transfer modelling.
The analyses of mean backscatter signatures from SIR-C/X-SAR has shown that:
Figure 1 illustrates this last point
by showing the scatter in amplitude of the complex HH-to-VV correlation
coefficient for fields imaged in the April mission. Bare soil
surfaces display high correlations between the HH & VV channels,
whereas in the presence of even a sparse and low vegetation canopy
the relative phases between the channels appear to have been significantly
randomised. This interpretation is supported by simulations
with the RT2 model which illustrate the greatly differing sensitivities
of L and C-band polarimetric radars to the presence of sparse
grain crop canopies (Figure 2). These early-season effects are
in marked contrast to the mid-season findings on crop discrimination
from the July 1991 AIRSAR experiment. There (Figure 3) it was
found that L-band provided a more effective single-wavelength
discrimination capability than C-band. The AIRSAR result is perhaps
not surprising given that it took place under conditions where
the penetrative capability of L-band radiation may be better matched
to the range of crop biomasses and structures present at that
time. It was found in 1991 that crop discrimination, even in
mid-season, was greatly enhanced through the use of two-frequency
C+L band data, and the SIR-C early-season observations continue
to point to the complementarity of a two-frequency polarimetric
radar incorporating L and C bands. It is a significant frustration
that no polarimetric X-band data were collected in the SIR-C/X-SAR
missions to demonstrate a possibly similar capability to C-band.
SIR-C / X-SAR is providing invaluable information for the design of more optimised satellite SARs for agricultural applications. Within the limitations imposed by its spatial resolution (i.e. so that we are broadly restricted to considering field-level effects rather than within-field variations) it indicates that we should expect future instrument designs to provide very significant improvements over current capabilities.
At present, exploiters of the European ERS satellites have used a multi-temporal strategy to construct temporal signatures for the C-VV backscatter behaviour of crops as they grow and ripen (Wooding et al. 1995). We may anticipate that the dual-polarised ASAR to be carried on Europe's Envisat (providing HH & VV or Co & Cross polarisations) will continue that strategy with improved results. However, even without a clear multi-temporal demonstration, it is apparent from our AIRSAR and SIR-C results that a fully-polarimetric C-band radar would be likely to provide very significant improvements in mapping capability over ERS. Indeed, the success of C-band polarimetry at very low biomass levels might be particularly attractive in the early-season mapping strategies. Going beyond a consideration of just C-band, we have seen that AIRSAR's L-band polarimetric sensor has an even more effective capability in a single-date mid-season trial, and a C+L-band combination a more effective capability yet.
The difficulty facing us at present
is in quantifying the expected enhanced capabilities of new sensor
designs when used as part of a realistic multi-temporal imaging
strategy. To what extent do we need a "one-shot" mapping
capability when we might derive high-quality classifications over
a number of revisits ? We do not have a quantitative answer to
that question. ERS-1 and 2 clearly do not give us all the discrimination
capability that we might want, even using multi-temporal imaging,
but do we really demand a C+L-band (or X+L-band) polarimetric
system ? The answer may be expected to lie in the needs of developing
applications for agriculture which go beyond simple crop identification.
If SAR is to be an effective tool for providing information to
the farmer rather than to government, and is to incorporate higher
spatial resolutions for imaging variations within fields, then
the greatly increased structural discrimination provided by fully-polarimetric
and possibly dual-frequency radars may be demanded.
Thetford Forest is a well understood and managed temperate plantation and is currently our test site with the most complete database. In addition to UK Forestry Commission inventory data and biomass data from ground survey (Baker & Mitchell 1992), the database contains georeferenced SAR data from AIRSAR (1989 and 1991), a series of 15 ERS-1 scenes (1992 and 1993) together with one JERS-1 scene (1993). Most recently it has been augmented by two sets of SIR-C data (1994). The individual forest stands are segmented in the database and form natural units for the analysis of imagery covering homogeneous areas of known age.
The Tapajos region of Para State in Brazil contains a variety of land cover including cattle ranches, subsistence farming, regenerating abandoned plots, and protected mature forest. Measurements of tree height, diameter and species composition were made in 1994 over areas representing a range of ages of regeneration from new regrowth to mature forest. Biomass densities were derived from 0.05 hectare plots by employing timber specific gravity values in published species-dependent regression equations (Luckman et al 1996). Figure 4 shows four georeferenced data sets from Landsat, ERS-1, JERS-1 and SIR-C over the Tapajos test site, the sites of ground biomass survey being indicated on the Thematic Mapper image.
The region to the north of Manaus is characterised by abandoned ranches, either side of the BR174 highway. Fieldwork in August 1993 and August 1995 concentrated on the Fazendas Agroman, Maringa, Dimona, Porto Alegre and Esteio (Lucas et al 1995). A similar measurement strategy to that at Tapajos was adopted, over 16 sample plots of 0.1 hectare.
SIR-C imagery of Thetford has been used to reinforce and extend results demonstrated previously with AIRSAR imagery (Baker et al 1994):
Analyses of Tapajos and Manaus SIR-C and JERS images demonstrate:
In both the UK and Brazillian test sites, the sensitivity of L-band SIR-C and JERS datasets to vegetation biomass variations has been demonstrated, although the existence of a universal biomass "calibration curve" covering managed coniferous forests and tropical forests is not supported. In view of the greatly differing structures and diversities of the forest components in the two regions this is not surprising.
In Brazil, L-band data discriminates
mature forest from newly felled forest and from agricultural land
uses, and can track a number of classes as abandoned agricultural
land reverts / regenerates to forest again. That number of classes
is small however (two or perhaps three) but such classification
does provide a very useful input to models of forest change and
ultimately to models of the carbon cycle and climate change.
The limitation to low biomass levels before L-band backscatter
saturates is not necessarily a major drawback because carbon uptake
in tropical forests is most rapid during the early stages of regrowth
(Uhl et al 1988). Nevertheless the saturation level at L-band
at around 50-80 tonnes per hectare is much lower than is desirable
for more local forest management applications. On grounds of
dynamic range there is still a desire for yet longer wavelengths
to be used in the monitoring of biomass change.
The agricultural studies with SIR-C
/ X-SAR data have been supported by the BNSC and Ministry of Defence.
The RT2 model was developed with the support of BNSC through
contract RAE 1b/89 and by ESA through contract 10644. Our forestry
studies have been supported by the Natural Environment Research
Council through the TIGER programme (Terrestrial Initiative in
Global Environmental Research). Field work in Brazil was conducted
in collaboration with Instituto Nacional de Pesquisas Espaciais
(INPE).
References
Baker, J.R. & Mitchell., P.L. " The UK Element of the MAESTRO-1 SAR Campaign", Int. J. Remote Sensing, Vol 13., pp 1593-1608, 1992.
Baker, J.R., Mitchell, P.L., Cordey, R.A., Groom, G.B., Settle, J.J. & Stileman, M.R., "Relationships Between Physical Characteristics and Polarimetric Radar Backscatter for Corsican Pine Stands in Thetford Forest, UK", Int. J. Remote Sensing, Vol. 15, pp. 2827 - 2849, 1994.
Luckman, A.J., Baker, J.R., Kuplich, T.M., Yanasse, C.C.F. & Frery, A.C., "A Study of the Relationship Between Radar Backscatter and Regenerating Forest Biomass for Spaceborne SAR Instruments", Remote Sensing of the Environment (In Press)
Lucas, R.M. & Honzak, M., "Secondary Forests at Manaus: Data Collected During a Field Campaign, July - August 1993", Report to INPE and INPA, University of Swansea, 1995.
Saich, P.J., Cordey, R.A., Quegan, S., Williams, M., Wielogorska, A., Baker, J.R., Luckman, A. & Wooding, M.G., "SAR Retrieval Algorithms for Land Applications", GEC-Marconi Research Centre MTR 95/36A for ESA Contract 10644, 1995.
Uhl, C, Buschbacher, R. & Serrao, E.A.S., "Abandoned Pastures in Eastern Amazonia. 1. Patterns of Plant Succession", Journal of Ecology, vol. 76, pp 663 - 681, 1988.
Wooding, M.G., Attema, E., Aschbacher,
J., Borgeaud, M., Cordey, R.A., de Groof, H., Harms, J., Lichtenegger,
J., Niewenhuis, G., Schmullius, C. & Zmuda, A.D., "Satellite
Radar in Agriculture: Experience with ERS-1", ESA SP-1185,
1995.
Figure Captions
Figure 1: Scatter plot of field-averaged backscatter from the April 1994 SIR-C experiment at Feltwell. The plot shows the very clear distinction in C-band HH-to-VV correlation amplitude between bare field surfaces and those containing even a minimal crop cover.
Figure 2: Output from the RT2 radiative-transfer model showing the predicted sensitivities of HH-to-VV correlation amplitude as a function of emergent grain crop height for C and L bands. Very low levels of interaction with the crop are predicted at L-band, while the C-band predictions show a rapid fall-off in correlation amplitude as interaction increases.
Figure 3: Pairwise crop cluster separabilities in multi-dimensional radar parameter space from the 1991 AIRSAR deployment at Feltwell. Clusters consisted of field-averaged backscatter estimates over five crop classes (a total of more than 600 fields). The separability measure is the Jeffries-Matusita distance, normalised to unity for notional perfect separation. The radar parameters shown on the X-axis are a selection intended to show trends with increasing numbers of polarimetric channels and changing radar wavelength.
Figure 4: Multi-sensor comparison of the Tapajos experiment site in Brazil. The Landsat TM image identifies fieldwork sample locations. The ERS and JERS images are both multi-temporal composites, while the SIR-C image is a single-date multi-parameter composite.
Figure 5: Plot of L-band radar backscatter vs. forest biomass density for the Tapajos sites. The plot includes a comparison of SIR-C and JERS results for their respective L-HH channels, and also the SIR-C L-HV channel which displays a higher dynamic range in backscatter.
Figure 6: Plot of L-band radar backscatter vs. forest biomass
density for the Manaus sites. The results shown here are not
significantly different from those for Tapajos (Figure 5)