Land Applications using ERS-1/2 Tandem data
Based on ERS-1 data acquired during 3-day repeat orbits a good potential of repeat-pass SAR interferometry for land applications such as landuse classification and change detection was identified (Wegmüller et al., 1995a, 1995b). It was found that the interferometric correlation is not just a measure of the phase noise of the interferogram but a valuable source of information on scene properties. With the ERS-1/2 Tandem mission repeat-pass SAR data useful for interferometric analysis became widely available: the 1-day acquisition time interval and the precise orbit control which allowed to almost permanently obtain short interferometric baselines for Tandem pairs resulted in data ideal for interferometry. In addition the 35-day repeat-orbits of the two satellites allowed to achieve nearly global coverage. With the available data interferometric techniques can now be widely applied. In this contribution it will be shown that Tandem data is very useful for landuse classification. In addition potential and limitations of Tandem data for change detection and monitoring will be discussed.
The Interferometric Correlation
The interferometric correlation is a measure of the phase noise of the interferogram. It depends on sensor parameters (wavelength, system noise, slant range resolution), parameters related to the imaging geometry (interferometric baseline, local incidence angle), and target parameters. Volume scattering and temporal change (i.e. random motion of the scatterers, change of the scatterers) decrease the interferometric correlation. The system and geometry dependent effects are pretty well understood and can be controlled by appropriate interferometric processing, as long as the system parameters are within a certain interval. The baseline dependence of the interferometric correlation, for example, may be eliminated in many cases by common spectral band filtering of the range spectrum.
The dependence of the interferometric correlation on target parameters is used to retrieve information on the target characteristics. The information available through the interferometric correlation is complementary to the information content of the backscatter coefficient as investigated and documented by Wegmüller et al. 1995a, 1995b.
The maximum likelihood estimator used to estimate the interferometric correlation tends to be biased at low interferometric correlation values. For an unbiased estimation a large estimation window is required at low interferometric correlation values as shown in Figure 1.
Figure 1: Dependence of interferometric correlation estimate on number of looks used in the estimation for areas representative for low to high interferometric correlation.
The landuse classification scheme presented is based on the normalized interferogram and the two backscatter intensity images of the interferometric image pair. In a first step these data are used to estimate (1) the interferometric correlation, (2) the average backscatter intensity, (3) the backscatter intensity change, and (4) the texture of one of the backscatter images. For a wide applicability the classification is done on a per pixel level. Therefore, appropriate estimation schemes and filtering are required to retrieve useful estimates at per pixel level.
As mentioned above the estimation of the interferometric correlation requires a sufficient number of looks. As a compromise between maintaining a high spatial resolution and accurate estimation an estimator with adaptive estimator window size was used. In a first step the interferometric correlation was estimated with a fixed, relative small estimator window size. In a second step the estimator window size was determined based on the first correlation estimate. To estimate lower correlation values larger estimation windows are used. Usually, the estimator size was varied between 3 x 3 and 9 x 9 pixels of a 5-look interferogram. In addition, a weighting function, decreasing linearly with increasing distance, was applied in the estimation. For test areas representing areas of low to high interferometric correlation the adaptive estimator is compared to non-adaptive estimation with estimation window sizes of 1 x 1 and 9 x 9 pixels of the 5-look interferogram (Figure 2).
Figure 2: Comparison of histogram of estimated interferometric correlation for areas representative for low to high interferometric correlation using (a) non-adaptive 3 x 3 pixel (b) non-adaptive 9 x 9 pixel and (c) adaptive 9 x 9 pixel estimators.
In order to reduce speckle effects and obtain a backscatter intensity estimate at the pixel level which is representative for the ensemble average of the area around that pixel the two registered SAR images of the interferometric pair are averaged. Minimum Mean Square Error (MMSE) filtering, as described by Frost et al., 1982, was then applied to the averaged image. Typically, the filter was applied to areas of 7 x 7 pixels of a 5-look image (5 azimuth looks). For the same test areas as used for Figure 2 the histogram of MMSE filtered backscatter intensity is compared to average filtered data (Figure 3).
Figure 3: Comparison of histogram of estimated backscatter intensity for areas representative for different landuse classes using (a) average filtering with 1 x 1 pixel window, (b) average filtering with 7 x 7 pixel window, and (c) MMSE filter with 7 x 7 pixel window.
The backscatter intensity change between the two images of the interferometric pair is defined as the absolute value of the ratio between the two images expressed in the logarithmic dB scale
The brackets stand for the ensemble average. Due to the speckle of the individual images it is essential to average sufficiently before the calculation of the ratio. Usually, at least 9 x 9 pixels of the 5-look images were used in the averaging step. In addition, a weighting function, decreasing linearly with increasing distance, was applied in the estimation. In the classification change is only distinguished versus no change. Therefore it is reasonable to take the absolute value of the ratio in dB. If the change is further interpreted this is of course not reasonable as the capability to distinguish backscatter decrease from backscatter increase is wasted. For the same test areas as used for Figure 2 the histograms resulting from different ratio estimators are compared (Figure 4).
Figure 4: Comparison of histograms of estimated backscatter intensity
ratios for areas representative for different landuse classes using (a) ratio of
unfiltered 5-look values, (b) after average filtering with 9 x 9 pixel window,
and (c) after average filtering with 9 x 9 pixel window and application of
linearly decreasing weighting function.
The texture of the backscatter image is defined as the ratio between the standard deviation and the average
Again, the estimation of the ensemble averages requires sufficiently large estimator windows. It turns out that extremely strong scatterers in an image have the unwanted effect the entire high texture is obtained over an area corresponding to the size of the estimator window. This effect can be avoided to some degree if the texture estimation is followed by filtering with a moving average filter of larger size than the texture estimator. As an example we used 15 x 15 pixels of the 5-look image for the initial texture estimation with a subsequent 25 x 25 pixel moving average filtering. In addition, weighting functions, decreasing linearly with increasing distance, were applied in the different steps. For the same test areas as used for Figure 2 the histograms resulting from different texture estimators are compared in Figure 5.
Figure 5: Comparison of histograms of estimated backscatter intensity
textures for areas representative for different landuse classes using (a)
texture estimate using 5 x 5 pixel window, (b) texture estimate using 15 x 15
pixel window, and (c) texture estimate using 15 x 15 pixel window with
application of linearly decreasing weighting function.
Based on the interferometric signatures a simple landuse classification
algorithm was developed. A hierarchical decision tree algorithm using the
criteria listed in Table 1 allowed to generate a landuse map. In order to
account for the specific conditions under which the data was acquired the
classification scheme needs to be slightly adapted.
Table 1. Decision rules of landuse classification algorithm. The criteria are applied hierarchically, in the order as listed. The value ranges used are indicated for the interferometric correlation (), the average backscatter intensity of the two images (), the backscatter intensity change between the two images, and the texture of the first backscatter image.
For comparison results based on ERS-1 data acquired during 3-day repeat orbits in November 1991 are used. The November 1991 data is ideal for the presented approach because of the relatively short 3-day acquisition time interval, the short 58 m baseline, and the acquisition during the winter season when the forest can best be distinguished from agricultural fields because the fields are bare or only sparsely covered with vegetation. The RGB color composite of the interferometric correlation (red), the backscatter intensity (green), and the backscatter change (blue) is shown in Figure 6, and the resulting landuse classification in figure 7. These data are compared with results achieved using Tandem data in November 1995, April 1996, and July 1995. The exact dates and baselines are listed in Table 2:
Table 2: Bern (CH) test site: Dates and baselines for interferometric pairs used.
The RGB composite (Figure 8) and the landuse classification (Figure 9) of the November 1995 Tandem data confirm the expected usefulness of Tandem data for landuse classification. Unlike with 3-day repeat data the approach lead to reasonably good results during spring (Figure 10,11) and summer (Figures 12,13) period, too. The shorter acquisition time interval results in an increase of the interferometric correlation of fields with grass or crops, improving the potential to distinguish fields from forest.
For the November 1991 data the result was validated with a conventional forest map. An accuracy for the forest/non-forest classification of around 90% was reported by Wegmüller et al. 1995c. Classification accuracy of the same order may be expected with Tandem data.
These examples allow to conclude that landuse classification based on interferometric signatures from ERS-1/2 Tandem data is feasible and has a high potential not only because of the quality of the results which may be achieved, but also because of the good spatial and temporal coverage with appropriate image pairs.
Figure 6: Bern (CH), ERS-1, 24/27 Nov. 1991: RGB composite of interferometric correlation (red), backscatter intensity (green), and backscatter change (blue).
Figure 7: Bern (CH), ERS-1, 24/27 Nov. 1991: Landuse classification based on SAR interferometric signatures.
Figure 8: Bern (CH), ERS-1/2, 26/27 Nov. 1995: RGB composite of interferometric correlation (red), backscatter intensity (green), and backscatter change (blue).
Figure 9: Bern (CH), ERS-1/2, 26/27 Nov. 1995: Landuse classification based on SAR interferometric signatures.
Figure 10: Bern (CH), ERS-1, 14/15 Apr. 1996: RGB composite of interferometric correlation (red), backscatter intensity (green), and backscatter change (blue).
Figure 11: Bern (CH), ERS-1, 14/15 Apr. 1996: Landuse classification based on SAR interferometric signatures.
Figure 12: Bern (CH), ERS-1, 9/10 Jul. 1995: RGB composite of interferometric correlation (red), backscatter intensity (green), and backscatter change (blue).
Figure 13: Bern (CH), ERS-1, 9/10 Jul. 1995: Landuse classification based
on SAR interferometric signatures.
Change Detection and Monitoring
Repeat-pass SAR interferometry is very sensitive to temporal change. To characterize the change the interferometric correlation and the backscatter intensity change are used. The ground resolution and the local incidence angle of corresponding areas in the image pair are identical. Therefore, the backscatter change can be reliably estimated, even without applying elaborate radiometric calibration algorithms that take into account the local pixel size and terrain slope. The backscatter intensity is a function of the geometry and permittivity of the scatterer. Different types of temporal change can be distinguished with repeat-pass SAR interferometry as reported by Wegmüller et al., 1995a, 1995b.
Coherent scatter intensity change (high interferometric correlation together with a significant positive or negative backscatter intensity change) results from an unchanged scatterer geometry in combination with a permittivity change. The permittivity of soil and vegetation is dominated by the high permittivity of liquid water. Coherent backscatter intensity change is observed as a result of changing soil moisture, freezing, and thawing. Decreases of about 3 dB as a result of freezing were observed. Smaller changes occur due to changes in soil moisture and plant water content.
Changing scatterer geometry causes a loss of coherence. Mechanical cultivation (ploughing, furrowing, harvesting) is an example for incoherent change. Quite often a loss of interferometric correlation is observed in combination with a backscatter intensity change. Nevertheless, it also occurs that no backscatter intensity change is observed in spite of geometric changes. In particular, for the case of random dislocation of the individual scatterers in a resolution cell the interferometric correlation is reduced without backscatter intensity change. Such a behavior is typically observed for forest stands and other dense vegetation (sugar beets, corn, potatoes).
If only a part of the scatterers within a resolution cell move or if the displacements are small compared to the radar wavelength partial coherence is maintained. This behavior is observed for canopies with a significant contribution of ground surface scattering. For certain fields a decay of the interferometric correlation due to vegetation growth was observed.
If multiple acquisitions are available over a test-site interferometry allows to monitor change (Wegmüller et al. 1995a. To monitor agricultural fields information is required at intervals of one to two weeks. In this respect ideal data was acquired by ERS-1 during the 3-day repeat-orbit phases. The interferometric correlation of consecutive interferometric pairs allows to monitor change on agricultural fields. An example is shown in Figure 14 for an agricultural area in Middle Zeeland (NL).
The temporal behavior of the fields together with knowledge about local crop calendars allows to map crop types and to detect key processes such as sowing and harvest for the different crop types. As discussed by Wegmüller, 1996, this technique also allows to improve the potential of SAR data for hydrological applications as the interferometric correlation allows to distinguish backscatter changes resulting from geometric change, i.e. changing surface roughness and vegetation cover changes, from permittivity change, i.e. soil moisture change and freezing.
Change occurring between the two data acquisitions of Tandem pairs can be detected as described. Nevertheless, monitoring applications are restricted with the acquisition scheme of the Tandem mission. The 1-day / 35-day acquisition intervals do not allow a continuos monitoring of change on agricultural fields due to the too long 35 day interval. Except for bare and very sparsely vegetated fields the interferometric correlation decreases too much after a 35 day period. In addition the interferometric baselines of the 35 day pairs are very often too large for this type of interferometric analysis. As a result no or very little additional information is obtained from the interferometric correlation.
Of course multi-temporal composites of the interferometric correlation of Tandem pairs may be generated and interpreted, for instance to study seasonal changes in the land cover. An example combining June 1995, November 1995, and April 1996 data is shown for the Bern test site in Figure 15. Low correlation is permanently observed over water and for certain forests. Over agricultural fields and certain forest stands (predominantly deciduous forest) seasonal changes related to changes in the canopy are observed.
Figure 14: Change monitoring on agricultural fields in Middle Zeeland (NL) winter 1994 with ERS-1 repeat-pass SAR interferometry. The multi-temporal image of the interferometric correlation (red: 6./15. Jan., green: 15./27 Jan., blue: 27. Jan./5. Feb.) allows to monitor farming activities.
Figure 15: Multi-temporal image of interferometric correlation over Bern (CH), based on Tandem pairs on 4./5. Jun. 1995 (red), 26./27. Nov. 1995 (green) and 14./15. Apr. 1996 (blue).
GAMMA Remote Sensing's Products and Services
GAMMA Remote Sensing Research and Consulting AG (GAMMA) is a Swiss corporation (Aktiengesellschaft - AG) located near Bern, Switzerland. It was founded in January 1995 by Charles Werner and Urs Wegmüller. The objectives of GAMMA are to conduct research studies and provide consulting services in the field of microwave remote sensing. In the following GAMMA's products and services in SAR and SAR interferometry are shortly introduced. The key personnel of GAMMA have extensive experience in remote sensing techniques, theoretical and empirical modeling, and application development. This experience has been gained during their work at the Universities of Bern and Zürich, Switzerland, and the Jet Propulsion Laboratory, Pasadena, USA.
GAMMA provides licenses for its SAR processing and interferometric processing software. The software is of high quality, portable, efficient (patch processing), user-friendly, and reliable (used at leading institutes). Overall, the design philosophy has been to achieve high quality processing of the data while still permitting processing of the data on a workstation computer in a reasonable amount of time. The software was successfully applied to spaceborne (SEASAT, ERS- 1/2, JERS, SIR-C, RADARSAT) and airborne data.
The SAR Processor is a modular software package to process synthetic aperture radar images from SAR raw data. The design philosophy was to achieve accurate range-Doppler processing of the data (phase conservative, radiometric calibration, well defined geometry, autofocus, motion compensation) while still permitting processing of the data on a workstation computer in a reasonable amount of time. The processor consists of a suite of ANSI-C programs. The ANSI-C language was chosen for its portability and efficiency in processing of large data sets such as full frames of ERS (100km x 100km). The main modules are prefiltering and range compression, autofocus, azimuth compression and multi-looking. For the processing of airborne data a motion compensation module is available.
The interferometric processing software includes the main steps of interferometric processing, i.e. baseline estimation from orbit data, precision registration of interferometric image pairs, interferogram generation (including common spectral band filtering), estimation of interferometric correlation, removal of flat Earth phase trend, adaptive filtering of interferograms, phase unwrapping, precision estimation of interferometric baselines, generation of topographic height, and rectification and interpolation of interferometric height and slope maps. The display of the final and intermediate products is supported with display programs and programs to generate easily portable images in SUN rasterfile format. Processing related parameters and data characteristics are saved as text files that can be displayed using commercial plotting packages.
GAMMA supports customers in the development of their applications. GAMMA conducts SAR processing, interferometric processing, and data analysis for customers, providing end to end support starting with the selection of appropriate data to the interpretation of the resulting SAR images and interferometric products.
SAR interferometric analysis of ERS-1 SAR data pairs acquired during the 3-day repeat-orbits of the commissioning and ice phases have shown a good potential for landuse classification and change monitoring. The main goal of this contribution was to investigate the usefulness of ERS-1/2 Tandem data for these applications. For landuse classification Tandem data has a very good potential. The shorter time interval and the permanently short baseline lead to good results with a much wider applicability of the approach due to the different acquisition mode.
ERS-1/2 Tandem data is less useful for change monitoring. Change occurring during the 1-day acquisition interval of the Tandem pairs can be detected. Nevertheless, continuous monitoring is not possible for most vegetated land surfaces because of a too strong decrease of the interferometric correlation during the 35 day interval to the next data pair. In addition the interferometric baselines of the 35 day pairs are very often too large for this type of interferometric analysis.
The main differences between application of ERS-1 repeat-pass data and ERS-1/2 Tandem data for landuse classification and change detection were summarized in Table 3.
In a quick overview over Gamma Remote Sensing's products and services it was announced that Gamma uses its know-how and software to support customers in their research and application activities in SAR and SAR interferometry.
Table 3: Differences between ERS-1 repeat-pass data and ERS-1/2 Tandem data and consequences for landuse classification, change detection, and change monitoring.
ERS-1/2 SAR raw and slc data provided under ESA-AO2.JRC101 and ESA/Contract 11740/95/NL/PB(SC). This work was supported by ESA ESTEC and the Swiss Federal Office for Education and Science.
Keywords: ESA European Space Agency - Agence spatiale europeenne, observation de la terre, earth observation, satellite remote sensing, teledetection, geophysique, altimetrie, radar, chimique atmospherique, geophysics, altimetry, radar, atmospheric chemistry
|Copyright 2000 - European Space Agency. All rights reserved.|