Assessment of land cover mapping potential in Africa using tandem ERS interferometry
The production of land cover maps from remotely-sensed images has always been perceived as one of the greatest contributions which satellite earth observations could make to both scientific and commercial exploitation of these data.
Since the launch in 1972 of the US ERTS-1 (renamed LANDSAT) programme, scientists have been using optical and near-infrared data to obtain 30m-100m resolution land cover maps. More recently from the LANDSAT-TM and SPOT-XS sensors.
The objectives here are to assess the potential of ERS tandem interferometry to provide an alternative source of land cover information for a semi-arid environment. Existing optical-NIR have poor land cover mapping quality in these environments due to their inability to deal with low Leaf Area Index areas (LAI<2). The clear advantage of this approach is it could permit land cover maps to be produced at anytime of year irrespective of cloud cover which is very high particularly during the rainy season when the Inter-Tropical Convergence Zone is present. It may also be used to validate land cover maps produced using coarse resolution sensors such as AVHRR [Loveland et al., 1991] and ATSR-2 [Higgins, 1995] as well as exploring scaling and generalisation issues [Barnsley et al., 1995]
Results are presented of a quantitative comparison between a land cover map derived using SPOT-XS, LANDSAT-TM and ERS SAR interferometric results. The method used to obtain the SAR interferometric results is briefly described together with an analysis of the results.
Both the SPOT-XS and LANDSAT-TM scenes were acquired on 25 September 1992. Details of these scenes as well as the HAPEX-SAHEL experiment can be obtained from Prince et al., 1995. Unfortunately, due to the limited availability of ERS tandem data from the Libreville, Gabon Receiving Station, only a dry season data-set was acquired which would have much greater amounts of bare earth present (see later). The location of the HAPEX-SAHEL test site is shown below (Figure 1) as well as a close up showing the processed area and the location of the anciliary data (Figure 2).
Figure 1. Overview map of the HAPEX-SAHEL study area using DCW map vectors
Figure 2. Close up of area in Figure 1 showing data sources used and test area
Table 1. Characteristics of the ERS tandem pairs used for land cover mapping.
An area of the full ERS-1,2 scene was selected for processing based upon the availablity of ground truth data from the HAPEX-SAHEL experiment. The commercial EDS-PulSAR system was used to focus the RAW data into an SLC product and further software was used to produce an interferogram and coherence image. In-house UCL software [Muller et al., this proceedings] was used to flatten the interferogram and geocode to a (lat,lon) grid using a 30 arc-second (~1km) DEM together with the precision state orbital vectors from the D-PAF. The tandem pair had a perpendicular baseline seperation (Bperp) of 102m which meant the elevation difference implied by a whole phase cycle was 92m. This resulted in well spaced fringes which appeared easily unwrappable (see Figure 3).
Figure 3. UCL-IfSAR geocoded flattened fringe pattern using
30 arc-second DEM
The phase coherence image (Figure 4) and histogram (Figure 5) computed show that the phase coherence values are high enough for our application.
Figure 4. Phase coherence image for test area (C) UCL/ESA 1996
Figure 5. Phase coherence histogram for test area (C) UCL/ESA 1996
The UCL-IfSAR system for unwrapping is similar to that described by [Goldstein et al., 1988] with the additional use of the coherence image to prevent unwrapping in areas of very low coherence. The residues were first calculated and then opposite-signed residues were linked to produce branch cuts which must not be crossed during unwrapping. Although it is possible to unwrap an interferogram using just a single phase seed point, some 25 seed points were chosen throughout the interferogram. These seeds were chosen by allocating a 'zeroth' fringe and then counting up or down the fringes depending on the grey scale gradient. Having joined the residues and produced a seed point file the algorithm was used to integrate the phase values throughout the interferogram and produce absolute values without the 2PI ambiguity inherent in the interferogram. The unwrapped interferogram was analysed visually in order to determine if any obvious blunders had occurred which would need further seed points to correct. Since there were no obvious blunders the unwrapped phase values were converted to heights using the coarse DEM and the orbital trajectory data and another UCL algorithm (see Figure 6).
Figure 6. UCL-IfSAR hill shaded DEM using 30 arc-second DEM for phase flattening (with overlay box showing test area) (C) UCL/ESA 1996
The ERS-1 and ERS-2 amplitude images were terrain geocoded to the same (lat,lon) coordinates of the coherence image (see Figures 7 and 8). A rectangular section enclosing all available data sources was selected for further analysis (see Figure 6 for location).
Figure 7. ERS1 amplitude image for test area (C) UCL/ESA 1996
Figure 8. ERS2 amplitude image for test area (C) UCL/ESA 1996
An amplitude difference image was then computed and is shown in Figure 9.
Figure 9. Amplitude difference image for test area (C) UCL/ESA 1996
The considerable speckle noise present in these images was considered potentially damaging to classification accuracies. Pratt (1978) suggested the use of a median filter in order to reduce speckle. A 5x5 kernel was selected after investigating the speckle-supression and edge-retention characteristics of a 3x3 and 7x7 kernel. Wegmueller and Werner, 1995 suggest the construction of a false colour composite using the coherence, a single data amplitude and amplitude difference images in order to facilitate visualisation of the thematic information contained within the various SAR products. A schematic diagram of this proposed method is shown in Figure 10.
Figure 10. Wegmueller and Werner proposed false colour composite procedure
Such a composite was produced using the geocoded SAR products described above (see Figure 11).
Figure 11. False colour composite using colour scheme shown in Figure 10. (C) UCL/ESA 1996
The false colour composite (Figure 11) showed several distinct classes and presented good evidence on its own of the land cover capabilities of IfSAR. A 50-class ISODATA unsupervised classification was then performed using ERDAS-IMAGINE on the dataset in order to enhance the class information present in the image in line with the classification procedure shown below in Figure 12.
Figure 12. Overview of classification procedure
Figure 13. Unsupervised classification of SAR datasets (C) UCL/ESA 1996
The result of this classification helped in the selection of 4 training areas for a maximum likelihood supervised classification shown in Figure 14.
Figure 14. Supervised land cover classification map of SAR with colour key (C) UCL/ESA 1996
A land cover map produced in 1992 using SPOT-XS data acquired on September 25th 1992 as part of the HAPEX-SAHEL experiment [Prince et al., 1994] was downloaded from the HSIS www site (http://www.orstom.fr/hapex/index.htm) (acknowledgement J.M d'Herbes, C. Valentin, B. Mougenot). The map which was derived using supervised classification was originally in 6 classes as shown in Table 2 below.
Table 2. Original classes of the HSIS map
The 6-class map was reduced to 4 classes which correspond to the IGBP classification scheme [Belward and Loveland, 1995]. Classes 3, 4 and 5 were merged, producing the final classification scheme shown in Table 3 and in Figure 15.
Table 3. HSIS classes after merging to 4-class IGBP scheme
Figure 15. Supervised land cover classification map of HSIS with colour key
The map was registered to the radar imagery using a cubic transformation with some 10 ground control points. A section of a 7-band Thematic Mapper scene was available and this was registered to the same (lat,lon) coordinates as the IfSAR products in the same way as the HSIS map. A 50-class ISODATA unsupervised classification was performed on the 6 non-thermal bands in order to help select 4 training areas for a maximum likelihood supervised classification (Figure 16).
Figure 16. Supervised land cover classification map of LANDSAT-TM with colour key
The footprints of the three data sources were not identical which necessitated the selection of a rectangular area enclosing all three sources for further analysis. The HSIS map was treated as a reference source and the common areas in the radar and TM classifications were extracted.
Congalton, (1991) recommended random stratified sampling in order to fairly assess the accuracy of classifications based on remotely sensed data. 40,000 random stratified samples ( representing approx. 5% of the pixels present in the area selected for quality assessment) were selected from the reference dataset. Error matrices were then produced using the radar and TM classifications (see next section). The correlation between the radar and TM classifications was also directly assessed. The construction of error matrices allowed a simple statistical analysis of the classification accuracies to take place.
WS= woody savannas, BA= barren, OS=open shrublands, WB=water bodies.
Table 4 below shows the error matrix computed for the SAR classification using HSIS as the reference source. HSIS classes are along the top, SAR classes are down the side.
Table 4. Error matrix for radar classification using HSIS as reference
PRODUCER'S ACCURACIES: 68%,22%,78%,67%
The producer's accuracies show the proportion of reference pixels correctly identified by the radar classification. The user's accuracies show how many pixels labelled a particular class by the radar classification actually are that class in the reference data. The two measures together are extremely useful as they give the commision and omission errors. For example, a map could be produced where every single pixel was labelled as `barren'. A simple test of the producer's accuracy for `barren' would give a figure of 100% - no omissions. However, the user's accuracy would reveal the true nature of the classification by showing the huge commision errors.
The figures for the radar classification are reasonable with the exception of `barren' - with producer's and user's accuracies of 22% and 53% respectively. The overall accuracy of 68% is encouraging however despite being weighted considerably by the good results for the large `open shrublands' class. The HSIS map was produced using imagery acquired during September, when vegetation is at a maximum. The radar imagery was acquired at the height of the dry season and therefore the classification results are expected to differ considerably.
Table 5. Error matrix for TM classification using HSIS as reference
Producer's accuracies: 72%,33%,68%,72%
Table 5 above shows that the barren class is again very badly represented in both omission and comission errors. The user's accuracies for `open shrublands' and `water bodies' are extremely high indicating that these classes had a very distinct profile in the TM imagery.
Table 6 below shows the error matrix computed for the SAR classification
using TM as the reference source.
Table 6. Error matrix for radar classification using TM as reference data
Producer's accuracies: 59%, 13%, 83%, 87%
This error analysis indicates that there is virtually no correlation between the areas marked as barren in the TM and radar imagery. The other classes are well correlated and the overall accuracy is fair.
Discussion and conclusions
The land cover maps produced using SAR interferometry show high correlation with both the LANDSAT-TM and SPOT-XS results for the vegetated classes. The differences may partly be due to the different times of year (this is particularly true for the barren class which are much higher for the SAR than the other two satellite-produced maps) and partly due to the fact that the SAR senses the volume scattering from within the vegetation canopy [Gatelli et al., 1994] as opposed to the response of the vegetation at different spectral wavelengths at optical/NIR wavelengths.
SAR interferometry has been shown to be an extremely effective method, from a single date of obtaining land cover information in a semi-arid environment. A quantitative assessment was made of SAR-derived land cover using the IGBP classification scheme using both SPOT-XS and LANDSAT-TM as "ground" truth. For vegetated and water classes the correlation was high whereas for barren the SAR consistently shows a greater coverage.
Further studies are underway to assess the land cover potential of IfSAR over other areas in Africa including wetlands (Morley & Muller, this proceedings), savannah and tropical forested areas as well as within Europe. Reports on these experiments will be made in upcoming Florence meeting.
Ben Mumford acknowledges receipt of a NERC Advanced Course studentship (GT3/95/140/E0) for a Master of Science programme in Remote Sensing. This work was supported by BNSC under the Application Development Programme with Logica Space Systems and by NERC under the framework of the TIGER-SVATS project. Finally, ERS raw data was supplied under the ESA data grant (AOL.UK201).
Barnsley, M.J., S.L. Barr, and T. Tsang. (1995). Producing large-area land cover maps from satellite sensor images: scaling issues and generalisation techniques., in Environmental Remote Sensing from regional to global scales, edited by G. Foody, and P. Curran, John Wiley & Sons.
Belward, A. and Loveland, T. (1995). The IGBP-DIS 1Km Land Cover Project. Proc. 21st annual conference of the remote sensing society, Southampton, UK. Remote sensing society, Nottingham. 1099-1106.
Congalton, R. (1991). A review of assessing the accuracy of classifications of remotely sensed data. Rem. Sens. Env., Vol.37. 35-46.
Gatelli, F., A.M. Guarnieri, F. Parizzi, P. Pasquali, C. Prati, and F. Rocca. (1994). The wavenumber shift in SAR interferometry., IEEE Trans. Geosci. Rem. Sensing, 32, 855-865.
Goldstein, R. M., Zebker, H. A. & Werner, C. L. (1988). Satellite radar interferometry: Two dimensional phase unwrapping". Radio Science, Vol.23, No.4, 713-720.
Higgins, N.A. (1995). Potential of the ATSR-2 on ERS-2 for monitoring land cover change., in 21st Annual Conference of the UK Remote Sensing Society, edited by P.J. Curran, and Y.C. Roberston, pp. 320-324, Remote Sensing Society, University of Southampton, 11-14 September.
Loveland, T.R., J.W. Merchant, D.O. Ohlen, and J.F. Brown (1991). Development of a Land Cover Characteristics Database for the Conterminous US, Photogrammetric Engineering and Remote Sensing, 57 (11), 1453-1463.
Pratt, W. K. (1978). Digital Image Processing. Wiley, New York.
Prince, S. D., Kerr, Y. H., Goutorbe, J.-P., Lebel, T., Tinga, A., Bessemoulin, P., Brouwer, J., Dolman, A. J., Engman, E. T., Gash, J. H. C., Hoepffner, M., Kabat, P., Monteny, B., Said, F., Sellers, P., & Wallace, J. (1995). Geographical, biological and remote sensing aspects of the Hydrologic Atmospheric Pilot Experiment in the Sahel (HAPEX-SAHEL). Rem. Sens. Env., 51(1), 215-234.
Wegmueller, U. and Werner, C. L. (1995). Land surface analysis using ERS-1 SAR interferometry. ESA Bull., No.81, Feb. 1995. 30-37
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.|