Research on the capabilities of ERS SAR for monitoring of land use changes in the Neotropics

Abstract

Land use maps at a scale of 1:200.000 exist for the whole of Costa Rica. They were produced in 1985 from Landsat-TM images and aerial photographs. Given the frequent cloud cover, regular updating of those maps via optical satellite data is problematic.
The north-east of Costa Rica is a region with a nearly flat relief and some isolated hills. Most of the land is covered with pasture and wet tropical forest. During the last decade a significant portion of the area has been converted into plantations, mostly of banana.
A procedure has been developed for producing land cover maps and an associated measure of accuracy. It consists of calibration, resampling, filtering, segmentation, principal component transformation and supervised classification of multitemporal radar images. As input to this processing chain any number of SAR.PRI images or derived texture images can be used. The output consists of a land cover map and an accuracy assessment. The procedure has been applied to a series of 4 SAR images, which yielded a map with an overall accuracy of 78 %. The high precision with which the large banana plantations can be mapped is most interesting, both for its economics importance as for environmental monitoring.
By combining this map with the land cover map of 1986 it is possible to detect the changes over the period 1986-1995.

Keywords: land cover change, banana plantation, ERS-SAR

Introduction

Land use maps are important tools for regional land use planning. Not only should these maps contain the correct thematic information, but this information should also be up-to-date.

In tropical areas the updating of the land use maps by use of optical satellite imagery (SPOT, Landsat,...), is often problematical because of the frequent cloud cover. Quite often it will take years in order to obtain a reasonably cloud free image, and in the mean while the value of the extant maps may decrease to a point where they are almost useless.

Since the launch of ERS-1 and ERS-2 by ESA, SAR-images are available that are weather independent and which cover the whole surface of the globe on a regular basis. So it is quite understandable that the question has been raised as to what extent these images can be used to produce new land cover maps, and with what accuracy this can be done.

These are the questions to which this article will try to find an answer.

Material and Methodology

Study area

The study area is situated approximately between 9° 30' N and 11° N, 83° W and 84° W (figure 1) in the Atlantic Zone of Costa Rica (Central America) and covers some 425 000 hectares. The area is virtually flat, with a few isolated hilltops reaching 170 meters. Most soils are well drained, with the exception of a more than ten kilometre wide zone along the Caribbean coast. The southern half is rather densely populated, with Puerto Viejo de Sarapiquí, Guápiles, Guácimo and Siquirres as main population centres.

Figure 1 : Study area

Until a few decades ago the area was completely covered in wet tropical forest. The vegetation in the coastal swamps is characterised by the presence of the yolillo-palm (Raphia taedigera) (Janzen, 1983). Recent colonisation of the area has brought important changes to the original vegetation. Nowadays land use in the southern part is dominated by cattle ranches. Less important activities include the cultivation of ornamental plants, palm heart (Bactris gasipaes), plantations of Gmelina arborea and Tectona grandis (Stoorvogel and Eppink, 1995).

During the last decade banana plantations have expanded dramatically, and the Atlantic zone has become a major production area. The banana production plays a very important role in Costa Rica's economy. It is the country's leading source of foreign exchange (Science in the Rainforest), with the total income for 1993 estimated at USD 560 million. The banana industry also is an important employer. An estimated 43 000 Costa Ricans are directly dependent on banana production for their livelihood, not even including part-time-employment or the domestic service industries that support the banana producers. But there is a less bright side to the picture too. The banana production is dependent on large volumes of pesticides, herbicides and fertilisers and much of this is washed away or dissipates before the plants can benefit from it. An estimated 90 % of fungicides are lost in this way. Another problem is posed by the solid waste generated during the banana production. For each ton of banana exported, three tons of waste is produced, mostly consisting of leaves and stems of the banana plants (Hernandez and Scott, 1996). Furthermore, streams are diverted and channelled, and contaminated with sediment and chemicals. It should be clear that monitoring of this activity is of interest to governmental as well as non-governmental institutions.

Available data

The available data consist of four ERS-1 images, dating from November 9 1992, August 16 1993, September 20 1993 and October 25 1993, and one cloud-free Landsat-TM image dating from February 6 1986. Collateral data consist of aerial photographs at scale 1:60 000 dating from 1992 and 1993 and more than 1200 field observations of land use for 1995.

Digital processing of the radar image

A suite of tools has been designed for the processing of ERS.SAR.PRI images and has been described in Verhoeye and De Roover (1996). One or more PRI-images, possibly together with synthetic channels such as texture images, serve as input. The output consists of a classified image and a measure of accuracy.

As a rule the chain has been implemented using commercially available software (Earth View by Atlantis Scientific Systems Group Inc., ISI-2 by Intergraph Corporation). Only when these are inadequate additional software programs has been developed. This is the case for the calibration, segmentation and evaluation modules.

First the images are read from tape onto the computer hard disk, after which the header files of the PRI-images are analysed in order to extract the parameters needed for the calibration. A program for calibration of ERS-images has been developed using the formulas proposed by ESA (Laur, 1992). Subsequently the calibrated images are imported into the image processing software. At this point a first data volume reduction is obtained by linearly rescaling the 16-bit data into 8-bit data.

After georeferencing, the images, originally measuring approximately 8000 by 8000 pixels, are reduced to 2000 by 2000 pixels by applying an averaging procedure. While at the one hand allowing faster processing by further reducing the data volume, the averaging also removes some of the speckle. The resolution of resultant images is 60 meters.

In an effort to further reduce the speckle, the images are subjected to filtering. Next a principal components transformation (PCT) is applied to the filtered images. The analysis yields Eigenvalues, which will be used as weighting factors during the segmentation.

Despite the fact that the application of filters can eliminate a considerable amount of speckle, it is accepted that the accuracy can be further improved if the radar image is segmented prior to the classification (ESA, 1995). Segmentation techniques partition the image into multi-pixel regions that represent discrete objects or regions in the image. The method applied here is based on the "Multiresolution Pixel Linking" algorithm, which uses a "pyramid" of images at successively lower resolutions. It establishes links between pixels at successive levels of the pyramid and uses these links to move the values of the top level down to the base level, resulting in a segmented image (Burt, 1984; Hong, 1982; Rosenfeld, 1984). The segmentation algorithm has been discussed in more detail in Verhoeye and De Roover (1996).

The segmented images are then used as input to a supervised classification, applying the Maximum Likelihood algorithm and using the field observations of 1995 as training data. In order to be able to assess the accuracy of the classification result, it has to be confronted with the ground truth, which will be provided by the results of the interpretation of the 1993 aerial photographs. Different measures of classification accuracy can be calculated after constructing a confusion matrix (Lillesand and Kiefer, 1994).

Results

Land cover map for 1993

An exhaustive evaluation of the processing chain revealed that the optimum classification result was obtained using the following settings :

• multitemporal input (four images)
• inclusion of textural information as a fifth channel
• sigma filtering using a window measuring seven by seven pixels
• principal components transformation
• segmentation using a pyramid consisting of seven levels, root level at level one, two root intervals
• classification using the maximum likelihood algorithm, six land cover classes
• no modus filtering.

Ground truth is provided by stereoscopic interpretation of aerial photographs at scale 1:60.000. The total area for which ground truth has been interpreted covers some 70.000 hectares. As can be seen in table 1, no ground truth data are available for the yolillo class. This is due to the fact that the aerial photographs taken over these extensive swamp forests contain no features (like cross-roads, bridges,...) that can be referenced to the topographical maps.

The classification result can be seen in figure 2. The patches "not classified" within the study area correspond to isolated hills that have been masked out because the effects of layover and radiometric distortion do not allow a correct classification. The general picture is largely consistent with the land cover as interpreted from the aerial photographs and the field observations : yolillo swamps along the Caribbean coast, forests in the northern half of the study area, extensive grasslands and large scale banana plantations in the south.

Figure 2 : Land cover 1993

A more detailed picture can be obtained by studying the confusion matrix (table 1), the producer's accuracies (table 2) and the consumer's accuracies (table 3). Overall accuracy is 78 %, which falls within the range (65 % to 80 %) cited by ESA (1995) for land cover classification in Thailand. Striking is the misclassification of villages. This should come as no surprise as houses in the study area are widely scattered along the roads and their surroundings are very heterogeneous, typically containing a lot of trees and small fields. Considering the spatial resolution of the images used (60 meters), pixels inevitably become mixed.

Quite a lot of confusion exists between forest and grassland. The largest error is that of commission, with 40 % of the forest land cover actually belonging to the grassland class. This is probably related to the ranching practices applied in the study area : even intensively used grassland contain a fair amount of remaining tree groups and stems. More extensively used pastures tend to be invaded rapidly by scrubs and bushes eventually leading to a dense and entangled secondary vegetation. This creates a rough surface and will lead to higher values of backscatter which apparently are confused with those of forest. Similar observations have been made in the Brazilian rain forest (Keil et al., 1996) and on Borneo (Kuntz et al., 1996).

On the other hand, the accuracies for classifying grassland and banana are fairly high : both producer's and consumer's accuracy range between 83 % and 87 %. Especially the monitoring of banana plantations could prove to be an important operational application, as these plantations represent important economic values as well as pose severe environmental threats (Science in the Rainforest; Hernandez and Scott, 1996). When the confusion matrix is evaluated after regrouping into two classes (banana versus non-banana) it becomes clear that the accuracy is extremely high (95 %) (table 4). Another indication of the reliability of ERS-images for monitoring banana plantations can be found when calculating the total surface of these plantations (44 600 hectares). This compares very well to the surface (44 187 hectares) calculated from the 1993 aerial photographs at scale 1 : 60 000 for the same area (Stoorvogel and Eppink, 1995).

The results discussed till now are produced by multitemporal input, meaning that images dating from several (in this case four) dates are processed simultaneously. It would be interesting to know how these compare to the results from monotemporal input. From table 4 it can be seen that the accuracy for discriminating banana from other land cover is equally high in monotemporal as in multitemporal processing!

Change detection for 1986-1993

The classification of the 1986 Landsat-TM image yields a land cover map with 90 % accuracy. By applying the post-classification comparison technique, it is possible to detect the changes in land cover that have occurred between 1986 and 1996.

Figure 3 : Land cover change 1986-1993

The expansion of the banana plantations is very obvious (from 4% to 11%). While there is only a small increase of grassland, the rate of forest loss is significant : 2% of the forested area disappears annually (figure 3). These figures would seem to confirm the opinion (Science in the Rainforest) that the expansion of the banana plantations occurs at the expense of the rain forest. A more detailed interpretation of the land cover change maps leads to a somewhat different conclusion : it appears that that the majority of the new banana plantations are established in former grasslands in the southern half of the study area, while new pastures are being created at the fringes of the forests in the north and the south-east.

Conclusions

It has been shown that land cover maps with a fair degree of accuracy can be produced using ERS.SAR.PRI as input data. This can be achieved by applying a series of digital operations, most of which can be done using commercially available software. A notable exception is the segmentation, for which a software module has been developed in house. Although the overall accuracy of the land cover map may not be as high as can be achieved using optical images, the accuracy with which banana plantations can be detected is very high. This holds true both for multi- and monotemporal datasets. These findings point to a potentially important operational application.

Acknowledgements

Funding for this project was provided by the Belgian Federal Office for Scientific, Technical and Cultural Affairs (OSTC). The ERS-1.SAR.PRI images were made available by ESA.

References

Burt, P.J., 1984:

The pyramid as a structure for efficient computation. In: Rosenfeld, A., Multiresolution Image Processing and Analysis. Springer-Verlag, Berlin Heidelberg New York Tokyo: pp 8-35.

ESA, 1995:

Satellite Radar in Agriculture. Experience with ERS-1. ESA SP-1185. ESA Publications Division. ESTEC, Noordwijk, The Netherlands. ISBN 92-9092-339-3.

Hernandez, C.E., Scott, G.W., 1996:

Evaluating and Managing the Environmental Impact of Banana Production in Costa Rica : A Systems Approach. Ambio, Vol. 25, May 1996 : pp. 171 - 178.

Hong, T.H., 1982:

Image smoothing and segmentation by multiresolution pixel linking : Further experiments and extensions. IEEE Transactions on Systems, Man, Cybernetics, pp. 611 - 622.

Janzen, D. H., 1983:

Costa Rican natural history. The University of Chicago Press. ISBN 0-226-39334-8.

Keil, M., Scales D., Winter R., Kux H., dos Santos J.R., 1996:

Tropical rainforest investigation in Brazil using multitemporal ERS-1 SAR data. Proceedings of the Second ERS Applications Workshop, London, UK, 6-8 December 1995 (ESA SP-383, February 1996) : pp. 77 - 82. ISBN 92-9092-215-X.

Kuntz, S., Streck C., Siegert F., 1996:

Multitemporal Evaluation of ERS-SAR Data for Monitoring Deforestation Tropical Forest. Proceedings of the Second ERS Applications Workshop, London, UK, 6-8 December 1995 (ESA SP-383, February 1996) : pp. 83 - 94. ISBN 92-9092-215-X.

Laur, H., 1992:

Derivation of backscattering 0 in ERS-1.SAR.PRI products. ESA/ESRIN.

Lillesand, T.M., Kiefer R.W., 1994:

Remote sensing and image interpretation. Third Edition. John Wiley & Sons, Inc. New York - Chichester - Brisbane - Toronto - Singapore. ISBN 0-471-57783-9.

Rosenfeld, A., 1984:

Some useful properties of pyramids. In: Rosenfeld, A., Multiresolution Image Processing and Analysis. Springer-Verlag, Berlin Heidelberg New York Tokyo: pp. 8 - 35.

Science in the Rainforest :

"Smart Banana" Program at http://www.pbs.org/tal/costa_rica/res2/costbanana.html

Stoorvogel, J., G.P. Eppink, 1995:

Atlas de la Zona Atlántica Norte de Costa Rica. Programa Zona Atlántica (CATIE-UAW-MAG). Guápiles, Costa Rica.

Verhoeye, J., B. De Roover, 1996:

Research on the Capabilities of ERS-SAR for Monitoring of Land Use Changes in the Neotropics : Preliminary Results. Proceedings of the Second ERS Applications Workshop, London, UK, 6-8 December 1995 (ESA SP-383, February 1996): pp. 463 - 468.

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