Images 8, 9 and 10 show respectively: those pixels corresponding to forest and classified as forest, those pixels classified as non-forest but corresponding forest areas, and finally those pixels classified as forest but corresponding to non-forest areas.

Fig. 8 Forest classified as forest	Fig. 9 Forest classified as non-forest
Fig. 10 Non-Forest classified as forest

As it can be observed, the major error occur for areas indicated as forest in the ground thruth map and not detected as forest in the classification process. In principle, there may be several reasons for this missclassification:

• a change in the surface (e.g. deforestation) occurred between the field work carried out for generating the land use map and the acquisition of the ERS data here used.;
• a missclassifcation in the ground thruth map;
• a bad tunning of the classification algorithm parameters (see fig.4);
• particular characteristics of the non-detected forest which make it appear in intensity and/or in coherence as non-forest areas

In order to investigate the reason for this particular miss-classification, the coherence and intensity mean histograms for the pixels indicated as forest in the ground truth map and classified as non-forest (i.e. pixels in fig. 9) were generated. They are provided in figures 11 and 12.

Fig. 11 Histogram of coherence for forest areas classified as non-forest



Fig. 12 Histogram of intensity mean for forest areas classified as non-forest
The above histograms show that most of the pixels indicated as forest in the land use map which have not been classified as such, present intensity values within the expectable dynamic range but coherence values much higher than what would be expected over forest areas. In other words, this means that these areas have probably been deforested between the generation of the land use map and the ERS acquisitions.

What has been here presented are the preliminary evaluation results, derived from the land use map of Zhengcheng County in 1990. However, in order to carry out a more precise evaluation of the obtained forest-non forest map, Landsat images and specific field work will be exploited and a complete evaluation based on them will be carried out in the future.

Conclusion
Several conclusions can be directly derived from the results presented in the previous section:

1. There is a considerably good agreement between the ground truth forest map and the obtained forest/non-forest map (75 % of accuracy).
2. The methodology used for the classification is highly based in the coherence behaviour of forest and non-forest surfaces, which is essential for reaching the above accuracy.
3. There is certain umbiguity between water and forest areas which is difficult to solve from the information used in this exercise (intensity, coherence and intensity change). The traditional way to solve this ambiguity relays in the higher change in intensity occurred over water bodies. However, in this particular case, water bodies within the scene are relatively calm in both acquition dates, and therefore the exploitation of the intensity change allows only a partial removal of the ambiguity between forest and water. Although there could be auxiliary ways of eliminating this umbiguity (e.g. using texture information), no major effort has been addressed to this matter since The Chinese Academy of Forestry has a map of water bodies that can be easily used to solve any umbiguity between water and forest.
4. The parameters used in the classification process have been obtained from a-priori analysis of different surface samples within this test site but it may be necessary to modify their values over a different area.
5. The results over Zengcheng County are indicative of the accuracy which is possible to achieve using ERS SAR data and interferometric techniques for forest/non-forest mapping in China. However, this accuracy depends mainly on the capability of coherence and intensity for distinguishing the radar reponse received from forest and from non-forest surfaces. Since several factors such as forest type and surface topography have a great influence on this capability (i.e. topography biases the radar intensity and coherence, and in addition increases layover areas, where clasification is difficult), it is expectable that the achievable accuracy will vary over different areas.

Acknowledegment
This work was performed within a joint project "Forest Mapping in China with ERS SAR Tandem data" between the Chinese Academy of Forestry and ESA ESRIN. Many thanks to ESA ESRIN for providing the ILU images used in this work.

References

• Askne, J., Dammert P., & Fransson, J., etal.,1995, Retrieval of forest parameters using intensity and repeat-pass interferometric SAR information, Proceeding of the International Symposium on Retrieval of bio- and geophysical parameters from SAR data for land application, Toulouse, France, 10-13 october 1995, pp. 119~129
• Urs vegmüller, Charles L. Werner, 1995, SAR Interferometric signatures of forest, IEEE Trans. Geosci. Remote Sensing. Vol., 33, no. 5, pp. 1153~1161
• Urs vegmüller, C. L. Werner, D. Nüesch, and M. Borgeaud, 1995, Land-surface analysis using ERS-1 SAR interferometry, ESA Bulletin, No. 81, pp. 30-37.
• Urs vegmüller, Charles L. Werner, 1996, Land Applications using ERS-1/2 Tandem data, Frange 96.
• Andr Beaudoin, Thierry Babaute, 1996, Forest monitoring over hilly terrain using ERS INSAR data, Frange 96.