ESA Earth Home Missions Data Products Resources Applications
    22-Jul-2014
EO Data Access
How to Apply
How to Access
FRINGE '96 Workshop: ERS SAR Interferometry, 30 September - 2 October 1996
Services
Site Map
Frequently asked questions
Glossary
Credits
Terms of use
Contact us
Search


 
 
 

FRINGE 96

Forest INSAR decorrelation and classification properties

J. Askne Chalmers University of Technology,
Dep. of Radio and Space Science,
S-41296, Göteborg, Sweden
askne@rss.chalmers.se
http://www.rss.chalmers.se
G. Smith Chalmers University of Technology
Dep. of Radio and Space Science
S-41296, Göteborg, Sweden
smith@rss.chalmers.se
http://www.rss.chalmers.se

Abstract

Large areas are covered by forest and the decorrelation properties of the forest are important to understand in order to use INSAR for deriving forest properties and determine the potential for DEM production. From the 3-day repeat cycle of ERS-1 in 1994 we have determined interferograms of forested areas in northern Sweden. Variations of coherence with time and stem volume have been determined. Discrimination between forest and clear cuts has been characterised. A model for the coherence variation is discussed and critical parameters for further studies identified.
Keywords:. Coherence analysis, forestry, land classification

Introduction

Estimation of forest biomass is important on a global scale as an input parameter to environmental models and on a local scale for assessment of forest properties. It is of great economical value for forestry companies to improve forest inventory methods.

A large number of observations of boreal forest properties have been performed based on the radar backscatter, e.g. using the ERS-1 satellite C-band SAR. A conclusion is that the radar backscatter at C-band saturates for relatively small values of the stem volume or biomass, see e.g. (Israelsson and Askne, 1994). However, the information content from the ERS satellites is increased by using interferometric repeat pass INSAR observations. From these we can derive the interferometric effective height of the forest layer and the coherence, see e.g. (Dammert et al., 1995; Hagberg et al., 1995). These properties have been found to vary with forest parameters and there is a need for analysis of the potential of interferometric measurements and to understand the basic phenomena related to the interferometric imaging of forest.

The most significant aspect of INSAR observations of forest areas is the typically low coherence. This complicates the analysis as the phase variations are then hard to estimate. But the low coherence of forests also makes it possible to use coherence to discriminate forest and non-forest, and we will also study some clear cuts and open fields.

In this presentation we will report on ERS-1 INSAR observations and extend the previous model approaches to understand the signal dependence on different parameters (Askne et al., 1995; Hagberg et al., 1995; Askne et al., to be publ.). An aim is also to identify those properties we need to investigate further in order to improve our modelling of the INSAR response of vegetation.

Observations

Field site

The test areas are centred around Hökmark (lat. 64 25', long. 21 15') in northern Sweden characterised by boreal forest. The most dominant species are conifers (Norway spruce and Scotch pine). A typical value of the stem volume in this part of Sweden is 100 m3/ha. An illustration of the forest at one of the test sites is given in Figure 1, and for a fairly typical clear cut with regrowth in Figure 2.

Figure 1. Illustrating an area with a stem volume estimated to 240 m3/ha. Note the gaps in the vegetation canopy.

Figure 2. Illustrating a clear-cut in the area.

Data set

Repeat-pass interferometry is dependent on the time interval between the acquisitions as temporal changes decrease the coherence. For this reason we have concentrated on observations during the 3-day repeat pass period of ERS-1 in February - March 1994. We have studied 55 interferogram pairs based on 11 SAR images. For 13 interferograms the baseline was too long, while 27 were characterised by too low coherence. Of the 15 remaining interferograms we have here concentrated on four interferograms with coherence values for the forests above the coherence bias (see below). These interferograms are based on six SAR images

INSAR measurements

In Figure 3 we have illustrated for five forest areas and six clear cuts of varying age. The values have been determined from ERS-1 SLC images and are estimated to have an accuracy of about 0.5 dB. The variation in illustrated by Figure 3 are partly related to temperatures changing from above zero to -15C. These variations are larger for the forest due to the dielectric constant changes than for the fields which are assumed to be frozen for the entire observation period.

Figure 3. Illustration of for five forest areas and six clear cuts of varying age (note stem volumes for clear cuts are preliminary estimates).

Coherence is defined by

g1 and g2 denote pixel values in each of the two images respectively and E{} denotes expectation value.

Observations of the coherence as function of stem volume is illustrated in Figure 4. The accuracy of the coherence measurements is determined by the averaging area, in our case 104 m2 = 1 ha. Only values above 0.3 have been illustrated as the coherence estimator is biased for lower values. As coherence is dependent on the baseline, Bn, we have divided the figure for two cases with Bn = 21 and 30 m and Bn = 175 and 203 m

Figure 4. Illustrating coherence variations with stem volume, V. a) small baselines, b) medium baselines.

An interferometric phase shift relative surrounding open fields can be observed over forested areas. By comparison with the open field phase values or a DEM over the forested area an interferometric forest height can be determined. Results have been reported in (Askne et al., to be publ.). This information is as important as the coherence for classification of the forest properties. However, in this presentation we will concentrate on the variation of coherence as function of stem volume.

Interpretation of observations

INSAR system model

The interferometric SAR observations are dependent on system parameters (e.g. system noise and baseline), on processing steps (e.g. resampling), and medium (forest) properties. Assuming the processing of the images to affect the coherence negligible we obtain (Ulander and Hagberg, 1995)

For reasonably high signal levels the noise part is also negligible. The spatial part includes decorrelation due to the surface and volume scattering. Of these the surface related part, the baseline decorrelation, can be corrected for by the technique in (Gatelli et al., 1994). All coherence values reported here have been corrected for baseline with local slopes determined by FFT analysis. The accuracy of the coherence measure is dependent on the resampling procedure, on the estimator properties, on correction for local topography etc. Further aspects are reported by Dammert in this conference. We will here concentrate on the variation of the volume and temporal decorrelation and investigate the coherence variations with stem volume.

INSAR scattering model

Models to determine the backscatter from forested areas have reached a maturity with the introduction of programs like MIMICS, (Ulaby et al., 1990a), and other similar approaches. Such models are useful for the forward model, but due to the number of included parameters the inverse model has to be based on observations from multi frequency, multi polarisation measurements. Including phase dependent interactions necessitates still further complex modelling and to obtain a more conceptual feeling for the influences of various factors we will return to a more basic and simple description of the radar backscatter model, the so called water cloud model, (Attema and Ulaby, 1978).Then the forest layer is described simply by a homogenous layer of scattering particles. The basic parameters in the model are the radar cross section and total attenuation cross section of each scattering "particle" in combination with the ground parameters. N is the number of particles per unit volume (here assumed proportional to the stem volume V).

However, in particular for sparse forests like boreal forests, radar scattering is obtained not only due to the radar wave penetrating the forest layer, but also through gaps in the vegetation (McDonald and Ulaby, 193). The water cloud model has been generalised to take this into account (Askne et al., 1995; Askne et al., to be publ.).

The decorrelation due to changes in the scattering is the most complex aspect of repeat pass interferometry. Aspects of interferometric decorrelation have been described in e.g. (Zebker and Villasenor, 1992). Very little is known about actual mechanisms. In the forest case we anticipate there are decorrelation phenomena on various time scales. On the short time scale we have the effect of wind moving the scatterers in a random manner, in a relatively short time scale we can have dielectric constant changes e.g. due to rain and temperature transitions around the ice/water transition point. On the longer time scales we have e.g. the growth of the vegetation, man made changes, storm damages and fires. The temporal decorrelation affects the ground surface and the forest layer differently. For some of the effects meteorological information can help in the interpretation, and this is important for the future development of the technique.

Our model assumptions are suffering from lack of observations or accurate observations of some parameter values. The aim of the model is therefore not to make a detailed analysis in order to support the measurements, but to test what properties are important for understanding the measurements by comparing the model with measurements. By including a large number of parameters in a model observations can be adjusted to fit any observations. Instead we try to use as few as possible and limit these to such which are reasonably easy to estimate and observe independently. This also means that each parameter is a complex function of many important phenomena in a more detailed future model.

Forest parameters

For simplicity we will describe the forest by means of the stem volume. This is a parameter which can with high correlation be related to many other basic properties such as biomass, basal area, height, and age. In particular the relation to the forest height is important in an interferometric model.

In the water cloud model the forest properties are described in a statistical manner as a continuous layer of properties. This is in contrast to the clumping of properties of an actual forest. In our case we have included an area fill parameter to characterise the leakage of electromagnetic energy through small gaps in the canopy layer down to the ground level. The area fill factor is estimated in a few cases from site inspections and photographs, see Figure 1. We have also used results reported by (Pulliainen et al., 1994) where their transmissivity is interpreted as a combination of attenuation through the vegetation layer and the leakage through vegetation gaps. The area fill factor can be assumed to be climatologically dependent and smaller in northern parts of Sweden than in southern parts.

Decorrelation properties

For short baselines the temporal decorrelation dominates over volume decorrelation. We find forest coherence values between 0.30 and 0.45 over time intervals up to nine days. For middle range baselines (170 - 300 m) the coherence values are of the same order and we have found coherence values above 0.3 for time intervals up to 15 days. This make us believe that the temporal decorrelation dominates over volume decorrelation.

Of the various decorrelation phenomena we believe the wind decorrelation is most important, acting on time scales of minutes and shorter. The important aspect is the stability of those scatterers causing the major part of the coherent backscatter, which is believed to be branches in the upper part of the canopy. As the wind is affecting the top of the trees most in a forest, and the lower part of the stems very little, we model the decorrelation as a function increasing from the top downwards.

Long term changes have not been dealt with specifically in the model. However observations show that the coherence change with time and the backscatter difference between the image pairs increase with time. Such changes may be related to local storms combined with precipitation and also to temperature changes. These effects may also cause irreversible changes resulting in very low coherence in image pairs acquired on each side of such an occasion.

Attenuation through vegetation layer

To measure the attenuation of a vegetation layer is very complex due to the variability of the vegetation and due to speckle effect. Very few observations are known to the authors and we estimated the attenuation based on results from (Ulaby et al., 1990b; Fleischman et al., 1992; Seifert et al., 1995) and some MIMICS simulations.

Model properties

By testing the sensitivity of the model results to variations of the various parameters it is possible to estimate the importance of each parameter and then the importance of increasing our knowledge of each of these "effective" parameters. Model assumptions are illustrated in Figure 5, and will be further discussed in future papers, see also (Askne et al., to be publ.). Each one of the illustrated parameters needs to be further observed in the future.

 

Figure 5 (a) area fill factor; (b) two way attenuation through the homogeneous vegetation layer, dotted line, and corrected for the denser vegetation by the area fill factor, solid line, x marks estimates from literature; (c) assumed temporal decorrelation factor due to wind (dash line), attenuation in a forest with a stem volume of 230 m3/ha (dotted line), and the resulting effective layer (the product of the two previous factors) giving rise to the coherent scattering (solid line).

Results of model calculations and comparison with observations

Some results of the model calculations are given in Figure 6. The coherence is determined by a combination of the ground layer coherence (as observed by open fields in the forest) and a lower coherence for a dense vegetation layer. The vegetation coherence is determined by the temporal decorrelation effect and by volume decorrelation. Local topography is not taken into account but can also contribute. The relative importance of the ground component and the vegetation component is determined by the attenuation through the vegetation layer and the gaps in the vegetation layer. The coherence is typically decreasing with stem volume but may oscillate due to the interference between the two contributing factors.

The interferometric effective height is similarly determined by the interference of the scattering from the ground and from the vegetation layer. The effective height of the forest is typically increasing with the true volume, but the difference between the true height and the interferometric effective height can be quite large as function of stem volume and then coherence. The coherence values are obtained as a combination of the vegetation and ground part with a phase difference related to the interferometric phase. As this phase is very sensitive to actual conditions we have also illustrated the case with the two coherence parts combine constructively and destructively. This represents some upper and lower bounds on the coherence values. As seen in Figure 7 the observations fall within these bounds.

Figure 6. The coherence (solid line: coherence; dotted line: ground coherence, dash-dot line: vegetation coherence) and interferometric effective height (solid line) and actual tree height (dotted line) as function of stem volume for a baseline normal to the viewing, Bn, of 200 m.

Figure 7. Illustrating in (a) and (b) coherence with observations. The modelled total coherence value: solid black line; ground coherence: dotted black line; vegetation coherence: dash-dotted black line. Upper and lower bounds of the coherence values due to interaction between ground and vegetation part illustrated as red dash-dotted lines.

Conclusions

Interferometric C-band SAR observations increase the information about forest conditions compared to traditional intensity observations. Such observations have problems due to the saturation even at low stem volume values and also due to the variability of open field signatures. The latter is a problem for identifying clear cuts relative forest. However, coherence values are sensitive to forest/non-forest, see also (Smith et al., 1996).

The backscatter coefficient as well as the coherence are most sensitive at small values of the stem volume. Decreasing the frequency from C-band to e.g. L-band is assumed to increase the sensitivity of both observations to higher stem volumes.

According to the model the interferometric effective forest height is more sensitive at high stem volumes. In this case C-band is expected to be more suitable than lower frequencies due to the lower ground penetration.

In this paper coherence values as function of stem volume have been reported for boreal forest in northern Sweden. Model calculations support the general nature of the observations. The main effect causing low coherence from forest areas is believed to be wind effects causing geometric changes, which are dominating over volume decorrelation. Localisation of storm or fire damages may also be identified due to the geometrical effects. A general decay of coherence with time is attributed to dielectric constant changes.

The low coherence over forest areas may cause problems in DEM derivation due to decreased accuracy. The coherence associated with boreal forest is characterised by vegetation as well as ground contributions giving rise to reasonable high coherence under several conditions. Dense forests like the tropical rain forests cause problems due to the lower coherence associated with only vegetation contributions and there are associated problems in the processing of interferometric SAR images in such cases

Models of the scattering is necessary for the inversion procedure. Interferometric SAR modelling stresses the need for further observations of such model parameter as, attenuation through the canopy, area fill, decorrelation due to wind and dielectric changes.

Acknowledgement

We would like to acknowledge stimulating discussions with Patrik Dammert, Hans Israelsson for MIMICS simulations of attenuation and Johan Fransson, SLU, for providing data on forest stands and for instructing us on how to observe forest parameters.

References

Askne, J., et al., 1995, Retrieval of forest parameters using intensity and repeat-pass interferometric SAR information. Retrieval of Bio- and Geophysical Parameters from SAR Data for Land Applications, Symposium held in Toulouse, on 1995, (ACTES), pp 119-129
Askne, J., et al., to be publ.: C-band repeat-pass interferometric SAR observations of forest, IEEE Transactions on Geoscience and Remote Sensing :
Attema, E. P. W. and F. T. Ulaby, 1978: Vegetation modelled as a water cloud, Radio Science 13(2): 357-364.
Dammert, P. B. G., L. M. H. Ulander and J. Askne, 1995, SAR interferometry for detecting forest stands and tree heights. European Symposium on Satellite Remote Sensing II, Symposium held in Paris, on 1995, pp
Fleischman, Toups and Ayasli., 1992, Summary of results from a foliage penetration experiment with a three-frequency polarimetric SAR. SPIE Surveillance Technologies II, Symposium held in on 1992, SPIE Vol 1693 Surveillance Technologies II pp 151-160
Gatelli, F., et al., 1994: The wavenumber shift in SAR interferometry, IEEE Transactions on Geoscience and Remote Sensing 32: 855-865.
Hagberg, J. O., L. M. H. Ulander and J. Askne, 1995: Repeat-Pass SAR Interferometry over Forested Terrain, IEEE Transactions on Geoscience and Remote Sensing 33: 331-240.
Israelsson, H. and J. Askne, 1994, Retrieval of Forest Biomass using SAR. 14th EARSeL Symposium, Symposium held in Göteborg, 6-8 June, 1994, on 1994, Sensors and Environmental Applications of Remote Sensing (Balkema), pp 97-102
McDonald, K. C. and F. T. Ulaby, 193: Radiative transer modelling of discontinuous tree canopies at microwave frequencies, Int. J. Remote Sensing 14(11): 2097-2128.
Pulliainen, J. T., et al., 1994: Backscattering properties of boreal forests at the C-and X-band, IEEE Transaction on Geoscience and Remote Sensing 32(September): 1041-1050.
Seifert, F. M., H. Kietzmann and M. Zink, 1995, Forest monitoring with SIR-C/X-SAR. Retrieval of Bio- and Geophysical Parameters from SAR Data for Land Applications, Symposium held in Toulouse, on 1995, (ACTES), pp 79-84
Smith, G., P. B. G. Dammert and J. Askne, 1996, Decorrelation mechanisms in C-band SAR interferometry over boreal forest. European Symposium on Satellite Remote Sensing III, Symposium held in Taormina, Italy, on 23-27 September, 1996 1996, pp
Ulaby, F. T., et al., 1990a: Michigan microwave canopy scattering model, International Journal of Remote Sensing 11(7): 1223-1253.
Ulaby, F. T., M. W. Whitt and M. C. Dobson, 1990b: Measuring the propagation properties of a forest canopy using a polarimetric scatterometer, Trans. IEEE Antennas and Propagation 38: 251-258.
Ulander, L. M. H. and J. O. Hagberg, 1995: Radiometric and interferometric calibration of ENVISAT-1 ASAR. Dep. of Radio and Space Science, Chalmers University of Technology,
Zebker, H. A. and J. Villasenor, 1992: Decorrelation of Interferometric Radar Echos, IEEE Transactions on Geoscience and Remote Sensing 30(5): 950-959.

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