Tropical Forest Biomass Density Estimation using ...

7 downloads 5753 Views 744KB Size Report
Tel : +44-1487-773381, Fax : +44-1487-773277, Email : [email protected], [email protected] .... along with the locations of the field campaign sample.
Tropical Forest Biomass Density Estimation using JERS-1 SAR : Seasonal Variation, Confidence Limits and Application to Image Mosaics Adrian Luckman 1, John R. Baker 1, Miroslav Honzák 2, Richard Lucas 2 Remote Sensing Applications Development Unit, British National Space Centre ITE Monks Wood, Abbots Ripton, Huntingdon, Cambridgeshire, PE17 2LS, UK. Tel : +44-1487-773381, Fax : +44-1487-773277, Email : [email protected], [email protected] 2 Department of Geography, University of Wales Swansea Singleton Park, Swansea, West Glamorgan, SA2 8PP, UK 1

Abstract This study describes the development of a semi-empirical model for the retrieval of above-ground biomass density of regenerating tropical forest using JERS-1 Synthetic Aperture Radar (SAR). The magnitude and variability of the response of the L-band SAR to above-ground biomass density was quantified using field data collected at Tapajós in central Amazonia and imagery from a series of dates. A simple backscatter model was fitted to this response and validated using image and field data acquired independently at Manaus, 500km to the west of Tapajós. The sources of variability in biomass density and SAR backscatter were investigated so as to determine confidence limits for the subsequent retrieval of biomass density using the model. This analysis suggested that only three broad classes of regenerating forest biomass density may be positively distinguished. While the backscatter appears to saturate at around 60 tonnes per hectare, the biomass limit for retrieval purposes which is tolerant to both speckle and image texture is only 31 tonnes per hectare. The spatial distribution of biomass density in central Amazonia was estimated by applying the model to a mosaic of 90 JERS-1 images. A favourable comparison of this distribution to a map of regeneration derived from NOAA AVHRR imagery suggested that L-band SAR will provide a useful method of monitoring tropical forests on a regional scale.

1.

Introduction

Forests are important to the environment largely because of the carbon locked up in their biomass and the potential for carbon exchange between this biomass and the atmosphere through forest destruction and regrowth. A number of studies have shown that the above-ground biomass density (hereafter simply referred to as biomass density) of forests may be monitored using Synthetic Aperture Radar (SAR) remote sensing at L-band (23cm) and longer wavelengths [1-5] Such instruments have been deployed for several years on spaceborne platforms such the Shuttle Imaging Radar missions and JERS-1 (the Japanese Earth Resources Satellite) allowing repeat datasets to be acquired. JERS-1 in particular has provided global coverage since 1992 and ongoing work is providing large scale image mosaics covering all the tropical forest areas of the globe. Future Lband SAR missions are also being considered by NASA and NASDA, the US and Japanese space agencies. Given this wealth of past and future L-band SAR data it should be possible to develop algorithms for the retrieval of biomass density of different types of forest using SAR and thereby provide important inputs into models of regional

carbon flux. With this in view, a recent study investigated the sensitivity of imagery from L-band spaceborne SAR instruments on JERS-1 and SIR-C (the third Shuttle Imaging Radar mission) to the biomass density of regenerating tropical forest [6]. This study used field measurements collected from the Tapajós region of central Amazonia in 1994 to quantify the empirical relationship between SAR backscatter and the biomass density of regenerating tropical forest. It was found that the biomass density threshold, above which there appears to be no further increase in L-band radar backscatter, was around 60 tonnes per hectare (t/ha). As carbon uptake by tropical forests is most rapid during early stages of regrowth [7], this suggested that JERS-1 is sensitive to those stages of forest regrowth which have a significant impact on local carbon flux between land surface vegetation and atmosphere. This paper extends the earlier work by considering how the biomass density of tropical forests may be retrieved from L-band SAR data. The objective of this study is to generate and test a biomass density retrieval scheme that can be used to map the biomass density of regenerating tropical forest using JERS-1 SAR. An important element of this objective is to quantify the expected accuracy and confidence limits of the retrieval scheme so that the true value of the SAR data for this application may be assessed. The method employed for the development of a biomass density retrieval scheme may be summarised by the following stages: 1) Derive a simple forest backscatter model and fit it to image and ground data of biomass density from a tropical forest test site in Amazonia. 2) Perform an error analysis so as to quantify the confidence limits for the subsequent retrieval of biomass density using the model. 3) Validate this model using field and image data from an independent test site in the same type of forest but at a different location within Amazonia. 4) Propose a biomass density retrieval scheme which takes into account seasonal variation in signature and estimated confidence limits. 5) Apply this scheme to a mosaic of JERS-1 SAR imagery from Amazonia and compare the resulting classification with a comparable map derived from optical and nearinfra-red earth observation data. These steps, and a description of the sources of field and image data are described in detail in the following sections.

2.1

2. Sources of Data Test sites and field data

The study uses field data and remotely sensed images from the Tapajós region of Pará State and the Manaus region of Amazonas State in Brazil and mosaics of imagery covering approximately 250,000 sq. km of central Amazonia. The geographical location of these study areas within South America is shown in Figure 1.

To quantify the temporal variation in SAR signature, six JERS-1 coverages of Tapajós, each comprising of a consecutive pair of images, were acquired from the above dates. A single JERS-1 coverage (also a pair of images) of the Manaus test site from 1/10/93 was also acquired. Each of these individual coverages was georegistered to the UTM projection so that image statistics could be easily extracted for areas corresponding to sample plots on the ground. Figure 2 shows examples of the geocoded JERS-1 imagery along with the locations of the field campaign sample measurement plots.

Manaus

Tapajos

Figure 1. Location of Tapajós and Manaus test sites and the JERS-1 Amazon mosaic within South America.

The Tapajós region contains a variety of land cover including cattle ranches, subsistence farming, regenerating abandoned plots and protected mature forest. During a field campaign in 1994, fifteen plots of 10m by 50m (0.05 ha) representing a suitable range of regrowth age were measured for diameter at breast height (dbh), species and tree height. Biomass density was derived by employing timber specific gravity values and published speciesdependent regression equations [6]. The region to the north of Manaus is characterised by abandoned ranches either side of the BR174 highway. Fieldwork campaigns were carried out in this area in August 1993 and August 1995 concentrating on the abandoned Fazendas Agroman, Maringa, Dimona, Porto Alegre and Esteio [8]. As at Tapajós, tree species, height and dbh were recorded within small sample plots but this time of 10m by 100m (0.1 ha) rather than 10m by 50m. Estimates of biomass density were derived in the same way as at Tapajós and the measurements from selected sample plots were later combined to give mean biomass density values for homogeneous forest areas. As a result, a total of 16 homogeneous areas of forest at Manaus were sampled. 2.2

Image data

The remotely sensed imagery used in the study included SAR data from JERS-1 and an extract from a map of land cover and regenerating forest biomass density derived from an Amazon-wide compilation of AVHRR data. Image 1 2 3 4 5 6

Date 22/8/92 14/2/93 30/3/93 26/6/93 27/7/94 10/10/95

Season dry wet peak wet dry dry peak wet

Figure 2. Geocoded JERS-1 images from Tapajós and Manaus showing the location of the biomass density sample plots. Tapajós (left): 50km x 100km scale Manaus (right): 60km x 80km scale

A 500km by 500km preliminary mosaic of ninety JERS-1 images from Amazonia was also employed in the study. This was acquired between February and March 1993 and processed jointly by the Japanese Ministry of International Trade and Industry (MITI) and National Space Development Agency (NASDA) [9]. The mosaic is presented as 8-bit 1 ha pixels averaged and quantised from 64 original 3-look 16-bit JERS-1 pixels and is shown in Figure 3. This preliminary mosaic was generated by an early version of the NASDA SAR processor which introduced a slight range-dependent calibration error appearing as striping between individual scenes and which will be corrected in future versions. Whilst this error has a small effect on the subsequent retrieval of biomass density from the mosaic, it does not significantly affect the conclusions of the study or the applicability of the technique to subsequent JERS-1 mosaics.

(approximately 0-40 t/ha, 40-80 t/ha and 80-120 t/ha) for which the confidence limits have yet to be fully evaluated. For this study, only the forest/non-forest boundary and the 0-40 t/ha class were considered. This map does not provide an ideal means of validation for the retrieval of biomass density from JERS-1 mosaics. It is based upon 1km data while the JERS-1 mosaic is presented at ten times this resolution, and has not been validated. However, remote sensing is the only method of acquiring information at such synoptic scales and this map provides a useful method of qualitatively assessing the use of JERS-1 mosaics for biomass density retrieval. 2.3

Figure 3. The JERS-1 mosaic from Central Amazonia.

As an independent source of data for the distribution of biomass within Amazonia, a map of regenerating forest derived from 1km NOAA (National Oceanic and Atmospheric Administration) AVHRR (Advanced Very High Resolution Radiometer) imagery was used [10]. This map was generated using atmospherically corrected AVHRR bands 1, 2 and 3 and a vegetation index and calibrated with the use of the field measurements collected at Manaus and a multiple time-series of Landsat TM data from various parts of the Amazon. The boundaries within the map between forest and non-forest were based upon the TREES (Tropical Ecosystem and Environment observation by Satellite) classification [11]. The original map included three categories of regenerating forest classified by above-ground biomass density

Extraction of image signatures

Image regions or polygons with a minimum area of 1 ha corresponding to the homogeneous forest areas sampled on the ground at both Tapajós and Manaus were located by a combination of image geocoding and GPS positioning. From first-hand knowledge of the field sites, further image areas were defined corresponding to undisturbed forest and to pasture. These were assigned nominal biomass density values of 400 t/ha and zero t/ha respectively and provide important data points for the subsequent curve-fitting procedure. For all of these polygons, the mean and variance in backscatter were extracted from the geocoded JERS-1 data (6 coverages for Tapajós and one for Manaus) so as to quantify the relationship between s0 (the backscattering coefficient) and biomass density. The values extracted from the imagery, and the corresponding biomass density estimates are given in Table I for Tapajós and in Table II for Manaus. The relationships between backscatter and biomass density for each Tapajós image are shown in Figure 4 while Figure 5 indicates the relative timing of image acquisitions, biomass density measurements and mean seasonal cycle at both Tapajós and Manaus.

Table I. Biomass density, JERS-1 images statistics and derived values for sampled forest polygons at Tapajós. Polygon Number

Above-Ground Biomass Density (t / ha) ±30%

Polygon Size (pixels)

Mean s0 (Linear Units)

Modelled s0 (Linear Units) Equation 1

Weighted Square Residual Equation 3

Sum of Squares of Within Polygon Variation Equation 2

Measured Variance

Estimated Texture Variance

1

62

1204

0.179

0.166

0.0562

1.628

0.00541

0.168

2

15

1260

0.127

0.117

0.0312

0.865

0.00275

0.171

3

62

972

0.163

0.166

0.0019

1.161

0.00478

0.179

4

8

92

0.100

0.096

0.0004

0.040

0.00175

0.171

5

54

816

0.140

0.163

0.1110

1.102

0.00540

0.275

6

82

2920

0.170

0.169

0.0022

4.227

0.00579

0.199

7

78

2344

0.165

0.168

0.0060

3.267

0.00558

0.204

8

104

1128

0.168

0.170

0.0005

1.511

0.00536

0.188

9

75

1160

0.155

0.168

0.0503

1.475

0.00509

0.212

10

181

1676

0.168

0.170

0.0029

3.327

0.00794

0.283

11

101

1524

0.175

0.170

0.0102

2.459

0.00645

0.211

12

42

500

0.170

0.157

0.0212

0.674

0.00539

0.185

13

89

5716

0.169

0.169

0.0006

7.553

0.00529

0.186

14

25

2040

0.133

0.140

0.0244

1.764

0.00346

0.196

15

387

6088

0.178

0.170

0.1025

11.843

0.00778

0.244

16

0

1028

0.052

0.053

0.0002

0.185

0.00072

0.262

17

400

3136

0.161

0.170

0.0618

5.738

0.00732

0.281

18

400

13688

0.171

0.170

0.0018

23.631

0.00691

0.236

mean weighted square residual =

0.0270

mean within-polygon sum of square variation =

4.025

Table II. Biomass density, JERS-1 images statistics and derived values for sampled forest polygons at Manaus. Polygon Number

Above-Ground Biomass Density (t / ha) ±30%

Polygon Size (pixels)

Mean s0 (Linear Units)

Modelled s0 (Linear Units) Equation 1

Weighted Square Residual Equation 3

Sum of Squares of Within Polygon Variation Equation 2

Measured Variance

Estimated Texture Variance

1

97

30280

0.158

0.170

3.7787

41.969

0.00554

0.221

2

93

10160

0.156

0.169

1.7539

14.405

0.00567

0.232

3

24

764

0.129

0.138

0.0575

0.903

0.00473

0.281

4

81

1808

0.170

0.169

0.0032

2.222

0.00492

0.170

5

99

3576

0.160

0.170

0.3032

4.514

0.00505

0.196

6

70

732

0.168

0.167

0.0010

0.684

0.00374

0.131

7

400

5940

0.188

0.170

1.9738

11.625

0.00783

0.220

8

0

3748

0.059

0.053

0.1219

0.763

0.00081

0.235

9

118

1248

0.182

0.170

0.1781

2.065

0.00662

0.200

10

113

676

0.195

0.170

0.4225

1.055

0.00624

0.164

11

113

1188

0.152

0.170

0.3722

1.366

0.00460

0.198

12

136

996

0.180

0.170

0.0879

1.708

0.00686

0.212

13

125

596

0.141

0.170

0.4923

0.744

0.00499

0.249

14

108

1636

0.175

0.170

0.0409

2.682

0.00656

0.214

15

144

4648

0.171

0.170

0.0008

0.798

0.00682

0.231

16

100

848

0.183

0.070

0.1487

1.936

0.00913

0.273

17

86

1540

0.184

0.169

0.3536

2.808

0.00729

0.215

18

60

704

0.180

0.165

0.1444

0.756

0.00430

0.132

mean weighted square residual =

0.5686

mean within-polygon sum of square variation =

Luckman et.al.

Tropical Forest Biomass Density Estimation Using JERS-1 SAR

s0 -7 (dB)-8

s0 -7 (dB)-8

-9 -10

-9 -10

-11

-11

-12

-12

-13 -14

-13 -14

-15

Image 1 (22/8/92)

-15

0 50 100150200250300350400

-7

(dB)-8

(dB)-8 -9

-10 -11

-10 -11

-12

-12

-13

-13

Image 3 (30/3/93)

-14 -15

Image 4 (26/6/93)

Biomass Density (t / ha)

Biomass Density (t / ha) Tapajós JERS1 image 5 Tapajós fieldwork

s0 -7 (dB)-8

-9

-9 -10

-11

-11

-12 -13

-12 -13

-15

Tapajós JERS1 image 4 Manaus fieldwork part 1 Manaus JERS1 image

0 50 100150200250300350400

-10

-14

Tapajós JERS1 image 2 Manaus JERS1 mosaic Tapajós JERS1 image 3

0 50 100150200250300350400

s0 -7 (dB)-8

Image 5 (27/7/94) 0 50 100150200250300350400

Biomass Density (t / ha)

Tapajós JERS1 image 1

-7

-9

-14 -15

Tropical Forest Biomass Density Estimation Using JERS-1 SAR

Image 2 (14/2/93) Biomass Density (t / ha)

s0

Luckman et.al.

Activity or Acquisition

0 50 100150200250300350400

Biomass Density (t / ha) s0

February 1997

5.167

-14 -15

Image 6 (10/10/95) 0 50 100150200250300350400

Biomass Density (t / ha)

4. The relationship between JERS-1 and FigureFigure between JERS-1 backscatter and backscatter estimated biomass 4 - The relationship estimated biomass atemployed Tapajós in for density at Tapajós for thedensity six images thethesix study. images The fittedemployed backscatterin models and error barsfitted indicating three times the standard error in the mean the study. The backscatter models and error bars indicating backscatter andthe +/- standard 30% of the biomass are shown. three times error indensity the mean backscatter and ± 30% of the biomass density are shown.

Manaus fieldwork part 2 Tapajós JERS1 image 6

Date

February 1997

Mean monthly rainfall (mm) J F M A M J J A S O N D J F M A M J J A S O N D J F M A M J J A S O N D J F M A M J J A S O N D

Figure 5. The timing of all SAR images acquisitions and ground data collection campaigns relative to the mean seasonal cycle.

Figure 5 - The timing of all SAR image acquisitions and ground data collection campaigns relative to the mean seasonal cycle

3.

Backscatter Model

It was necessary to fit a model to the biomass-backscatter responses to enable them to be used in the retrieval of biomass density from further images. A simple though plausible model for the relationship between biomass density and microwave backscatter from vegetation is: -(bB+c)

s0 (linear units) = a - e

(1)

where B is the biomass density, a, b and c are constants and the incidence angle is assumed not to vary significantly. This equation is based loosely on the water cloud model [12] which essentially represents the extinction of microwave radiation as it passes through a layer of vegetation made up of elements containing water. Parameter a of the model corresponds exactly to the saturation value of s0 in linear units. Parameter b describes the gradient of the low biomass-density part of the curve while parameter c relates to the residual backscatter at zero biomass (i.e. the nominal backscatter from bare soil). As a model to fit empirically to the JERS-1 responses at Tapajós, it has two advantages: · It is robust to the limited number of sampled biomass density measurements as it has only 3 parameters. · It has the potential of allowing residual backscatter at zero biomass density to allow for soil surface scattering at low vegetation cover (although the attenuation of the soil scattered radiation on the path back through the vegetation is not accounted for). This model was fitted to each of the six JERS-1 responses by the Levenberg-Marquardt non-linear least-squares method, weighted by the standard error of the mean [13]. The resulting curves are shown in Figure 4. To compare these fitted models and hence the JERS-1 responses at different times of the year, several parameters were extracted from the curves. These included the three free variables (a, b and c), s0 at saturation (in dB) and the biomass density giving rise to s0 1dB below saturation. These parameters are given in Table III for each of the six JERS-1 images along with the RMS residual error of the fit of the curve to the data. Table III. Parameters of the fitted backscatter model for each Tapajós JERS-1 image. Image

Parameter of Curve a b c

RMS Residual Error (linear units)

Saturation s0 (dB)

Biomass Density 1dB Below Saturation (t / ha)

1

0.148

0.047

2.235

0.045

-8.29

26.9

2

0.165

0.037

2.266

0.051

-7.83

29.8

3

0.166

0.060

2.319

0.027

-7.80

17.6

4

0.165

0.036

2.157

0.051

-7.84

33.8

5

0.170

0.053

2.146

0.038

-7.69

22.8

6

0.156

0.048

2.081

0.081

-8.06

28.2

The biomass density corresponding to s0 1dB below saturation gives a relative indication of the maximum biomass density that may be reliably retrieved using these empirical relationships and thus indicates in which season JERS-1 data is best acquired for the purposes of biomass density estimation. As might be anticipated due to the relatively smaller contribution from the soil backscatter, the images from the dryer parts of the seasonal cycle in general

give the greater dynamic range in backscatter between low and high biomass areas and the higher biomass densities at which saturation begins to occur. Image 4 (June - dry season) seems to have the most favourable response indicating that the dry season is the most suitable period of acquisition over Amazonia for this application. Any change in biomass density, especially in the younger forest areas, due to growth or extraction within the four year period of JERS-1 image acquisition (before or after the field campaign) could not be accounted for. Hence the relationship obtained with the highest confidence is the one derived from the image closest in date to the biomass sampling (Image 5, July 1994). The clearest dissimilarity between each of the fitted curves is between the exhibited saturation levels and this difference is independent of the accuracy of the biomass density estimates. As mature tropical forest is believed to present very stable backscattering properties [14], the differences in saturation level is most probably due to calibration inconsistencies. The maximum difference between any two of the images is 0.6dB which is well within the specified JERS-1 one-sigma calibration accuracy of 1.86dB [15]. 4. Confidence Limit Analysis A critical part of the retrieval of biomass density from remotely sensed data is the quantification of confidence limits. These depend on errors in the sources of data used to develop the retrieval model and in the variability of new data acquisitions. 4.1

Variability in the retrieval model

Potential sources of error in the retrieval model include those in the estimation of the sample plot biomass density and in the fit of the simple backscatter model to the image data. The confidence in biomass density estimates is difficult to quantify but taking the known sources into account, the error is estimated to be no worse than 30% of the biomass density value [6]. The adequacy of the model may be investigated by examining the residuals of the model fit to check that there is no systematic variability in s0 other than that explained by the expected behaviour. The residuals may also be compared to the within-polygon image variances to test whether the deviation from the model is adequately explained by the speckle and texture features within the image. This comparison is only valid if the variance is distributed evenly throughout the data points. Although the multiplicative nature of the noise concentrates the variance at large backscatter values, this is offset by the relatively larger sizes of polygon in the older forest areas. The residual deviation of the data points from the model developed from the Tapajós data are given in Table I. The mean of these residuals does not deviate significantly from zero and they show no systematic variation with either the measured backscatter or the estimated biomass density so the model appears satisfactory from this point of view. The within-polygon sum of square variations about the mean are shown in Table I and are given by:

2

SS m = ån =m1 (x n - x m ) N

(2)

where Nm is the number of pixels in polygon m, xn is the pixel value and x m is the polygon mean. The weighted mean value for all polygons is 4.03 (Table I). This value may be compared to the weighted square residual deviation of the polygon means from the model which is given by: weighted square residual = åm =1 N m (xˆ m - x m ) M

where

xˆ m

2

(3)

is the value of backscatter predicted by the

model at a particular biomass density value. The mean weighted square residual for all polygons is only 0.027 thus demonstrating that the deviation from the model is adequately explained by the variation due to speckle and texture within the image areas and that there is no evidence here that there are contributions from other sources. The fit of the backscatter model to the data from Tapajós Image 5 is shown graphically in Figure 6 along with error bars indicating twice the standard error of the mean. Luckman et.al.

Tropical Forest Biomass Density Estimation Using JERS-1 SAR

February 1997

Figure 6. A representation of the quantised biomass density retrieval scheme showing the Tapajós images 5 and Manaus data points and error bars at twice the standard error of the mean backscatter.

-7 saturation at -7.7 -8 1.2 dB interval

-9

Sigma0 (dB)

-10 Tapajós field data Model fit to Tapajós data

-11

Manaus field data

-12 1

2

6

10 13

31

100

Areal Density of above-ground biomass (t/ha)

500

Variability in s0 for subsequent images Potential variability in measured s0 arises from five possible sources: 4.2

Figure 6 - A representation of the quantised biomass density retrieval scheme showing the Tapajós Image 5 and Manaus data points and error bars at twice the standard error of the mean backscatter

· · · · ·

JERS-1 calibration inconsistency Topographic effects on SAR backscatter Image speckle Target inhomogeneity leading to image texture Variability due to factors other than biomass density (e.g. local moisture conditions, species distribution)

The first of these error sources may be compensated for by assuming that any calibration error is multiplicative (i.e.

additive in the logarithmic or dB scale) and that mature tropical forest exhibits constant backscatter at L-band regardless of time or season [14]. This assumption is supported by the relatively small differences between saturation values in each of the six JERS-1 image from Tapajós as noted earlier. Thus image cross-calibration may be performed based on the difference in s0 between large forest areas in a new image and the mature forest saturation value for Tapajós Image 5 (which is taken nominally as 7.69dB). Variations in backscatter due to the influences of topography on per-pixel scattering area may be corrected using a DEM [16, 17]. Variations caused by the dependence of different scattering mechanisms on incidence angle is target dependent and therefore less well quantified but can also be compensated for using simple land-cover-dependent empirical models [18, 19]. In this study, the terrain at Tapajós and at Manaus has little variation in height and did not cause a significant problem for the test areas. However, a DEM will be required to compensate for topography for biomass density retrieval from areas with large variations in topography of which there are some within the JERS-1 mosaic area and many throughout the Amazon basin. Biomass density and topography are unlikely to be the only factors within regenerating and mature tropical forest areas to influence the level of backscatter. Other sources of variability might for instance include the distribution of species within the forest leading to structural changes in the canopy, or changes in the dielectric constant of the soil and plant matter as a result of variations in their moisture content. Although such a limited study cannot hope to identify and quantify these other factors, it is possible to assess whether they have a significant effect on the results by comparing the responses from the two test sites investigated. The range of biomass densities sampled at Manaus is not as wide as would be desired for a rigorous validation of the model. However, the within-polygon variations at Manaus and their mean square sum (5.17) is larger than the weighted mean square residual deviation of the Manaus polygon means to the Tapajós-based model (0.569). Therefore the Manaus data gives no reason to reject the model for not being valid over a wider area. However, it should be noted that Manaus residuals are larger than those at Tapajós (0.270). The data from the Manaus test site is given in Table II and its comparison to the Tapajós-based model is shown in Figure 6. 4.3

Variability due to speckle and texture

Image speckle arises from the coherent nature of SAR imaging while image texture is caused by inhomogeneities in the target on a spatial scale greater than one pixel. Both give rise to statistically random multiplicative variations in the measurement of s0 whose distributions are indistinguishable from Gaussian for large numbers of pixels [20]. Given the relatively large image areas used to characterise the backscattering coefficients in this study, the speckle variation is likely to be insignificant compared to that due to texture. The combined variance of both effects is described by the multiplicative model [21] and is given by:

1ö æ VT sd 2 = m 2 ç + VT + ÷ N N è ø

where sd2 is the expected variance due to speckle and texture which depends on the mean s0 (m), the texture variance (VT) and the number of independent samples (N). Rearranging Equation (4) gives the ratio of the standard deviation to the mean backscatter or the coefficient of variation (CV): sd VT 1 = + VT + N m N

(5)

The number of independent samples (N) is given by the look averaging of the image multiplied by the number of pixels in the image area used to estimate s0. Although the nominal look averaging for JERS-1 is 3, in reality these looks are not entirely independent. An analysis of JERS-1 imagery of areas of ground that are believed to contain minimal textural variations and appear to obey the Square Root of Gamma distribution gave an estimate for the equivalent number (ENL) of looks as 2.8 [22]. For biomass retrieval purposes, the size of area must be defined. In this study areas of 64 pixels of 12.5m by 12.5m representing 1ha of ground are used because this corresponds to the resolution of the JERS-1 image mosaics and also defines a meaningful size of forest plot. Hence N in this case is 179.2 and the variability due to speckle will be very small. The texture variance (VT) may be estimated from the image areas used to characterise the biomass density-backscatter relationships by rearranging Equation (5) to give: 2

æ sd ö 1 çç ÷ m ÷ø N è VT = 1 1+ N

æ sd 1 ö ÷÷ = +0.55dB or - 0.63dB (7) * 10 * log10 çç1 ± 2 * m 64 è ø

(4)

(6)

where sd2 is the measured variance within each polygon, m is the mean s0 and N is the number of pixels in the textured area multiplied by the ENL. The image texture might be expected to increase with biomass density because the homogeneity of the forest decreases as pioneer species are gradually replaced by late secondary or primary species exhibiting a higher variability in canopy structure [23]. Hence the variability in measured s0 due to texture should be lower for younger forest areas. However, in this case the worst case texture variance appears to occur in the middle of the range and is used to define error margins for all forest areas so as to account for any possible textural variations in other images. Tables I and II gives the values for measured variance and estimated texture variance at Tapajós and Manaus using Equation (6). The worst case texture variance occurs within Polygon 10 at Tapajós and has a value of 0.2826 . As the overall deviation is dependent on the mean backscatter, the variability due to speckle and texture will be constant in the logarithmic (dB) scale. An image area of 64 pixels exhibiting worst-case texture of 0.2826 will have an expected CV (sd/m) from Equation (5) of 0.5378. Confidence limits in dB of twice the standard error of the mean for such an image region are given by:

Based on worst-case texture and speckle predictions alone therefore, we can expect 95% of all estimates of s0 to lie within the region +0.55dB to -0.63dB from the mean response. 4.4

Summary

The variability in s0 due to calibration and topographic factors may be compensated for by the use of image cross calibration and a DEM. The residual error in the fit of the simple backscatter model is adequately explained by the texture and speckle in the sample forest areas and data collected independently at Manaus does not give any reason to reject the backscatter model. Using the worst-case texture found at Tapajós and Manaus and the expected levels of speckle, confidence limits of +0. 55dB and -0. 63dB (2 standard errors) may be given for the estimation of s0 from 1 ha regions of further JERS-1 images. The corresponding confidence limits of retrieved biomass density may be derived from the model and combined with the expected biomass density error of +/- 30%.

5.

Biomass Density Retrieval Scheme

A suitable biomass density retrieval scheme therefore comprise of the following steps:

would

1) Cross-calibrate the gain of a new image to the July 1994 data such that mature forest areas have the same backscatter level of -7.69dB. 2) Compensate for scattering area with the use of a DEM and knowledge of the instrument track and altitude. Calibrate for incidence-angle dependence of scattering mechanisms if suitable models can be derived for tropical forest. 3) Calculate the mean backscatter in homogeneous 1 ha (or larger) areas of regenerating and mature tropical forest (the scheme is not valid for other land cover types). Use the model fitted to the July 1994 data to retrieve biomass density. 4) Calculate the confidence limits in the retrieved biomass density using the model and the expected variability in s0. By quantising the measured backscatter into bins corresponding to limited ranges of biomass density, the retrieval scheme may be used to classify JERS-1 data according to the biomass density of regenerating forest. The size of bins should be constant in the logarithmic (dB) scale and equal to the confidence interval calculated for worst case texture and speckle based on 1 ha samples. This quantised scheme is shown in Figure 6 and indicates the limits of biomass density bins that may be retrieved with confidence at the 95% level. The quantised retrieval scheme is summarised in Table IV.

Table IV. Biomass density and backscatter thresholds for the quantised retrieval scheme. Lower s0 Threshold

Upper s0 Threshold

Typical Land Cover

Regenerating Forest Biomass Density Class

noise floor

-11.9 dB

inland water

-11.9 dB

-10.7 dB

pasture and crops

-10.7 dB

-9.5 dB

young regrowth

6- 13 t/ha

-8.3 dB

established regeneration

13 - 31 t/ha

-7.1 dB

old regeneration to primary forest

> 31 t/ha

maximum

flooded forest and urban areas

-9.5 dB -8.3 dB -7.1 dB

6.

Application to Image Mosaic

To assess its usefulness for the large scale mapping of biomass density in regenerating tropical forests, the biomass density retrieval scheme was applied to the JERS-1 Amazonia mosaic. The mean s0 was recomputed by a simple relationship from the 8-bit pixels and the retrieval scheme was applied to this value to predict the mean biomass density within each 1 ha area of forest. Figure 7 shows the retrieved biomass density classification calculated from the Amazonia JERS-1 mosaic using the quantised retrieval scheme. The part of the classified image within the box has been enlarged for clarity. The retrieval scheme allows biomass density of regenerating tropical forest to be estimated in the three classes shown but further tentative classes may be proposed based on experience of JERS-1 data from this area and on local knowledge. These classes and the corresponding thresholds in s0 are shown in Table IV. It is assumed that backscatter values below 11.9dB are most likely to be associated with still inland water as this generally exhibits very low backscattering properties and that image values above -7.1dB are most likely to be associated with urban areas or flooded forest because of the additional double-bounce scattering mechanism that occurs at these targets [24, 25].

non-forest limits are similar although the AVHRR classification includes flooded forest in this category. Areas of regenerating forest appear to correspond well, especially along the BR174 to the north of Manaus (in the enlarged areas of Figures 7 and 8) where several abandoned ranches are to be found. It should be noted that the AVHRR and JERS-1 classes are not absolutely congruent in biomass density limits but they are close enough to make a useful comparison. The most notable differences in the maps include the cerrado areas to the north-west, the mountainous areas to the north-east and the regions along the Rio Negro and Solimões, all of which indicate regeneration where experience suggests it does not occur. These areas highlight the need for a priori knowledge of land cover so that only tropical forest areas may be considered, and the need for topographic calibration using a DEM to prevent misclassifications due to terrain effects on SAR backscatter.

Figure 8. Land cover derived from AVHRR imagery. Left: Land cover from AVHRR (500km x 500km scale) Right: Enlarged area (125km x 250km scale)

A quantitative comparison was made by considering the proportions of the classes detected by each method and examining a contingency matrix between the two classifications. This was carried out on the enlarged area within Figure 7 (125km by 250km) as it avoids the boundaries of the JERS-1 mosaic. Table V shows the relative proportions of each class detected while Table VI shows the contingency table between the AVHRR and JERS-1 classifications. Table V. Proportions of land cover classes found within the image subarea by the AVHRR and JERS-1 classification schemes.

Figure 7. Biomass density and land cover retrieved using the mosaic of JERS-1 images from Amazonia. Left: Land cover and biomass from JERS-1 (500km x 500km scale) Right: Enlarged area (125km x 250km))

The part of the AVHRR map corresponding to the JERS-1 mosaic is shown in Figure 8 along with an enlargement of the area from Manaus City in the south to the Balbinas Reservoir in the north. There appears to be a general correspondence between the JERS-1 and AVHRR derived biomass density maps. The

Data Source

Mature Forest

AVHRR

77.9% Forest > 31 t/ha

JERS-1 Mosaic

92.3%

Regenerating Forest

Non-Forest

0-40 t/ha

(+ some flooded forest)

1.1%

21.0%

Regenerating Forest 6-31 t/ha 2.4%

Non-Forest 5.2%

Table VI. Contingency table between the AVHRR and JERS-1 classifications within the mosaic image sub-area. JERS-1 Regenerating Forest

Non-Forest

Mature Forest

82.4% (46582 pixels)

1.1% (629 pixels)

16.5% (9339 pixels)

Regenerating Forest

35.3% (524 pixels)

1.1% (17 pixels)

63.6% (944 pixels)

Non-Forest

19.1% (613 pixels)

0.4% (13 pixels)

80.5% (2589 pixels)

Mature Forest

AVHRR

There is a reasonable correspondence between the two classifications of the proportions of land in each class although the flooded forest associated the Balbinas hydroelectric scheme (in the north of the enlarged area) and the Rio Negro is classified as non-forest in the AVHRR map and thereby reduces the apparent amount of mature forest. The contingency matrix shows a satisfactory match between mature and non-forest areas but a poor correspondence between the areas of regeneration detected in each image. This may be due to a variety of factors including the differences in class definition and imperfect image registration.

7.

Results and Discussion

This study has employed JERS-1 image data from a series of dates and field data collected at Tapajós and Manaus in the Brazilian Amazon to develop a semi-empirical model for the retrieval of above-ground biomass density within regenerating tropical forest. A detailed analysis of error sources has led to a retrieval scheme with well-quantified confidence limits that has been shown to produce comparable results to other remotely sensed data when applied to a large-scale data set. The need for accurately calibrated and topographically corrected data has been demonstrated as well as the need for a priori land cover information to determine where the biomass density retrieval scheme may be applied. The potential for retrieving extra information from JERS-1 data, such as the location of flooded forests, and the independence of this data from meteorological effects, gives this method an advantage over optical remote sensing techniques. While AVHRR can only be employed as a combination of cloud-free pixels from different days, JERS-1 images can take a high resolution snapshot of the forest in any given 44 day period. The biomass density threshold for tropical forest at which the L-band JERS-1 SAR begins to saturate appears to be lower than that for other forest types such as a coniferous plantation [1] but comparable to other studies of broadleaf forest [26]. The estimated confidence limits allow an evaluation to be made of the maximum realistically retrievable biomass density and appear to be determined not by the level of speckle, which becomes negligible when considering area of 1 ha or more, but by the inhomogeneities in the forest which result in image texture. Although the biomass threshold may be too low to allow a detailed inventory of mature forest, it will certainly allow the extent of regenerating forest areas to be quantified and the biomass density to be categorised. This will provide a

method of measuring those stages of forest regrowth which have a large impact on local carbon uptake from the atmosphere. This study validates the ongoing work of NASDA and NASA to provide mosaics of JERS-1 covering the whole of the Brazilian Amazon and other tropical forest areas.

8.

Aknowledgements

This study was funded by the UK Natural Environment Research Council as part of the Terrestrial Initiative in Global Environment Research (TIGER). Fieldwork was carried out in collaboration with the Instituto Nacional de Pesquisas Espaciais (INPE - the Brazilian Space Agency), the Instituto Nacional de Pesquisas da Amazônia (INPA - Brazil’s Amazon Research Institute), Sheffield University and the NERC ABRACOS project. Essential assistance was given in the field by the Instituto Brasileiro do Meio Ambiente (IBAMA - Brazil's environmental and renewable resources agency), the Superintendência do Desenvolvimento da Amazônia (SUDAM - Brazil's agency for the development of the Amazon) and the Smithsonian Institute. Landsat Images used for calibrating the AVHRR classification were provided through INPE by Thelma Krug and Silvana Amaral. JERS-1 images and mosaics were provided by NASDA. We are indebted to many people for their help in this study, not least those who helped in the ground data collection at Tapajós and Manaus. These include Bruce Nelson, Ieda do Amaral, Corina Yanasse, Tatiana Mora Kuplich, Pedro Hernandez Filho, Geoff Groom, Saira Luckman and Kevin Grover.

9.

References

[1] T. Le Toan, A. Beaudoin, and D. Guyon, “Relating Forest Biomass to SAR Data,” IEEE Transactions on Geoscience and Remote Sensing, vol. 30, pp. 403-411, 1992. [2] E. J. Rignot, R. Zimmerman, and J. J. van Zyl, “Spaceborne Applications of P Band Imaging Radars for Measuring Forest Biomass,” IEEE Transactions on Geoscience and Remote Sensing, vol. 33, pp. 1162-1169, 1995. [3] P. A. Harrell, L. L. Bourgeau-Chavez, E. S. Kasischke, N. H. F. French, and N. L. Christensen, “Sensitivity of ERS-1 and JERS-1 Radar Data to Biomass and Stand Structure in Alaskan Boreal Forest,” Remote Sensing of Environment, vol. 54, pp. 247-260, 1995. [4] G. Sun and K. J. Ranson, “Relating Multifrequency Radar Backscattering to Forest Biomass: Modeling and AIRSAR Measurement,” presented at Third Annual JPL Airborne Geoscience Workshop, Pasadena, USA, 1992. [5] K. J. Ranson and G. Sun, “An Evaluation of AIRSAR and SIR-C/X-SAR Images for Mapping Northern Forest Attributes in Maine, USA,” Remote Sensing of Environment, vol. 59, pp. 203-222, 1997. [6] A. J. Luckman, J. R. Baker, T. M. Kuplich, C. C. F. Yanasse, and A. C. Frery, “A Study of the Relationship between Radar Backscatter and Regenerating Forest Biomass for Spaceborne SAR Instruments,” Remote Sensing of Environment, vol. 60, pp. 1-13, 1997. [7] C. Uhl, R. Buschbacher, and E. A. S. Serrao, “Abandoned Pastures in Eastern Amazonia. I. Patterns of Plant Succession,” Journal of Ecology, vol. 76, pp. 663-681, 1988.

[8] R. M. Lucas and M. Honzàk, “Secondary Forests at Manaus: Data collected during a field campaign, JulyAugust, 1993. Report to the Instituto Nacional de Pesquisas Espaciais (INPE) and the Instituto Nacional de Pesquisas da Amazônia INPA.,” Department of Geography, University of Swansea 1995 1995. [9] Y. Hashimoto and K. Tsuchiya, “Investigation of Tropical Rain Forest in Central Amazonia, Brazil based on JERS-1 SAR Images,” Journal of Geography, Japan, vol. 104, pp. 827-842, 1995. [10] R. M. Lucas, “Integration of NOAA AVHRR and fine spatial resolution imagery for tropical forest monitoring,” in Advances in the use of AVHRR Data for Land Applications, J. P. Malingreau and A. Belward, Eds. Dordrecht: Kluwer Academic, 1996, pp. 371-394. [11] J. P. Malingreau and A. S. Belward, “Recent Activities in the European-Community for the Creation and Analysis of Global AVHRR Data Sets,” International Journal of Remote Sensing, vol. 15, pp. 3397-3416, 1996. [12] E. P. W. Attema and F. T. Ulaby, “Vegetation modeled as a water cloud,” Radio Science, vol. 13, pp. 357-364, 1978. [13] D. W. Marquardt, Journal of the Society for Industrial and Applied Mathematics, vol. 11, pp. 431-441, 1963. [14] I. J. Birrer, E. M. Bracalente, G. J. Dome, J. Sweet, and G. Berhold, “Sigma-zero Signature of the Amazon Rainforest Obtained from the Seasat Scatterometer,” IEEE Transactions on Geoscience and Remote Sensing, vol. 20, pp. 11-17, 1982. [15] Shimada, “Radiometric and Geometric Calibration of JERS-1 SAR,” Advanced Space Research, vol. 17, pp. 7988, 1996. [16] J. J. van Zyl, B. D. Chapman, P. Dubois, and J. Shi, “The Effect of Topography on SAR Calibration,” IEEE Transactions on Geoscience and Remote Sensing, vol. 3, pp. 1036-1043, 1993. [17] A. J. Luckman and J. R. Baker, “The effects of topography on radar scattering mechanisms from coniferous forest and upland pasture,” presented at MAC-

Europe final results workshop, Lenggries, Munich, Germany, 1994. [18] J. J. van Zyl, “The Effect of Topography on Radar Scattering from Vegetated Areas,” IEEE Transactions on Geoscience and Remote Sensing, vol. 31, pp. 153-160, 1993. [19] A. J. Luckman, “The Effects of Topography on Mechanisms of Radar Backscatter from a Coniferous Forest and Upland Pasture,” IEEE Transactions on Geoscience and Remote Sensing, 1996. [20] F. T. Ulaby, R. K. Moore, and A. K. Fung, Microwave Remote Sensing; Active and Passive, vol. III. Dedham, MA: Artech House Inc., 1986. [21] C. J. Oliver, “Information from SAR Images,” Journal of Physics D: Applied Physics, vol. 24, pp. 1493-1514, 1991. [22] P. R. Vieira, “Desenvolvimento de Classificadores de máxima verossimilhança pontuais e para imagens de abertura sintética,”: Instituto Nacional de Pesquisas Espaciais, São José dos Campos, Brasil, 1996. [23] A. J. Luckman, A. C. Frery, C. C. F. Yanasse, and G. B. Groom, “Texture in Airborne SAR Imagery of Tropical Forest and its Relationship to Forest Regeneration Stage,” International Journal of Remote Sensing, 1996. [24] L. L. Hess, J. M. Melack, S. Filoso, and Y. Wang, “Delinieation of inundated area and vegetation along the amazon floodplain with the SIR-C synthetic aperture radar,” IEEE Transactions on Geoscience and Remote Sensing, vol. 33, pp. 896-904, 1996. [25] J. A. Richards, P. W. Woodgate, and A. K. Skidmore, “An explanation of enhanced radar backscattering from flooded forests,” International Journal of Remote Sensing, vol. 8, pp. 1093-1100, 1987. [26] M. L. Imhoff, “Radar Backscatter and Biomass Saturation: Ramifications for Global Biomass Inventory,” IEEE Transactions on Geoscience and Remote Sensing, vol. 33, pp. 511-518, 1995.