Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the...

21
International Journal of Remote Sensing Vol. 33, No. 3, 10 February 2012, 710–729 Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China QI-SHENG HE†‡§, CHUN-XIANG CAO*, ER-XUE CHEN, GUO-QING SUN, FEI-LONG LING, YONG PANG, HAO ZHANG, WEN-JIAN NI†§, MIN XU†§, ZENG-YUAN LIand XIAO-WEN LIState Key Laboratory of Remote Sensing Science, Jointly Sponsored by the Institute of Remote Sensing Applications of the Chinese Academy of Sciences and Beijing Normal University, Beijing 100101, PR China Department of Geographical Information Sciences, School of Earth Sciences and Engineering, Hohai University, Nanjing 210098, PR China §Graduate School of the Chinese Academy of Sciences, Beijing 100049, PR China ¶Institute of Forest Resource Information Techniques, Chinese Academy of Forestry, Beijing 100091, PR China Studies are needed to evaluate the ability of Advanced Land Observing Satellite (ALOS) Phased Array type L-band Synthetic Aperture Radar (PALSAR) for for- est aboveground biomass (AGB) extraction in mountainous areas. In this article, forest biomass was estimated at plot and stand levels, and different biomass grades, respectively. Light detection and ranging (LiDAR) data with about one hit per m 2 were first used for forest biomass estimation at the plot level, with R 2 of 0.77. Then the LiDAR-derived biomass, as prior knowledge, was used to investigate the rela- tionship between ALOS PALSAR data and biomass. The results showed that at each biomass level, the range of the back-scatter coefficient in HH and HV polar- ization (where H and V represent horizontal and vertical polarizations, respectively, and the first of the two letters refers to the transmission polarization and the second to the received polarization) was very large and there was no obvious relationship between the synthetic aperture radar (SAR) back-scatter coefficient and biomass at plot level. At stand level and in different biomass grades, the back-scatter coef- ficient increased with the increase of forest biomass, and a logarithm equation can be used to describe the relationship. The main reason may be that forest structure is complex at the plot level, while the average value could partly decrease the influ- ence of forest structure at stand level. Meanwhile, terrain radiometric correction (TRC) was investigated and found effective for forest biomass estimation. 1. Introduction Forest biomass is an essential factor in environmental and climate modelling. Also, standing forest biomass forms an essential part of active carbon pool participation in the global carbon cycle. Mapping the amount and geographic distribution of forest biomass and its change with time is important for understanding the development of the carbon cycle. Light detection and ranging (LiDAR) has been used successfully for forest parame- ter estimation (Nilsson 1996, Lefsky et al . 2002, Hyde et al . 2005). LiDAR data have *Corresponding author. Email: [email protected] International Journal of Remote Sensing ISSN 0143-1161 print/ISSN 1366-5901 online © 2012 Taylor & Francis http://www.tandf.co.uk/journals http://dx.doi.org/10.1080/01431161.2011.577829

Transcript of Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the...

Page 1: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

International Journal of Remote SensingVol. 33, No. 3, 10 February 2012, 710–729

Forest stand biomass estimation using ALOS PALSAR data based onLiDAR-derived prior knowledge in the Qilian Mountain, western China

QI-SHENG HE†‡§, CHUN-XIANG CAO*†, ER-XUE CHEN¶, GUO-QINGSUN†, FEI-LONG LING¶, YONG PANG¶, HAO ZHANG†, WEN-JIAN NI†§,

MIN XU†§, ZENG-YUAN LI¶ and XIAO-WEN LI††State Key Laboratory of Remote Sensing Science, Jointly Sponsored by the Institute ofRemote Sensing Applications of the Chinese Academy of Sciences and Beijing Normal

University, Beijing 100101, PR China‡Department of Geographical Information Sciences, School of Earth Sciences and

Engineering, Hohai University, Nanjing 210098, PR China§Graduate School of the Chinese Academy of Sciences, Beijing 100049, PR China

¶Institute of Forest Resource Information Techniques, Chinese Academy of Forestry, Beijing100091, PR China

Studies are needed to evaluate the ability of Advanced Land Observing Satellite(ALOS) Phased Array type L-band Synthetic Aperture Radar (PALSAR) for for-est aboveground biomass (AGB) extraction in mountainous areas. In this article,forest biomass was estimated at plot and stand levels, and different biomass grades,respectively. Light detection and ranging (LiDAR) data with about one hit per m2

were first used for forest biomass estimation at the plot level, with R2 of 0.77. Thenthe LiDAR-derived biomass, as prior knowledge, was used to investigate the rela-tionship between ALOS PALSAR data and biomass. The results showed that ateach biomass level, the range of the back-scatter coefficient in HH and HV polar-ization (where H and V represent horizontal and vertical polarizations, respectively,and the first of the two letters refers to the transmission polarization and the secondto the received polarization) was very large and there was no obvious relationshipbetween the synthetic aperture radar (SAR) back-scatter coefficient and biomassat plot level. At stand level and in different biomass grades, the back-scatter coef-ficient increased with the increase of forest biomass, and a logarithm equation canbe used to describe the relationship. The main reason may be that forest structureis complex at the plot level, while the average value could partly decrease the influ-ence of forest structure at stand level. Meanwhile, terrain radiometric correction(TRC) was investigated and found effective for forest biomass estimation.

1. Introduction

Forest biomass is an essential factor in environmental and climate modelling. Also,standing forest biomass forms an essential part of active carbon pool participation inthe global carbon cycle. Mapping the amount and geographic distribution of forestbiomass and its change with time is important for understanding the development ofthe carbon cycle.

Light detection and ranging (LiDAR) has been used successfully for forest parame-ter estimation (Nilsson 1996, Lefsky et al. 2002, Hyde et al. 2005). LiDAR data have

*Corresponding author. Email: [email protected]

International Journal of Remote SensingISSN 0143-1161 print/ISSN 1366-5901 online © 2012 Taylor & Francis

http://www.tandf.co.uk/journalshttp://dx.doi.org/10.1080/01431161.2011.577829

Page 2: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Prior knowledge-based retrieval from remote-sensing data 711

been successfully used to estimate Douglas fir/western hemlock biomass (Lefsky et al.1999a), temperate mixed deciduous forest biomass (Lefsky et al. 1999b), tropical for-est biomass (Drake et al. 2002, 2003), hardwood forest biomass (Lim et al. 2003),conifer forest biomass (Popescu et al. 2004) and so on (Nelson et al. 1997, Meanset al. 2000, Lim and Treitz, 2004). The regression method has been used for LiDAR-derived biomass by several workers (Nelson et al. 1988, 2004, Naesset 1997, Lucaset al. 2006). The regression model was different for each special forest type and area,so the regression method had to be established for different forest types and areas.Because LiDAR can directly measure components of vegetation canopy structure,such as canopy height, it is a promising approach for forest biomass estimation.

On the other hand, forest biomass mapping from synthetic aperture radar (SAR)data has become one of the most promising applications of radar remote sensing tovegetation studies. Radar data have important roles in forest aboveground biomass(AGB) estimation, especially in areas with frequently cloudy conditions (Lu 2006).Many studies have shown that it is possible to retrieve forest biomass using SAR data.The relationship between SAR back-scatter and forest AGB or forest volume has beeninvestigated in various forest types (Dobson et al. 1992, Le Toan et al. 1992, Luckmanet al. 1997, Kasischke et al. 1997). Generally speaking, the correlation between forestAGB or volume and SAR back-scatter for a single polarization/wavelength has beenfound to have a saturation point beyond which the back-scatter no longer increaseswith AGB or volume. The level of saturation has been found to vary as a functionof the wavelength and polarization of the SAR sensor and with different forest types.Until now, different saturation levels have been reported from 60 to 600 t ha−1 forthe L band (Rauste et al. 1994, Imhoff 1995, Luckman et al. 1998, Fransson andIsraelsson 1999, Kurvonen et al. 1999, Paloscia et al. 1999, Watanabe et al. 2006). Thesaturation level may depend on the tree species and forest types as well as the groundsurface type. Also, different R2 values (between biomass and back-scatter power) havebeen reported (Kuplich et al. 2000, Jenet et al. 2003), and multi-temporal SAR datacan increase the R2 value (Yrjo 2005).

These studies concentrated on relatively flat areas, where terrain effects were notsignificant. Studies should be extended to evaluate the ability of SAR to extract forestbiomass in mountainous areas, and the terrain radiometric correction (TRC) shouldbe evaluated for forest biomass extraction (Castel et al. 2001).

From the above studies on the relationship of L-band back-scatter and forestbiomass, it can be seen that the saturation point and the correlation coefficient vary fordifferent forest types and ground surface types. Furthermore, seasonal effects in thisrelationship have also been observed, although these observations are rather occa-sional (Rauste 2005). The LiDAR data, as prior knowledge, supply an opportunity toanalyse the relationship of the L-band back-scatter and forest biomass at plot andstand levels. In this article, the LiDAR data with about one hit per m2 were firstused for forest biomass estimation at plot level. Then the LiDAR-derived biomass,as prior knowledge, was used to investigate the relationship between Advanced LandObserving Satellite (ALOS) Phased Array type L-band Synthetic Aperture Radar(PALSAR) data and forest AGB. As is known, field surveying is very difficult, espe-cially at the stand level, so the LiDAR-derived biomass at plot level supplied anopportunity to achieve relatively high accuracy in stand biomass estimation. The aimsof this study were:

1. to investigate the ability of airborne LiDAR with about one hit per m2 forextracting forest biomass in mountain areas,

Page 3: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

712 Q.-S. He et al.

2. to estimate the relationship of the ALOS PALSAR L-band back-scatter andforest AGB in mountainous areas at plot and stand levels,

3. to analyse whether the TRC can increase the biomass inversion accuracy ofALOS PALSAR data and

4. to study the potential of combining multi-temporal L-band SAR data for forestAGB mapping.

2. Site description and data

The study site of Dayekou is situated in the Qilian Mountain area, with its geographiccoordinates ranging from 38◦ 29′ to 38◦ 35′ N latitude and from 100◦ 12′ to 100◦20′ E longitude, in Gansu province, western China (figure 1). The elevation varies from2500 to 3800 m above sea level. The slope of the forest area is mostly greater than 30◦.The area has a temperate continental mountainous climate. In winter, the atmosphericcirculation is controlled by the Mongolian anticyclone; the conditions are cold anddry, with little precipitation. When the atmospheric circulation is controlled by thecontinental cyclone in the summer, the diurnal difference of temperature is dramatic.The difference in precipitation between summer and winter is large, and annual pre-cipitation takes place mainly in the summer. Influenced by the climate and the terrain,the prevalent vegetation types in the study area are mountainous pastures and forests.The dominant vegetation includes Picea crassifolia, Sabina przewalskii and grassland.Vegetation density varies with terrain, soil, water and climate factors (Zhou et al.2007). The coverage of pure forest stands of P. crassifolia occupies about 95% of thewhole forest land over the test site, so this article only focuses on this forest type. Theforest structure is complex owing to the presence of stands at various stages of regen-eration and degradation, although these are often dominated by only P. crassifolia.

One regional remote-sensing campaign was carried out in this area in June 2008.The airborne sensor data acquired were small footprint LiDAR and high resolu-tion colour charge coupled device (CCD). The space-borne sensor data were ALOSPALSAR (HH and HV, where H and V represent horizontal and vertical polarizations,respectively, and the first of the two letters refers to the transmission polarization andthe second to the received polarization), Système Probatoire d’Observation de la Terre(SPOT-5) and Advanced Spaceborne Thermal Emission and Reflection Radiometer(ASTER). An ortho-rectified CCD image was produced using a digital elevationmodel (DEM) produced from the LiDAR data. The ALOS PALSAR level 1.1 datawere acquired over the whole research site in 2007 and 2008. The image and systemcharacteristics are summarized in table 1.

LiDAR data were acquired on 23 June 2008 using a laser scanner Riegl LMS-Q560and a Litemapper 5600 system (RIEGL Inc., Horn, Austria), operating at a flightaltitude of 800 m, configured to acquire data using a narrow scan angle of <0.5 mradeither side of nadir and with a point density of about 1 return m−2 to maximize canopypenetration and minimize any potential scan angle effect. The x, y, z position (easting,northing and elevation) and intensity of each pulse were supplied for the first and lastpulse and geo-referenced to a World Geodical System (WGS) 84, Universal TransverseMercator (UTM) north 48 projection system. The accuracy report that accompaniedthe LiDAR data indicated the accuracy in the x–y position as 0.10 m and in the z posi-tion as 0.03 m. Like most discrete return LiDAR systems, the Riegl LMS-Q560 recordsintensity for each pulse in the near infrared (1550 nm) region. The intensity of eachreturn pulse, sometimes referred to as laser amplitude, represents the reflected energyfrom a highly culminated beam of light (footprint of 0.2 m if the sensor’s operating

Page 4: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Prior knowledge-based retrieval from remote-sensing data 713

China(a)

(b)

(c)

Gansu Province

0

0 80 160

S

EW

N

320 480

100° 14′ E

38

° 3

4′ N

38

° 3

2′ N

38

° 3

0′ N

38

° 34′ N

38

° 32′ N

38

° 30′ N

100° 16′ E 100° 18′ E

100° 14′ E0 1 2 3 4 5

km

100° 12′ E 100° 16′ E 100° 18′ E

640km

290 580 1160

N

E

S

W

1740 2320km

Figure 1. Location of the study area. (a) Map of China, in which the yellow area is the loca-tion of Gansu province; (b) map of Gansu, in which the red square area is the location of theDayekou area; and (c) charge coupled device (CCD) image of Dayekou, in which the red flagsare the forest sample plots, and there is no data in the black area.

Page 5: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

714 Q.-S. He et al.

Table 1. PALSAR data over the research site and some SAR system parameters.

Imaging Polarization Azimuth/range (m) Incidence (◦) Orbit

10 August 2007 HH, HV 3.20/9.37 38.7 Ascending12 May 2008 HH, HV 3.20/9.37 38.7 Ascending27 June 2008 HH, HV 3.20/9.37 38.7 Ascending

height is 800 m). The LiDAR intensity provides a concentrated measurement of theobject’s reflectivity, unaffected by shadows or occlusions.

In order to investigate the capability of PALSAR data for forest biomass estimation,ground truth data were collected through field work from July to August 2007 andfrom June to July 2008. The high resolution airborne CCD image mosaic was usedto identify each forest stand through manual interpretation. Some forest stands wereselected for field plot data measurement. The size of sample plots was limited to20 × 20 m or 25 × 25 m. The height (H) and diameter at breast height (DBH) of1.3 m of each individual tree of each plot were measured and defined as ground truthdata at plot level. The location of the sample plot was positioned using a differentialglobal positioning system (GPS). Among all the plots surveyed, only the plots that canbe thought of as pure P. crassifolia forest plots were used for this study; the total plotnumber selected in this way was 95. Furthermore, only the plots whose location wasfixed by the differential GPS were selected. Finally, 83 plots were left for this study.Although the field data were not in agreement with the SAR data, for the P. crassifo-lia forest, the growth was so slow that the change of forest biomass can be neglected.So the field data were directly used to calculate the biomass with no growth correction.

3. Method

Figure 2 shows the main method used in this study illustrated as a flow chart. First,the ability of LiDAR data with about one hit per m2 for forest biomass estimation inmountain areas was estimated, and a biomass distribution map of high resolution was

LiDAR data

Biomass estimation

Biomass inversion

Relationship betweenSAR and AGB

Plot level Stand level

Result and discussion

Biomass Multi-temporal

CCD

Forest plot and stand

PALSAR

GTC and TRC

Figure 2. The main method used in this study, illustrated as a flow chart.

Page 6: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Prior knowledge-based retrieval from remote-sensing data 715

derived. Using this LiDAR-derived biomass map, the relationship between the forestbiomass of P. crassifolia forest and the PALSAR back-scattering coefficient can bediscussed at plot level, stand level and biomass grade. Meanwhile, the effect of TRCof SAR data on the relationship between forest biomass and PALSAR data was alsoexplored at both levels. Finally, the usefulness of multi-temporal PALSAR SAR wasinvestigated for forest biomass mapping.

3.1 Forest biomass calculation

Biomass of all the aboveground organs of one P. crassifolia tree were estimated, basedon its DBH and height (H), using empirically relative growth equations (equations(1)–(4)), following Wang et al. (1998):

stock biomass = 0.0478 × ((DBH)2 × H)0.8665, (1)

branch biomass = 0.0061 × ((DBH)2 × H)0.8905, (2)

leaf biomass = 0.2650 × ((DBH)2 × H)0.4701, (3)

fruit biomass = 0.0342 × ((DBH)2 × H)0.5779. (4)

Then, the AGB of each individual tree was obtained by summing the stock, branch,leaf and fruit biomass. By summing the biomass of the individual trees of each plot, itwas possible to obtain the plot biomass (t ha−1).

3.2 LiDAR data processing

The first step involved the elimination of pulses identified as below the nominal groundsurface or above the expected canopy height. The remaining pulses were dividedinto those that reached the ground surface and those that did not. Ground hitswere removed using Terrascan and MicroStation software (Terrasolid Ltd., Helsinki,Finland). This algorithm identified ground hits based on iterative slope analysis ofLiDAR returns (Axelsson 1999). Grid cell size and maximum slope of the area wererequired input parameters. The grid cell size is the smallest cell size for which a groundreturn can be extracted. In this article, grid cell size and maximum slope of the areawere set as 30 m and 88%, respectively. Then ground points were used to generate aground DEM (Kraus and Pfeifer 1998). First, a triangulated irregular network (TIN)was constructed for the point cloud, based on a Delaunay triangulation of its eleva-tion data. Then a rectangular grid of pixels was extracted from each TIN using linearinterpolation with a constant sampling interval of one metre. Finally, the raster DEMof 0.5 × 0.5 m spatial resolution was generated. To assess the accuracy of extractedground hits, the DEM created by ground hits was compared with the points measuredwith an electronic total station. Compared with 1546 field measurements, the extremesof their difference are −95.1 and +94.8 cm, the mean value is −21.8 cm and the rootmean square error (RMSE) is ±22.7 cm.

Non-ground hits, designated as vegetation hits, were normalized for varying terrainelevations, thereby enabling volume and biomass models to incorporate actual LiDARpoint heights (Means et al. 2000). This was done by calculating the actual return heightabove a LiDAR-derived 0.5 m DEM of the study area. The actual height of each veg-etation hit was calculated as the difference between the vegetation hit and the bilinear

Page 7: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

716 Q.-S. He et al.

interpolated height of the four corner cells of the DEM cell directly beneath each hit.Typically, a small shrub can grow up to about 1.3 m in height, and so only returns withan actual height above this threshold were associated with woody vegetation.

In this article, height quantiles of the cumulative percentage were selected as thestatistical variable (Naesset 2004). Finally, 19 vegetation height classes (based on 5%intervals of the cumulative percentage values) were calculated.

Crown cover (CC) was selected as another statistical variable. To generate equivalentCC estimates from LiDAR data, returns greater than 1.3 m in height were consideredas tree crown elements. All points were interpolated to a raster image. When the gridunit had multiple echoes, the maximum value was selected as the interpolation value.According to the point cloud density, the digital surface model (DSM) was interpo-lated into the resolution of 0.5 m. A canopy height model (CHM) was obtained toindicate the difference between the DSM and the DEM. A 1.3 m height thresholdwas used to conform to definitions of forest cover, with all 0.5 × 0.5 pixels abovethis threshold coded as either 1 or 0. For each field plot, the canopy cover percent-age (CC%) was calculated as the sum of all cells with a value of 1 as a percentage ofthe total.

The other statistical variables were the mean, variance, skewness, kurtosis, meanabsolute deviation and standard deviation contained in an N-element vector X; j isthe jth number in vector X. When X = (x0, x1, x2, . . . , xN−1), the various parametersare defined as follows:

mean = x = 1N

N−1∑j = 0

xj, (5)

variance = 1N − 1

N−1∑j = 0

(xj − x)2, (6)

skewness = 1N

N−1∑j = 0

(xj − x√

(variance)

)3

, (7)

kurtosis = 1N

N−1∑j = 0

(xj − x√

(variance)

)4

− 3, (8)

mean absolute deviation = 1N

N−1∑j = 0

∣∣xj − x∣∣, (9)

standard deviation = √(variance). (10)

Finally, these variables were selected for biomass estimation.To integrate the information on vegetation height of different cumulative per-

centages, CC and mean, variance, skewness, kurtosis, mean absolute deviation andstandard deviation, both forward and backward stepwise linear regressions wereevaluated for 83 plots.

3.3 SAR data processing

All the PALSAR data of this test site were acquired as level 1.1 products, so all thefollowing processing steps, such as radiometric calibration, multi-looking, geocodedterrain correction (GTC) and TRC, were started from single look complex (SLC) data.

Page 8: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Prior knowledge-based retrieval from remote-sensing data 717

For rectifying the PALSAR data, the LiDAR-derived DEM was resampled to 20 ×20 m in pixel size.

3.3.1 Radiometric calibration. The σ 0 radiometric calibration function provided bythe Japan Aerospace Exploration Agency (JAXA) is as follows:

σ 0 = 10 log10(I2 + Q2) + G − C, (11)

where I and Q are the real and imaginary parts of the complex SAR image pixel value;G and C are calibration constants, G = −83.2 for HH polarization and −80.2 forHV polarization, C = 32.0. Here we define K = G – C, and will not get the deci-bel (dB) value of the radiometric calibrated intensity value before analysing the data.So equation (12) is used as the first step of the data processing:

σ 01 = (I2 + Q2) · 10

K10 , (12)

where K = −115.2 for HH polarization and −112.2 for HV polarization, the subscriptof σ 0 means the σ 0 calibration result of each SLC pixel is in the intensity domain,not dB. For dual polarization PALSAR data, after this calibration, only the intensityimage of each polarization is given.

3.3.2 Multi-looking. Applying multi-look processing to the single look σ 01 image is

reasonable for both depressing speckle noise and reducing the size of the SAR data.We multi-looked the σ 0

1 image with 10 looks in the azimuth direction and two looksin the range direction to produce one image of pixel size about 30 m in azimuth andrange directions. Taking the dual polarization data listed in table 1 as an example, theresulting image is denoted as σ 0

1_ml.The pixel size of the DEM used for SAR image simulation was 20 × 20 m. After

GTC and TRC processing, σ 01_ml is resampled to the DEM’s coordinate space with the

same pixel size. The pixel size of σ 01_ml is about 30 × 30 m, which is a little bit bigger

than the pixel size of the DEM. The small difference in the pixel size of the DEM andthe multi-looked SAR image will benefit the procedure of DEM-based SAR simula-tion and automatic co-registration of the multi-looked σ 0

1_ml image and the simulatedmulti-looked SAR image because of the more recognized texture feature.

3.3.3 Geocoded terrain correction. A range-doppler (RD) based SAR imagingmodel was established for the geocoding of the PALSAR slant range image. All theRD model parameters were extracted from the SAR metadata file. GTC processingbased on the DEM and RD model was developed. The RD model and the DEM wereused to generate one simulated SAR image corresponding to the real SAR image σ 0

1_ml.The simulated SAR image is very similar to the real SAR image because of image tex-ture caused by the terrain variance (layover, shadow and foreshortening). Throughautomatic image to image registration, the simulated SAR image can be registeredto the real SAR image with sub-pixel accuracy. The GTC-produced SAR image isdenoted by σ 0

1_GTC.

Page 9: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

718 Q.-S. He et al.

3.3.4 Terrain radiometric correction. TRC is used to correct the effect of differ-ent effective back-scattering surface areas caused by the local topography and SARimagery geometry. In a well-calibrated SAR image, the same forest stand of pixelsize 20 × 20 m in different locations of a hilly region should have a similar back-scattering coefficient level no matter where it is located. But in reality, the effectiveback-scattering area has a strong effect on the total back-scattered power. TRC aims tocorrect this kind of effect by multiplying a TRC correction factor (F) with the σ 0

1_GTCimage (equation (13)). This has already been dubbed ‘the slope correction’ by severalresearchers (Ulander 1996, Shimada and Hirosawa 2000).

σ 0TRC = σ 0

1_GTCF . (13)

The GTC and TRC images of three PALSAR image scenes are shown in figure 3.

3.4 Regression methods

The conventional multivariate linear regression model can be expressed as follows:

Y = X0 + a1X1 + . . . aiXi, (14)

where Y is the dependent parameter to be predicted, X0 is the intercept, i is the numberof independent variables and a1, . . .,i and X1, . . .,i are the regression coefficients andvalues of independent variables, respectively.

In this study, Y refers to the forest AGB and X1,. . .,i are the statistical variables fromairborne low density LiDAR data. Considering that there are 26 possible indepen-dent variables which could be used in the regression based on LiDAR estimates (see§3.2), a backward stepwise linear regression method was used to select the most signif-icant variables (probability of F-to-remove = 0.1). The backward stepwise regressionmethod starts with the maximum model, eliminates variables one by one, and endswhen no more variables can be removed from the model at significance level 0.1. Notethat the F-test criterion was used in this procedure. The forward stepwise linear regres-sion method starts with an ‘empty’ model with no explanatory variables, and variablesare added one by one until the model cannot be improved significantly by addinganother variable.

4. Results

4.1 Biomass estimation from LiDAR

To integrate the information on vegetation height of different cumulative percentages,CC and mean, variance, skewness, kurtosis, mean absolute deviation and standarddeviation, both forward and backward stepwise linear regressions were evaluated for83 plots. Finally, five height variables (H15, H45, H70, H75 and H80, where the sub-scripts mean the cumulative percentage), CC, variance and mean absolute deviationwere found to be the most significant for estimating AGB and backward stepwise lin-ear regressions were established. The LiDAR-derived biomass is therefore describedby equation (14):

Biomass = − 8.179H15 + 18.686H45 + 17.429H70 − 24.851H75 + 9.405H80

+ 40.481(CC) + 3.543(Var) + 52.673(MAD) + 25.638,(15)

where Var means variance and MAD means mean absolute deviation.

Page 10: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Prior knowledge-based retrieval from remote-sensing data 719

100° 14′ E38°

34

′ N38°

32

′ N38°

30

′ N

38° 3

4′ N38° 3

2′ N38° 3

0′ N

100° 16′ E 100° 18′ E

(a) (b)

(c) (d)

(e) (f)

100° 14′ E 100° 16′ E 100° 18′ E

100° 14′ E

38°

34

′ N38°

32

′ N38°

30

′ N

38° 3

4′ N38° 3

2′ N38° 3

0′ N

100° 16′ E 100° 18′ E

100° 14′ E 100° 16′ E 100° 18′ E

100° 14′ E

38°

34

′ N38°

32

′ N38°

30

′ N

38° 3

4′ N38° 3

2′ N38° 3

0′ N

100° 16′ E 100° 18′ E

100° 14′ E 100° 16′ E 100° 18′ E

100° 14′ E

38°

34

′ N38°

32

′ N38°

30

′ N

38° 3

4′ N38° 3

2′ N38° 3

0′ N

100° 16′ E 100° 18′ E

100° 14′ E 100° 16′ E 100° 18′ E

100° 14′ E

38°

34

′ N38°

32

′ N38°

30

′ N

38° 3

4′ N38° 3

2′ N38° 3

0′ N

100° 16′ E 100° 18′ E

100° 14′ E 100° 16′ E 100° 18′ E

100° 14′ E

38°

34

′ N38°

32

′ N38°

30

′ N

38° 3

4′ N38° 3

2′ N38° 3

0′ N

100° 16′ E 100° 18′ E

100° 14′ E 100° 16′ E 100° 18′ E

0 1 2 3 4 5km

Figure 3. The GTC and TRC images of three scenes. (a) 20070810GTC, (b) 20070810TRC,(c) 20080512GTC, (d) 20080512TRC, (e) 20080627GTC and (f ) 20080627TRC. The eight digitsshow the date of acquisition in the order year, month, day and the three letters the type ofcorrection. The GTC image was produced using the DEM and RD model. The TRC image wasproduced by multiplying the GTC by the slope correction factor. In these images, the green areasrepresent forest. Compared with the GTC images, slope effects such as layover and shadow havebeen rectified in the TRC images.

Page 11: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

720 Q.-S. He et al.

Figure 4. Field-measured biomass versus LiDAR-predicted biomass using a stepwise linearregression for all plots combined (R2 = 0.77, standard error = 16.28 t ha−1).

100° 14′ E

38°

34

′ N

38° 3

4′ N38° 3

2′ N38° 3

0′ N

38°

32

′ N38°

30

′ N

100° 16′ E 100° 18′ E

100° 14′ E 100° 16′ E 100° 18′ E

0 230 t ha–10 2 3 4 5

km

Figure 5. The biomass inversion map by LiDAR data. The biomass ranges from 0 to230 t ha−1.

The use of equation (14) for estimating biomass was considered sufficiently robust(R2 = 0.77, standard error = 16.28 t ha−1, n = 83; figure 4). Through this equa-tion, fully automated procedures were applied across the entire area to estimate forestbiomass. The biomass inversion map is given in figure 5. Then the LiDAR-derivedbiomass, as a prior knowledge, was used to investigate the relationship between ALOSPALSAR data and biomass at the plot and stand levels.

Page 12: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Prior knowledge-based retrieval from remote-sensing data 721

4.2 Relationships between SAR back-scattering coefficient and AGB at plot level

First, about 1800 LiDAR AGB grid data were selected for analysis. The results areshown in figure 3 for the SAR data from 10 August 2007. One reason for taking thisimage as an example is that its relationship with forest AGB is the highest. Anotherreason is that the changing trend of the back-scattering coefficient with forest AGB isalmost the same for all the three scenes of the SAR image. The relationship betweenSAR back-scatter and LiDAR-derived biomass is shown in figure 6. Also, the rela-tionship between SAR back-scatter and the field-measured biomass is shown in figure7. From figures 6 and 7, it can be seen that at each biomass level, the data range ofthe back-scatter coefficient in HH and HV polarization is very large. There is no obvi-ous relationship between the SAR back-scatter coefficient and the LiDAR-derivedbiomass. The reason may be that in this study area, the forest structure is complex.At the same biomass level, the tree density can be very different, which affects theSAR signal.

The TRC was also analysed. From figure 8, it can be seen that before TRC, theSAR back-scattering coefficient decreases with the local incidence angle in both HHand HV polarization. But after TRC, the trend becomes unclear, especially for HV

250(a) (b)

(c) (d)

200

150

100

50

0

–18

–22 –20 –18 –16 –14 –12 –10 –8 –6 –22–24 –20 –18 –16 –14 –12 –10 –8 –6

–16 –14 –12

HH (GTC) back-scatter (dB)

Bio

mass (

t ha

–1)

–10 –8 –6 –4 –2 0

250

200

150

100

50

0

–18 –16 –14 –12

HH (TRC) back-scatter (dB)

Bio

mass (

t ha

–1)

–10 –8 –6 –4 –2 0

250

200

150

100

50

0

HV (GTC) back-scatter (dB)

Bio

mass (

t ha

–1)

250

200

150

100

50

0

HV (TRC) back-scatter (dB)

Bio

mass (

t ha

–1)

Figure 6. Scatter diagrams between SAR back-scatter and LiDAR-derived biomass at plotlevel. (a) Relationship between SAR back-scatter of HH polarization and LiDAR-derivedbiomass for the GTC image; (b) relationship between SAR back-scatter of HH polarizationand LiDAR-derived biomass for the TRC image; (c) relationship between SAR back-scatter ofHV polarization and LiDAR-derived biomass for the GTC image; and (d) relationship betweenSAR back-scatter of HV polarization and LiDAR-derived biomass for the TRC image.

Page 13: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

722 Q.-S. He et al.

200(a) (b)

(c) (d)

150

100

50

0–12

–20 –18 –16 –14 –12 –10 –18 –16 –14 –12 –10–8

–10 –8 –6 –4 –2 0

HH (GTC) back-scatter (dB)

–10 –8 –6 –4 –2 0

HH (TRC) back-scatter (dB)

HV (GTC) back-scatter (dB) HV (TRC) back-scatter (dB)

Bio

mass (

t ha

–1)

200

150

100

50

0

Bio

mass (

t ha

–1)

200

150

100

50

0

Bio

mass (

t ha

–1)

200

150

100

50

0

Bio

mass (

t ha

–1)

Figure 7. Scatter diagrams between SAR back-scatter and field-measured biomass at plotlevel. (a) Relationship between SAR back-scatter of HH polarization and field-measuredbiomass for the GTC image; (b) relationship between SAR back-scatter of HH polarizationand field-measured biomass for the TRC image; (c) relationship between SAR back-scatter ofHV polarization and field-measured biomass for the GTC image; and (d) relationship betweenSAR back-scatter of HV polarization and field-measured biomass for the TRC image.

polarization. So the TRC can be considered effective. Although at each biomass levelthe range of the back-scatter coefficient in HH and HV polarization is very large, fromthe total trend, there is an increase of the back-scattering coefficient with increase inforest biomass.

4.3 Relationships between SAR back-scattering coefficient and AGB at stand level

A forest stand represents a group of trees occupying a given area and sufficientlyuniform in species composition, structure, site quality and condition so as to be dis-tinguishable from the forest in adjoining areas. Figure 9 shows the sketch map of theforest plot and stand. The thin red square represents the forest plot, with a size ofabout 20 × 20 m, while the thick red polygon represents the forest stand. In this study,by image segment and visual interpretation of the high resolution CCD image, a totalof 43 forest stands were created for analysis. The results are shown in figure 10 for theSAR data from 10 August 2007. The reason that this image was taken as an example isalso because the relationship between the SAR back-scattering coefficient and forestbiomass is highest and the trend is almost the same as in the two other scenes. Anothertwo scenes of PALSAR data were also analysed, with the results shown in table 2, but

Page 14: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Prior knowledge-based retrieval from remote-sensing data 723

HH

(G

TC

) back-s

ca

tter

(dB

)H

V (

GTC

) back-s

ca

tter

(dB

)

HH

(T

RC

) back-s

ca

tter

(dB

)H

V (

TR

C)

back-s

ca

tter

(dB

)

0(a) (b)

(c) (d)

–2

–4

–6

–8

–10

–12

–14

–16

–18

–6

–8

–10

–12

–14

–16

–18

–20

–22

–8

–10

–12

–14

–16

–18

–20

–22

–24

0

–2

–4

–6

–8

–10

–12

–14

–16

–180 15 30

Local incidence angle (°)

45 60 75 90

0 15 30

Local incidence angle (°)

45 60 75 90 0 15 30

Local incidence angle (°)

45 60 75 90

0 15 30

Local incidence angle (°)

45 60 75 90

Figure 8. Scatter diagrams between SAR back-scatter and the local incidence angle at plotlevel. (a) Relationship between SAR back-scatter of HH polarization and local incidenceangle for the GTC image; (b) relationship between SAR back-scatter of HH polarization andlocal incidence angle for the TRC image; (c) relationship between SAR back-scatter of HVpolarization and local incidence angle for the GTC image; and (d) relationship between SARback-scatter of HV polarization and local incidence angle for the TRC image.

the exponential function and the scattering figures are not shown in figure 10 due tospace limitations.

From figure 10, it can be seen that the scattering coefficient increased with theincrease of biomass. The exponential function was suitable for this analysis. Table 2shows the correlation coefficients of three scenes of PALSAR data. From table 2, itcan be seen from the correlation coefficient that HV polarization was better than HHpolarization, and that the TRC image was better than the GTC image, for biomassestimation in the three images, so the TRC method applied in this article can increasethe accuracy for forest AGB estimation at the stand level.

4.4 Relationships between the SAR back-scattering coefficient and AGB in differentbiomass grades

Average back-scatter coefficients were calculated in different biomass grades, that is,20–25, 25–30, 30–35 t ha−1 and so on. The results are shown in figure 11. It can be seenthat back-scatter coefficients in HV polarization increase with increase of biomass.It can also be seen that in HV polarization the saturation was about 60 t ha−1. Theresult in HV polarization may be in agreement with other research (Imhoff 1995).The reason may be that the average value could partly decrease the influence of forest

Page 15: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

724 Q.-S. He et al.

100° 13′ 10″ E

38°

32

′ 50

″ N

100° 13′ 20″ E0 0.1 0.2

km

100° 13′ 30″ E 100° 13′ 40″ E

100° 13′ 10″ E

38°

33

′ N

38° 3

2′ 50″ N

38° 3

3′ N100° 13′ 20″ E 100° 13′ 30″ E 100° 13′ 40″ E

Figure 9. Sketch map of the forest plot and stand. The thin red square represents the forestplot, with a size of about 20 × 20 m, while the thick red polygon represents the forest stand.

Table 2. The correlation coefficients of three scenes ofPALSAR data.

HH HV

070810GTC 0.546 0.687070810TRC 0.713 0.789080512GTC 0.532 0.646080512TRC 0.667 0.729080627GTC 0.590 0.732080627TRC 0.688 0.770

Note: The six digits show the date of acquisition in theorder year, month, day and the three letters the type ofcorrection.

structure at stand level. For HH polarization, the trend almost agreed with the HVpolarization, but the point distribution was not continuous.

4.5 Multi-temporal regression model

From the above analysis, the exponential function is shown to be suitable for analysingthe relationship between the SAR back-scattering coefficient and forest biomass.In order to analyse whether multi-temporal SAR data were useful, ‘ln(biomass)’ wasselected as a dependent variable, while the independent variables were the scatteringcoefficients of different images in different polarization. In this article, backward step-wise linear regressions were used for this analysis. The results are shown in table 3,

Page 16: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Prior knowledge-based retrieval from remote-sensing data 725

200(a) (b)

(c) (d)

y = 231.1e0.145x

R2 = 0.546

y = 335.6e0.196x

R2 = 0.713

y = 574.6e0.156x

R2 = 0.687

y = 790.5e0.184x

R2 = 0.789

150

100

50

0–16

–22 –20 –18 –16 –14 –12 –10 –8 –6 –22 –20 –18 –16 –14 –12 –10 –8 –6

–14 –12 –10 –8 –6 –4 –2 0

HH (GTC) back-scatter (dB)

HV (GTC) back-scatter (dB)

–16 –14 –12 –10 –8 –6 –4 –2 0

HV (TRC) back-scatter (dB)

HH (TRC) back-scatter (dB)

Bio

mass (

t ha

–1)

200

150

100

50

0

Bio

mass (

t ha

–1)

200

150

100

50

0

Bio

mass (

t ha

–1)

200

150

100

50

0

Bio

mass (

t ha

–1)

Figure 10. Relationship between the SAR back-scattering coefficient and the LiDAR-derivedAGB at stand level. (a) Relationship between the SAR back-scattering coefficient of HH polar-ization and LiDAR-derived biomass for the GTC image; (b) relationship between the SARback-scattering coefficient of HH polarization and LiDAR-derived biomass for the TRC image;(c) relationship between the SAR back-scattering coefficient of HV polarization and LiDAR-derived biomass for the GTC image; and (d) relationship between the SAR back-scatteringcoefficient of HV polarization and LiDAR-derived biomass for the TRC image.

200(a) (b)

180

160

140

120

100

80

60

40

20

0–10 –16 –15 –14 –13 –12 –11 –10–9 –8 –7 –6 –5

Bio

mass (

t ha

–1)

200

180

160

140

120

100

80

60

40

20

0

Bio

mass (

t ha

–1)

HH (TRC) back-scatter (dB) HV (TRC) back-scatter (dB)

Figure 11. Relationship between SAR back-scatter and LiDAR-derived biomass in differentbiomass grades. (a) Relationship between SAR back-scatter of HH polarization and LiDAR-derived biomass for the TRC image in different biomass grades and (b) relationship betweenSAR back-scatter of HV polarization and LiDAR-derived biomass for the TRC image indifferent biomass grades.

Page 17: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

726 Q.-S. He et al.

Table 3. Correlation coefficients between ln(biomass) and PALSAR data.

R R2 Adjusted R2 Model

070810HH 0.845 0.713 0.706 5.816 + 0.196 × 070810HH070810HV 0.888 0.789 0.784 6.673 + 0.185 × 070810HV080512HH 0.817 0.667 0.659 5.653 + 0.181 × 080512HH080512HV 0.854 0.729 0.722 6.479 + 0.175 × 080512HV080627HH 0.856 0.733 0.726 5.707 + 0.184 × 080627HH080627HV 0.878 0.770 0.765 6.547 + 0.176 × 080627HV070810HH,

070810HV,080512HH

0.918 0.842 0.830 ∗7.212 + 0.197 × 070810HH+ 0.274 × 070810HV– 0.282 × 080512HH

Notes: The six digits show the date of acquisition in the order year, month, day and the twoletters the type of polarization.∗The equation was obtained through backward stepwise linear regressions.

from which it can be seen that multi-temporal SAR data can increase the biomassestimation accuracy.

5. Discussion

The reason for poor performance at the plot level may be the complexity of the foreststructure and the terrain correction effect. Through natural growth, trees will not beof an even height in a forest. Although terrain correction (the TRC image) can correctthe effect of different back-scattering surface areas caused by the local topographyand SAR imagery geometry, because of the complexity of the forest structure, thecorrection may be inadequate.

From figures 10 and 11, SAR back-scatter did not correlate well with high biomass,which means that these results are in agreement with those of several other work-ers in having proved that the correlation between forest biomass and back-scatterreaches a saturation point beyond which the back-scatter no longer increases withbiomass or volume (Rauste et al. 1994, Imhoff 1995, Luckman et al. 1998, Franssonand Israelsson 1999, Kurvonen et al. 1999, Paloscia et al. 1999, Watanabe et al. 2006).From figures 10 and 11, the saturation point was about 60 to 70 t ha−1, which meansSAR is of limited use for estimating high biomass forest. SAR has promise for forestbiomass estimation with low biomass, but at plot level the performance is poor, causedby the complexity of the forest structure and the terrain correction effect. Althoughterrain correction (the TRC image) can correct the effect of different effective back-scattering surface areas caused by the local topography and SAR imagery geometry,because of the complexity of the forest structure, the correction may be inade-quate, which was the limitation of SAR-based biomass estimation in mountainousareas.

Although this method cannot estimate high biomass, as is shown in the exist-ing literature, this article strengthens our knowledge of forest biomass estimationusing SAR data at different scales in mountainous areas. In mountainous areas,owing to the effects of terrain and complex forest structure, biomass estimation atplot level becomes infeasible. However, a relatively feasible result can be achievedat stand level, where a region biomass range can be estimated with relatively highaccuracy.

Page 18: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Prior knowledge-based retrieval from remote-sensing data 727

Field measurements contain errors, field biomass estimation using equations (1)–(4)contains errors and LiDAR-estimated biomass also contains errors. Field measure-ment errors comprise instrument error and operator error, but these errors can be con-trolled. Field biomass estimation using equations (1)–(4) contains errors, but becauseit was not permitted to cut down forest, the error was not validated. Equations (1)–(4)have high correlation coefficients of 0.9887, 0.9568, 0.8622 and 0.9340, respectively.Field biomass estimation using equations (1)–(4) can be considered as truth. Thisarticle used SAR back-scatter to relate to the LiDAR-estimated biomass. The mainerrors derived from the LiDAR-estimated biomass, with 16.28 t ha−1. At plot level, therelationship between SAR back-scatter and the field-measured biomass or LiDAR-derived biomass has a similar trend, and it can be considered that the error didnot affect the trend. The relationship between SAR back-scatter and LiDAR-derivedbiomass at stand level and biomass grade was credible.

6. Conclusion

In this article, the potential of ALOS PALSAR L-band HH and HV polarizationfor forest AGB was estimated in the Qilian Mountain, western China. The LiDARdata with about one hit per m2 were first used for forest biomass estimation at plotlevel, with R2 of 0.77. Then the LiDAR-derived biomass, as prior knowledge, wasused to investigate the relationship between ALOS PALSAR data and biomass atplot level, stand level and in different biomass grades. The results showed that atplot level, there is no obvious relationship between the SAR back-scatter coefficientand LiDAR-derived biomass. The reason may be that in this study area, the for-est structure is complex. At the same biomass level, the tree density can be verydifferent, which affects the SAR signal. But at stand level and in different biomassgrades, an exponential function can be used to describe the relationship between for-est biomass and the back-scattering coefficient. The reason may be that the averagevalue could partly decrease the influence of forest structure at stand level. Meanwhile,terrain radiometric correction is useful for forest biomass estimation, which makes therelationship between SAR signals and forest biomass more obvious. The results alsoshowed that when biomass reached about 60 t ha−1, the ALOS PALSAR HH and HVsignals arrived at saturation in different biomass grades.

AcknowledgementsThe ALOS PALSAR data were provided by JAXA through the ALOS PI project(no. 315). This article has been supported by the National State Key Basic ResearchProject (grant no. 2007CB714404), the Natural Science Foundation of China (grantno. 40871173), a Special Grant For Prevention and Treatment of Infectious Diseases(grant no. 2008ZX10004-012) and the Key Science and Technology R&D Programmeof Qinghai Province (grant no. 2006-6-160-01). The authors also wish to thank all thepeople who participated in the field experiment and all the people who have given helpfor the article.

ReferencesAXELSSON, P., 1999, Processing of laser scanner data-algorithms and applications. ISPRS

Journal of Photogrammetry and Remote Sensing, 54, pp. 138–147.CASTEL, T., BEAUDOIN, A., STACH, N. and STUSSI, N., 2001, Sensitivity of space-borne SAR

data to forest parameters over sloping terrain: theory and experiment. InternationalJournal of Remote Sensing, 22, pp. 2351–2376.

Page 19: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

728 Q.-S. He et al.

DOBSON, M.C., ULABY, F.T., LE TOAN, T., BEAUDOIN, A., KASISCHKE, E.S. and CHRISTENSEN,N., 1992, Dependence of radar back-scatter on coniferous forest biomass. IEEETransactions on Geoscience and Remote Sensing, 30, pp. 412–415.

DRAKE, J.B., DUBAYAH, R.O., CLARK, D.B., KNOX, R.G., BLAIR, J.B., HOFTON, M.A.,CHAZDON, R.L., WEISHAMPEL, J.F. and PRINCE, S.D., 2002, Estimation of tropi-cal forest structural characteristics using large-footprint LiDAR. Remote Sensing ofEnvironment, 79, pp. 305–319.

DRAKE, J.B., KNOX, R.G., DUBAYAH, R.O., CLARK, D.B., CONDIT, R., BLAIR, J.B. andHOFTON, M., 2003, Aboveground biomass estimation in closed canopy Neotropicalforests using LiDAR remote sensing: factors affecting the generality of relationships.Global Ecology and Biogeography, 12, pp. 147–159.

FRANSSON, J. and ISRAELSSON, H., 1999, Estimation of stem volume in boreal forests usingERS-1 C- and JERS-1 L-band SAR data. International Journal of Remote Sensing, 20,pp. 123–137.

HYDE, P., DUBAYAH, R., PETERSON, B., BLAIR, J.B., HOFTON, M., HUNSAKER, C., KNOX, R.and WALKER, W., 2005, Mapping forest structure for wildlife habitat analysis usingwaveform LiDAR: validation of montane ecosystems. Remote Sensing of Environment,96, pp. 427–437.

IMHOFF, M.L., 1995, Radar back-scatter and biomass saturation: ramifications for globalbiomass inventory. IEEE Transactions on Geoscience and Remote Sensing, 33,pp. 511–518.

JENET, M.A., BRENDAN, G., KIMBERLY, P. and VAN, N., 2003, Estimating forest biomass usingsatellite radar: an exploratory study in a temperate Australian eucalyptus forest. ForestEcology and Management, 176, pp. 575–583.

KASISCHKE, E.S., MELACK, J.M. and DOBSON, M.C., 1997, The use of imaging radars forecological applications – a review. Remote Sensing of Environment, 59, pp. 141–156.

KRAUS, K. and PFEIFER, N., 1998, Determination of terrain models in wooded areas with air-borne laser scanner data. ISPRS Journal of Photogrammetry and Remote Sensing, 53,pp. 193–203.

KUPLICH, T., SALVATORI, V. and CURRAN, P., 2000, JERS-1/SAR back-scatter and its relation-ship with biomass of regenerating forests. International Journal of Remote Sensing, 21,pp. 2513–2518.

KURVONEN, L., PULLIAINEN, J. and HALLIKAINEN, M., 1999, Retrieval of biomass in borealforest from multi-temporal ERS-1 and JERS-1 SAR images. International Journal ofRemote Sensing, 37, pp. 198–205.

LEFSKY, M.A., COHEN, W.B., ACKER, S.A., PARKER, G.G. and SHUGART, H.H., 1999a, LiDARremote sensing of the canopy structure and biophysical properties of Douglas firwestern hemlock forests. Remote Sensing of Environment, 70, pp. 339–361.

LEFSKY, M.A., COHEN, W.B., PARKER, G.G. and HARDING, D.J., 2002, LiDAR remote sensingfor ecosystem studies. BioScience, 52, pp. 19–30.

LEFSKY, M.A., HARDING, D., COHEN, W.B., PARKER, G.G. and SHUGART, H.H., 1999b,Surface LiDAR remote sensing of basal area and biomass in deciduous forest of easternMaryland, USA. Remote Sensing of Environment, 67, pp. 83–98.

LE TOAN, T., BEAUDOIN, A., RIOM, J. and GUYON D., 1992, Relating forest biomass to SARdata. IEEE Transactions on Geoscience and Remote Sensing, 30, pp. 403–411.

LIM, K., TREITZ, P., BALDWIN, K., MORRISON, I. and GREEN, J., 2003, LiDAR remote sensingof biophysical properties of tolerant northern hardwood forests. Canadian Journal ofRemote Sensing, 29, pp. 658–678.

LIM, K.S. and TREITZ, P.M., 2004, Estimation of aboveground forest biomass from airbornediscrete return laser scanner data using canopy-based quantile estimators. ScandinavianJournal of Forest Research, 19, pp. 558–570.

LU, D., 2006, The potential and challenge of remote sensing-based biomass estimation.International Journal of Remote Sensing, 27, pp. 1297–1328.

Page 20: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Prior knowledge-based retrieval from remote-sensing data 729

LUCAS, R.M., CRONIN, N., LEE, A., MOGHADDAM, M., WITTE, C. and TICKLE, P., 2006,Empirical relationships between AIRSAR back-scatter and LiDAR-derived forestbiomass, Queensland, Australia. Remote Sensing of Environment, 100, pp. 407–425.

LUCKMAN, A., BAKER, J., HONZAK, M. and LUCAS, R., 1998, Tropical forest biomass densityestimation using JERS-1 SAR: seasonal variation, confidence limits and application toimage mosaics. Remote Sensing of Environment, 63, pp. 126–139.

LUCKMAN, A., BAKER, J., KUPLICH, T.M., YANASSE, C.C.F. and FRERY, A.C., 1997, A studyof the relationship between radar back-scatter and regenerating tropical forest biomassfor space-borne SAR instruments. Remote Sensing of Environment, 60, pp. 1–13.

MEANS, J.E., ACKER, S.A., FITT, B.J., RENSLOW, M., EMERSON, L. and HENDRIX, C.J., 2000,Predicting forest stand characteristics with airborne scanning LiDAR. PhotogrammetricEngineering and Remote Sensing, 66, pp. 1367–1371.

NAESSET, E., 1997, Estimating timber volume of forest stands using airborne laser scanner data.Remote Sensing of Environment, 61, pp. 246–253.

NAESSET, E., 2004, Practical large-scale forest stand inventory using a small-footprint airbornescanning laser. Scandinavian Journal of Forest Research, 19, pp. 164–179.

NELSON, R., KRABILL, W. and TONELLI, J., 1988, Estimating forest biomass and volume usingairborne laser data. Remote Sensing of Environment, 24, pp. 247–267.

NELSON, R., ODERWALD, R. and GREGLIRE, T.G., 1997, Separating the ground and airbornelaser sampling phases to estimate tropical forest basal area, volume and biomass.Remote Sensing of Environment, 60, pp. 311–326.

NELSON, R., SHORT, A. and VALENTI, M., 2004, Measuring biomass and carbon in Delawareusing an airborne profiling LiDAR. Scandinavian Journal of Forest Research, 19,pp. 500–511.

NILSSON, M., 1996, Estimation of tree heights and stand volume using an airborne LiDARsystem. Remote Sensing of Environment, 56, pp. 1–7.

PALOSCIA, S., MACELLONI, G., PAMPALONI, P. and SIGISMONDI, S., 1999, The potentialof C-band and L-band SAR in estimating vegetation biomass: the ERS-1 andJERS-1 experiments. IEEE Transactions on Geoscience and Remote Sensing. 37,pp. 2107– 2110.

POPESCU, S.C., WYNNE, R.H. and SCRIVANI, J.A., 2004, Fusion of small-footprint LiDAR andmulti-spectral data to estimate plot-level volume and biomass in deciduous and pineforests in Virginia, USA. Forest Science, 50, pp. 551–565.

RAUSTE, Y., 2005, Multi-temporal JERS SAR data in boreal forest biomass mapping. RemoteSensing of Environment, 97, pp. 263–275.

RAUSTE, Y., HAME, T., PULLIAINEN, J., HEISKA, K. and HALLIKAINEN, M., 1994, Radar-basedforest biomass estimation. International Journal of Remote Sensing, 15, pp. 2797–2808.

SHIMADA, M. and HIROSAWA, H., 2000, Slope corrections to normalized RCS using SAR inter-ferometry. IEEE Transactions on Geoscience and Remote Sensing, 38, pp. 1479–1484.

ULANDER, L.M.H., 1996, Radiometric slope correction of synthetic-aperture radar images.IEEE Transactions on Geoscience and Remote Sensing, 34, pp. 1115–1122.

WANG, J.Y., JU, K.J., FU, H.E., CHANG, X.X. and HE, H.Y., 1998, Study on biomass of waterconservation forest on north slope of Qilian Mountains. Journal of Fujian College ofForestry, 18, pp. 319–323 [in Chinese].

WATANABE, M., SHIMADA, M., ROSENQVIST, A., TADONO, T., MATSUOKA, M., ROMSHOO,S.A., OHTA, K., FURUTA, R., NAKAMURA, K. and MORIYAMA, T., 2006, Forest struc-ture dependency of the relation between L-band σ 0 and biophysical parameters. IEEETransactions on Geoscience and Remote Sensing, 44, pp. 3154–3165.

YRJO, R., 2005, Multi-temporal JERS SAR data in boreal forest biomass mapping. RemoteSensing of Environment, 97, pp. 263–275.

ZHOU, Y., ZHU, Q. and CHEN, J.M., 2007, Observation and simulation of net primary produc-tivity in Qilian Mountain, western China. Journal of Environmental Management, 85,pp. 574–584.

Page 21: Forest stand biomass estimation using ALOS PALSAR data based on LiDAR-derived prior knowledge in the Qilian Mountain, western China

Copyright of International Journal of Remote Sensing is the property of Taylor & Francis Ltd and its content

may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express

written permission. However, users may print, download, or email articles for individual use.