Estimating montane forest above-ground biomass in the upper reaches of the Heihe River Basin using...

This article was downloaded by: [Monash University Library]On: 06 December 2014, At: 06:53Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number: 1072954 Registeredoffice: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of RemoteSensingPublication details, including instructions for authors andsubscription information:http://www.tandfonline.com/loi/tres20

Estimating montane forest above-ground biomass in the upper reaches ofthe Heihe River Basin using Landsat-TMdataXin Tianab, Zengyuan Lia, Zhongbo Sub, Erxue Chena, Christiaanvan der Tolb, Xin Lic, Yun Guoad, Longhui Lie & Feilong Lingda Research Institute of Forest Resource Information Techniques,Chinese Academy of Forestry, Beijing 100091, PR Chinab Faculty of Geo-Information Science and Earth Observation,University of Twente, Enschede 7500 AA, The Netherlandsc Cold and Arid Regions Environmental and Engineering ResearchInstitute, Chinese Academy of Sciences, Lanzhou, Gansu 730000,PR Chinad Key Laboratory of Spatial Data Mining & Information Sharing ofMinistry Education, Fuzhou University, Fuzhou, Fujian 350002, PRChinae Plant Functional Biology and Climate Change Cluster, Universityof Technology, Sydney, NSW 2007, AustraliaPublished online: 03 Nov 2014.

To cite this article: Xin Tian, Zengyuan Li, Zhongbo Su, Erxue Chen, Christiaan van der Tol, Xin Li,Yun Guo, Longhui Li & Feilong Ling (2014) Estimating montane forest above-ground biomass in theupper reaches of the Heihe River Basin using Landsat-TM data, International Journal of RemoteSensing, 35:21, 7339-7362, DOI: 10.1080/01431161.2014.967888

To link to this article: http://dx.doi.org/10.1080/01431161.2014.967888

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (theContent) contained in the publications on our platform. However, Taylor & Francis,our agents, and our licensors make no representations or warranties whatsoever as tothe accuracy, completeness, or suitability for any purpose of the Content. Any opinionsand views expressed in this publication are the opinions and views of the authors,

http://www.tandfonline.com/loi/tres20http://www.tandfonline.com/action/showCitFormats?doi=10.1080/01431161.2014.967888http://dx.doi.org/10.1080/01431161.2014.967888

and are not the views of or endorsed by Taylor & Francis. The accuracy of the Contentshould not be relied upon and should be independently verified with primary sourcesof information. Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilities whatsoever orhowsoever caused arising directly or indirectly in connection with, in relation to or arisingout of the use of the Content.

This article may be used for research, teaching, and private study purposes. Anysubstantial or systematic reproduction, redistribution, reselling, loan, sub-licensing,systematic supply, or distribution in any form to anyone is expressly forbidden. Terms &Conditions of access and use can be found at http://www.tandfonline.com/page/terms-and-conditions

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

http://www.tandfonline.com/page/terms-and-conditionshttp://www.tandfonline.com/page/terms-and-conditions

Estimating montane forest above-ground biomass in the upper reachesof the Heihe River Basin using Landsat-TM data

Xin Tiana,b, Zengyuan Lia, Zhongbo Sub, Erxue Chena*, Christiaan van der Tolb, Xin Lic,Yun Guoa,d, Longhui Lie, and Feilong Lingd

aResearch Institute of Forest Resource Information Techniques, Chinese Academy of Forestry,Beijing 100091, PR China; bFaculty of Geo-Information Science and Earth Observation, University

of Twente, Enschede 7500 AA, The Netherlands; cCold and Arid Regions Environmental andEngineering Research Institute, Chinese Academy of Sciences, Lanzhou, Gansu 730000, PR China;dKey Laboratory of Spatial Data Mining & Information Sharing of Ministry Education, FuzhouUniversity, Fuzhou, Fujian 350002, PR China; ePlant Functional Biology and Climate Change

Cluster, University of Technology, Sydney, NSW 2007, Australia

(Received 23 January 2014; accepted 1 July 2014)

In this work, the results of above-ground biomass (AGB) estimates from LandsatThematic Mapper 5 (TM) images and field data from the fragmented landscape ofthe upper reaches of the Heihe River Basin (HRB), located in the Qilian Mountains ofGansu province in northwest China, are presented. Estimates of AGB are relevant forsustainable forest management, monitoring global change, and carbon accounting. Thisis particularly true for the Qilian Mountains, which are a water resource protectionzone. We combined forest inventory data from 133 plots with TM images andAdvanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) globaldigital elevation model (GDEM) V2 products (GDEM) in order to analyse the influ-ence of the sun-canopy-sensor plus C (SCS+C) topographic correction on estimationsof forest AGB using the stepwise multiple linear regression (SMLR) and k-nearestneighbour (k-NN) methods. For both methods, our results indicated that the SCS+Ccorrection was necessary for getting more reliable forest AGB estimates within thiscomplex terrain. Remotely sensed AGB estimates were validated against forest inven-tory data using the leave-one-out (LOO) method. An optimized k-NN method wasdesigned by varying both mathematical formulation of the algorithm and remote-sensing data input, which resulted in 3000 different model configurations. Followingtopographic correction, performance of the optimized k-NN method was compared tothat of the regression method. The optimized k-NN method (R2 = 0.59, root meansquare error (RMSE) = 24.92 tonnes ha1) was found to perform much better than theregression method (R2 = 0.42, RMSE = 29.74 tonnes ha1) for forest AGB retrievalover this montane area. Our results indicated that the optimized k-NN method iscapable of operational application to forest AGB estimates in regions where fewinventory data are available.

1. Introduction

Forests are important contributors to the Earths surface radiation budget and to terrestrialcarbon, water, and nitrogen cycles. As such, they are relevant actors in the feedbackmechanisms between climate and biology. A key indicator of the status of the forest isabove-ground biomass (AGB) (Labrecque et al. 2006). AGB impacts all other ecosystemprocesses (Brown, Schroeder, and Kern 1999; Lu, Batistella, and Moran 2005). For

*Corresponding author. Email: [email protected]

International Journal of Remote Sensing, 2014Vol. 35, No. 21, 73397362, http://dx.doi.org/10.1080/01431161.2014.967888

2014 Taylor & Francis

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

example, variations in AGB have an effect on gross primary production (GPP) and netprimary production (NPP), climate change, the radiation balance, water interception, andeven air quality (Houghton et al. 2000; Keller, Palace, and Hurtt 2001; Palacios-Oruetaet al. 2005; Sivrikaya, Kele, and akir 2007; Stephens et al. 2007). AGB is, in turn,impacted by these processes (Routa, Kellomki, and Strandman 2012; Dai et al. 2013;El-Masri et al. 2013; Vanderwel, Coomes, and Purves 2013). AGB is also relevant forcommercial forestry (Ximenes, Gardner, and Kathuria 2008; Bouchard, Landry, andGagnon 2013). Therefore, spatially distributed information of forest AGB is of greatvalue for eco-hydrological process understanding and for scientific and practical tasks.

Traditional sampling-based methods for estimating forest AGB depend on tree allo-metric equations, normally calculated as a function of tree height, the diameter at breastheight (DBH), and other attributes that are all species dependent. To assess regional forestAGB, intensive forest inventories and statistical or geo-statistical interpolation (i.e. Kriginginterpolation) are required. Obtaining conventional inventories of forest parameters isusually tough, costly, time consuming, and requires a large amount of manpower, particu-larly if forested areas are located at high altitudes within steep and rough terrain. When ashorter-temporal-interval and a finer-spatial-resolution assessment of forest AGB isdemanded for ecological construction and environmental protection, remote-sensing tech-niques are viewed as a valuable alternative source of information for mapping forest AGBand other attributes (Maselli et al. 2005; Labrecque et al. 2006). Utilizing multiple sets ofremotely sensed imagery from different time periods (Gallaun et al. 2010; Powell et al.2010; Morel et al. 2011; Saatchi et al. 2012; Wang et al. 2013), temporal changes over largegeographic areas can be mapped. With the purpose of developing remote-sensing productsfor AGB, numerous research projects have focused on the relationship between forest fielddata and optical, multispectral, synthetic aperture radar (SAR), and light detection andranging (lidar) images from airborne and spaceborne sensors (Lefsky et al. 2002; Popescuand Wynne 2004; Salas et al. 2010; Soenen et al. 2010; Heumann 2011; Mattioli et al. 2012;Hauglin et al. 2013; Riegel et al. 2013; Peregon and Yamagata 2013).

Early studies mainly focused on forest AGB retrieval by so-called parametric methods establishing direct or indirect regression relationships between remotely sensed signalresponses and sampled forest AGB. The application of such parametric methods over alarge area requires one to assume spatially homogeneous relationships that link landsurface and remote-sensing signals. The parametric method is easy to perform. Whetherthe results are accurate largely depends on the statistical robustness. When the forestlandscape is fragmented, further complications arise due to topographic irregularities, soilvariability, and understory vegetation (Maselli and Chiesi 2006).

In contrast to a regression method that requires a predefined statistical form, non-parametric methods, such as the k-nearest neighbour (k-NN) method, do not require anyprior formulations of the statistical basis that describe the relationship between the targetparameter and the predictors (Fehrmann et al. 2008). The method has been developed tooperationally monitor forest resource information and has been successfully applied togenerate the spatial distribution of forest attribute estimates in countries such as Finland(Tomppo and Halme 2004), Sweden (Reese et al. 2003), Norway (Gjertsen 2007), Ireland(McInerney and Nieuwenhuis 2009), the USA (Franco-Lopez, Ek, and Bauer 2001;McRoberts et al. 2007), and China (Tian et al. 2012).

However, for both parametric and non-parametric methods, a well-known drawback isthat predictions outside the range of reference data inherently result in large uncertainties(McRoberts 2009; Magnussen, Tomppo, and McRoberts 2010; Baffetta, Corona, andFattorini 2012; Breidenbach, Naesset, and Gobakken 2012). Therefore, that the range of

7340 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

variability in forest AGB in the region under study is represented in the reference data (i.e.in the forest inventory data used to train the algorithm) is often considered to beimportant. As a result, relationships between forest measurements and remotely sensedrepresentations should be established for a large population of forest inventory data. Sucha prerequisite is particularly problematic for parametric methods. If simple linear or non-linear regression formulae are built on a few field measurements located within a largearea and if the independent variables are measured with error, the range of AGB will besuppressed and the variation underestimated. Remote-sensing data yield the same out-come (Gilichinsky et al. 2012). Since multispectral signals saturate at low forest AGBlevels, the fitted results will be worse when applying multispectral remote-sensing data.All of this is less of a problem for non-parametric models. For the k-NN algorithm, the knearest pixels within the training set and the smallest values of spectral distancemeasured, d, are used to compute the AGB of the remote-sensing pixel as a weightedaverage, which, to some extent, may alleviate the impact of the small range of variability.

Since data sets of field measurements are still relatively small and since random andbootstrap sampling methods require much larger data sets, these sampling methods are notsuitable. As an alternative method for calibration and validation, the leave-one-out (LOO)method can be used for assessing prediction errors. Since it is almost unbiased, LOO isfrequently used as a statistical estimator for the performance of a learning algorithm(Lachenbruch 1967; Arlot and Celisse 2010; Meijer and Goeman 2013).

In an earlier work, we integrated forest measurements with multi-parameter remote-sensing data from the Satellite Pour lObservation de la Terre (SPOT) and also L-bandadvanced land observation satellite (ALOS) phased array type L-band synthetic apertureradar (PALSAR) data, to investigate the feasibility of parametric and non-parametricmethods for estimating forest AGB. The work was conducted within an airborne campaignarea, in the small watershed of the Heihe River Basin (HRB) in northwest China (Tian et al.2012). One of the main reasons for performing this study is that the forest in the upperreaches of the HRB is essential for soil and water conservation, water resource protection,and hydrology regulation. The HRB serves as the water tower for downstream urban areasand agriculture in an otherwise extremely dry region. Since the 1990s, the HRB has beenused as a scientific study region for integrated watershed and land-atmosphere studies.Comprehensive experiments such as the Heihe Basin Field Experiment (HEIFE) (Hu et al.1994), Watershed Allied Telemetry Experimental Research (WATER) (Li et al. 2009), andHeihe Watershed Allied Telemetry Experimental Research (HiWATER) (Li et al. 2013)have been conducted. However, a forest resource monitoring study for the entire QilianMountain region has not yet been conducted. In our previous work (Tian et al. 2012), SPOThigh resolution visible (HRV) data only covered a portion of the area. In an attempt toextend the approach to a larger area, we considered that the Landsat Thematic Mapper (TM)provides more useful spectral information for estimating forest structure than SPOT HRV,albeit with a poorer spatial resolution (Cohen and Spies 1992; Labrecque et al. 2006).Therefore, for the task of routinely monitoring forest resources, TM imagery may be auseful alternative to expensive SPOT-5 images. Here, we examined forest AGB using freeand long-time-continuity TM imagery of the complex montane region of the upper HRB.

The goals of this work were to: (i) introduce an effective radiometric terrain correctionin order to alleviate the impacts of topographical environments on quantifying themontane forest AGB; (ii) apply regression and k-NN methods to TM images for mappingthe regional forest AGB; and (iii) test the estimation abilities of regression methods anddifferent configurations of the k-NN algorithm with and without radiometric terrainrectification. The performance of the regression and k-NN methods was systematically

International Journal of Remote Sensing 7341

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

compared and evaluated using the LOO. Finally, the optimal k-NN configurations wereconducted for estimating the regional forest AGB.

2. Study area and data

2.1. Study area

As the second largest inland river basin in the arid region of northwestern China, the HRBconsists of the following three major geomorphic units: (1) the southern QilianMountains, (2) the middle Hexi Corridor, and (3) the northern Alxa Highland. TheQilian Mountains, located in the upper reaches of the HRB, were selected for this studybecause this conserved forested area is hydrologically and ecologically the most importantarea. The Qilian Mountains function as the water source for the irrigation of agriculture inthe Hexi Corridor and maintain ecological viability in the northern Alxa Highland. TheQilian Mountains, for thousands of years, have been said to be the Mother Mountain ofthe Hexi Corridor, because glacial melt water has brought richness to regional people andthe ecosystem. However, due to wood denudation, extensive grazing, and the excessivedevelopment of water resources, the Qilian Mountains are in a state of degradation.Deforestation in the southern Qilian Mountains during the past half century has notonly resulted in the deterioration of the ecological environment but has also caused theweakening of developed soils that, covered by vegetation, conserve water.

The Qilian Mountains span an area of 10,400 km2 and consist of 72% alp, 27% fluvialarea, and 1% oasis. The elevation varies from 1500 m to 6000 m above sea level (seeFigure 1). This area has a temperate continental mountainous climate. During winter,

Figure 1. The location and terrain conditions of the Qilian Mountains within the upper reach of theHeihe River Basin.

7342 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

atmospheric circulation is controlled by the Mongolian anticyclone and conditions are coldand dry with little precipitation. When the atmospheric circulation is controlled by acontinental cyclone during summer, the diurnal difference in temperature is dramatic. Thedifference in precipitation between summer and winter is large, and most of the yearsprecipitation falls during the summer with the annual rainfall being between 350 mm and495 mm. Influenced by the climate and the terrain, prevalent vegetation types in the area aremountainous pastures, shrubs, and forests. The forest, consisting of Picea crassifolia mixedwith a fairly small fraction of Sabina przewalsk trees, only survives on shady slopes(between 2500 m and 3300 m in altitude), while sparse grass inhabits sunlit slopes.Vegetation density varies with terrain, soil, water, and climate factors. Due to a variety ofenvironmental factors, the site is often used as a field laboratory for the development ofremote-sensing biophysical parameter models.

2.2. Ground and remote-sensing data

For this work, we used data collected under the framework ofWATER, performed in the HRBin northwest China during 2008 (Li et al. 2009, 2011). In the WATER project, simultaneousairborne and satellite-borne remote-sensing observations, as well as ground-based measure-ments, were performed. Two forest inventory surveys were conducted from June to August inboth 2007 and 2008 one in 16 permanent forest plots (25 m 25 m) and 69 rectangularforest plots with two sizes (20 m 20 m and 25 m 25 m); and the other in 58 circular forestplots (with diameters ranging from 10 m to 28 m). Measurements obtained included treeheight (H) (m), DBH (cm), the crown width in two orthotropic directions, the first livingbranch height, and the stand density. Trees with a DBH below 3 cm were excluded from thesurvey. Using the following criteria, a total of 133 forest plots were selected from themeasurement database: (1) the prevalence of forest plots dominated by Picea crassifoliaand (2) forest plots with geographic independency with one another required to avoid spatialautocorrelation. Picea crassifolia is the most prevalent tree species within the study area.According to our inventory records, it occupies 99.39% of total measured trees (8667 trees),and therefore, the Picea crassifolia growth equations from Wang et al. (1998), calibrated forthe study area, were applied for calculating the forest plot AGB.

For remote-sensing data, four scenes of TM images with ortho-rectification (L1Tproducts on 10 August 2006, 17 July 2009, and two on 11 August 2009) that fully spannedthe Qilian Mountains in the HRB were acquired from the USGS (http://glovis.usgs.gov).Corresponding Advanced Spaceborne Thermal Emission and Reflection Radiometer(ASTER) global digital elevation model (GDEM) V2 (GDEM, 30 m) products wereobtained from the Japan Aerospace Exploration Agency (JAXA) (http://gdem.ersdac.jspa-cesystems.or.jp). Due to cloudy weather, not a single TM image of satisfactory quality couldbe obtained for the area. We, therefore, chose images during the same seasons of 2006 and2009. TM images were further processed using radiometric, atmospheric correction(FLAASH model) (Adler-Golden et al. 1999), radiometric terrain correction, and a normal-ization process in preparation for forest discrimination and forest AGB retrieval.

3. Methodology

3.1. The SCS+C radiometric terrain correction

Determining topographic effects on classification, image interpretation, and parameterestimation requires quantitative information from satellite imagery and is still problematic.


Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

http://glovis.usgs.govhttp://gdem.ersdac.jspacesystems.or.jphttp://gdem.ersdac.jspacesystems.or.jp

This problem can be very serious in mountainous areas due to large variations in the sunincident angle relative to the ground slope (Ghasemi, Mohammadzadeh, and Sahebi2013).

In this study, the Sun-Canopy-Sensor Plus C (SCS+C) correction model (Soenen,Peddle, and Coburn 2005), a modification of the SCS model (Gu and Gillespie 1998), wasemployed. For this model, the parameter C (defined in Equation (3) below) was added tothe C-correction method. The topographically corrected radiance (TC radiance), H, wasthen calculated using the following equations:

H TcosP cos Z C

cos i C; (1)

where T is the radiance of the uncorrected pixel, P is the terrain slope of the pixel, Z isthe solar zenith angle, and i is the incidence angle, and a linear relationship existsbetween the T and the cosi of the following form:

T a b cos i; (2)

The parameter C is a function of the regression slope (b) and the intercept (a), asfollows:

C ab ; (3)

cos i cosP cos Z sin P sin Z cos; (4)

where is the difference between the solar azimuth angle (a) and the surface aspect ofthe slope angle (S).

The values of P, Z, i, and a can be obtained from the metadata or header file of theremote-sensing image (in this case, the TM image) and S can be derived from DEMinformation (in this case, the GDEM).

C was added to better characterize diffuse sky irradiance (the downwelling spectralirradiance at the surface due to scattered solar flux in the atmosphere). As a result, thistreatment can reduce the overcorrection of faintly illuminated pixels (Soenen, Peddle, andCoburn 2005; Dimitrov and Roumenina 2013). In this study, the constant, C, was onlyadded for coniferous forests (Picea crassifolia). Approximately 14,000 pixels randomlyselected from four scenes (roughly 3500 pixels for each) were used, and, for each pixel,the radiances in the six bands and the value of cosi were extracted. Finally, a separatevalue for C was determined for each spectral band.

3.2. Forest discrimination

To reduce the influence of the other vegetation types (i.e. pastures, shrubs) onestimations of forest AGB, the generation of a forest/non-forest map is necessary foridentifying target pixels in TM images. Due to the complex terrain and the fragmen-tized landscape, we opted for a decision-tree classifier for discriminating betweenforest and non-forest by integrating TM images with GDEM information.Classification of vegetated and non-vegetated areas was performed using the ratio ofTM bands 4 and 3. Due to strong absorption by chlorophyll in the red band and strong

7344 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

reflection by mesophyll tissue in the near infrared band, the ratio of the reflectance inthe near infrared band to that in the red band is high for vegetation; in other words,the ratio vegetation index (RVI) is high and is clearly different from that in non-vegetated areas. Generally, the RVI is higher than 2.0 in vegetated areas and lowerthan 2.0 in non-vegetated areas. As a result of comparative analysis, the mean textureinformation of the fourth band was helpful for further distinguishing forests (2.5)from the other vegetation types (>2.5). Moreover, as mentioned previously, niches ofPicea crassifolia only exist between 2500 m and 3300 m. Using GDEM information,forest/non-forest categories can finally be determined (see Figure 2). A total of 133forest plots and 43 polygons (with sizes of about 100600 pixels each) were chosenfrom high-resolution images in Google Earth for the validation. The criteria forselecting the test polygons were based on visual interpretation of the aerial photo-graphs (acquired in 2008) and archive forest compartment map (20032007). The

Y

Y

N

ASTER GDEMTM images

Y

RVI Aspect

Altitude

Textures

RVI > 2

2500 m < Altitude< 3300 m

0 < Aspect < 45 or135 < Aspect < 360

M4 < 2.5

Non-forest Non-forest

Non-forest

Forest

NN

Non-forest

Y

Figure 2. The decision tree for the forest and non-forest classification (RVI is the ratio vegetationindex; M4 is the mean texture of band 4 of TM image).

Table 1. Confusion matrix for the forest/non-forest map.

Forest Non-forest TotalProducers

accuracy (%)Users

accuracy (%)

Forest 6017 613 6630 91.68 89.80Non-forest 546 5538 6084 89.99 91.03Total 6563 6151 12714 90.88Kappa coefficient 0.81


Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

overall accuracy of the forest/non-forest map (see Figure 3) was 90.88%, with thevalue of the kappa coefficient being 0.81 (Table 1).

3.3. Regression method

Stepwise multiple linear regression (SMLR) is a classical parametric method. For ourstudy, the remote-sensing signals and the indices were the independent variables and theforest AGB was the dependent variable. The goal of retrieving AGB in this way was to beable to compare the results with and without a radiometric terrain correction, and theresults of k-NN estimates with those of a conventional parametric method.

The conventional multivariate regression model can be expressed as follows:

Y X0 a1X1 aiXi; (5)

where is the dependent parameter to be predicted, X0 is the intercept, i is the number ofindependent variables, and a1 i and X1 i are the regression coefficients and the values ofthe independent variables, respectively.

Multicollinearity commonly exists between different types of remote-sensing informa-tion and tends to degenerate the quality of parameter retrievals. In order to estimatevegetation parameters, a large number of vegetation indices, geographic details, andtexture information should be included in the regression. To address this issue, TasseledCap transformation and principal component (PC) processes were conducted, and multiplevegetation indices such as the infrared index (IRI), the normalized difference vegetationindex (NDVI), the RVI, and others, were calculated. Some important texture information

Figure 3. The forest/non-forest map of the study area.

7346 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

(i.e. mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, andcorrelation) for each band within the TM image were extracted using the co-occurrencemodule in the ENVI software (Exelis Visual Information Solutions, Inc., Boulder, CO,USA). In addition, some terrain details (i.e. aspect, slope, and altitude) were extractedfrom the GDEM. By selecting the most significant variables (the probability of F-to-enter = 0.05; the probability of F-to-remove = 0.1), a total of 88 feature variables werecollected and the SMLR was executed in order to establish the regression relationshipbetween these variables and the forest AGB.

3.4. The k-NN method

The k-NN technique is a non-parametric, multivariate approach for processing observa-tions or combinations of observations from sampling units for obtaining estimations ormapping units (McRoberts et al. 2007; McRoberts 2012; Gilichinsky et al. 2012). Theattractiveness of the method is that it is distribution free in that it does not rely on anyunderlying probability distribution for estimations, but on forest conditions (Mattioli et al.2012). When there is good representation of ground sample plots, the method hasperformed well for biomass estimations. As a result of this performance, the techniqueis widely used (Labrecque et al. 2006). The k-NN predicts the unknown value of Y for thejth target pixel (Yj) as being the weighted mean of a set of values of Y (in this case: valuesof the AGB) for the k reference pixels (Yj,i) nearest to the jth target pixel in the multi-dimensional space defined by auxiliary (remotely sensed) variables, as follows:

Yj Xki1

wiYj;i; (6)

where k (

where xj and xr are vectors that contain the values of all of the T feature space variables (inthis case: the spectral information) for the target pixel, j, and for the reference plot, r,respectively.

Contrary to ED, MD takes correlations between variables and the variance of thevariables into account, and it is scale invariant. The variancecovariance matrix of thefeature space variables, C, is used to correct for the multicollinearity factor of the featurespace variables, as follows:

dYjMD ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffixj xr0C1xj xr

q; (9)

where the 0 is transpose of the matrix.The FD enhances the importance of the most informative bands for the specific

parameters to be estimated; and is the modification of the MD (Maselli 2001; Chirici et{C} al. 2008), as follows:

dYjFD ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffixj xr0C1xj xr

q; (10)

where C* is the fuzzy variancecovariance matrix, which is defined as follows:

C PRr1

Fjxr mxr m0

PNj1

Fj

; (11)

where R is the number of training pixels and m* is the fuzzy mean spectral vector of thetraining pixels, which is calculated as follows:

m PRr1

Frxr

PRr1

Fr

; (12)

and Fr is the membership grade of each reference pixel, j, as follows:

Fr 21=2D1p e1=2ZpMp2=Dp ; (13)

where Zp is the value of the parameter to be estimated at training pixel, j; Mp is the meanvalue of the parameter, and Dp is the standard deviation of the parameter.

3.5. Construction of an optimal k-NN

Several factors impact the accuracy of k-NN estimates. These are as follows: (1) the type andnumber of feature space variables, (2) multidimensional distance measures, (3) the actualnumber of the k nearest neighbours, and (4) the size of the sampling window (Chirici et al.2008). To constitute the optimal configuration of k-NN for the sake of improving forest AGBestimation accuracy, both input data (the feature types) and mathematical set-up (the k value,the distance measure, and the extracting window size) of the algorithm were varied. The LOOcross-validation was executed by evaluating the performance of each k-NN configuration atthe pixel scale. The feature types that were varied included the original reflectance in different

7348 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

bands, the derived indices, and ancillary information. Regarding feature space variables, 10different compositions were composed by comparing the sensitivities of 88 feature variablesto the forest AGB. These 10 combinations of feature space variables were used incombination with the three types of distance measure; k varied from 1 to 25 and thewindow sizes from 1 1 to 7 7 (pixels), resulting in a total of 3000 configurations(Table 2) that were tested. Since they can describe 95.6% of the original spectral variance,the first three PCs of the TM images were selected. The optimal k-NN configuration wasdetermined with pixel-wise accuracy of the estimates with the highest Pearson correlationindex (R) and the lowest root mean square error (RMSE). Finally, the best-performing k-NN configuration (out of the 3000 judged based on LOO performance) was selected forestimating and assessing AGB in the study area.

4. Results

4.1. The regression estimation

The consideration of 88 possible independent variables from the TM and GDEM data, andfrom the SMLRwith strict control parameters (probability of F-to-enter = 0.05; probability ofF-to-remove = 0.1), was done in order to determine sensitive variables for fitting theregression of forest AGB. To analyse the influence of the terrain on quantitative retrieval,two regressions were established using feature variables from the uncorrected and corrected(processed by SCS+C) TM data. With respect to this process, the range of variability in themeasured forest AGB should correspondingly represent both training and test set. Therefore,the complete reference data set from the 133 plots was divided into a training set (88 plots) anda test set (45 plots) representing, therefore, nearly two-third and one-third of the plots,respectively using a stratified sampling method based on forest AGB grades. As the forestAGB range was 18.85 tonnes ha1 to 220.65 tonnes ha1, the reference data was separated intofive subclasses (50 tonnes ha1 per stratum): 050 tonnes ha1, 51100 tonnes ha1, 101

Table 2. The k-NN configurations used in this study.

kvalue

Distancemeasures Feature types

Featureextraction method

1-5 ED, MD, FD (1) Altitude, M4, W, B5, B7 Pixel-wise; 3 3, 5 5 and 7 7windows

(2) Altitude, IRI, NDVI, PC1PC3(3) W, ARVI, IRI, B3, B4, B5(4) IRI, NDVI, PC1PC3(5) Altitude, W, S5, E7(6) Altitude, M4, W, PC3s(7) Altitude, IRI, W, M4, E7, B4/B7(8) Altitude, IRI, ARVI, B2/B7, B4/B7, M5

(9) Altitude, IRI, B2, B5, B7(10) Altitude, PC1PC3, M4, B2/B7

Note: M4 and M5 are the mean texture of band 4 and band 5 of TM image, respectively; W is the wetnesscomponent of the Tasseled Cap transformation; B2, B3, B4, B5, and B7 are the band 2, band 3, band 4, band 5,and band 7 reflectances, respectively; ARVI is the atmospherically resistant vegetation index; IRI is the infraredindex; NDVI is the normalized difference vegetation index; PC1PC3 are the first three PCs; S5 is the secondmoment of band 5; E7 is the entropy texture of band 7; and B2/B7 and B4/B7 are the ratios of band 2/band 7 andband 4/band 7, respectively.


Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

150 tonnes ha1, 151200 tonnes ha1, and 201250 tonnes ha1. Within each subclass, thereference data were randomly assigned to either the training or test data set in the ratio ofapproximately 2:1.

Without SCS+C, the fitting model was, as follows:

Y 652:685 0:195 DEM 62:662S70:772S; (14)

which had a modeling accuracy corresponding to of a coefficient of determination, R2,equal to 0.50 and an RMSE of 26.74 tonnes ha1. (DEM and S indicate the altitude andthe slope derived from the GDEM, respectively; and S7 is the texture of the secondmoment in TM band 7).

In contrast, using SCS+C, the following regression model was derived:

Y 4235:235 0:199 DEM 81:107M5 41:238S7: (15)

For this model, R2 = 0.53 and RMSE = 25.99 tonnes ha1. (Here, M5 is the meantexture of TM band 5.)

The two regressions above were validated at the pixel level using the selected test set.It could be seen that the model that incorporated variables from the SCS+C-corrected TMtextures (with R2 = 0.33 and RMSE = 26.44 tonnes ha1) performed better than theuncorrected model (with R2 = 0.15 and RMSE = 29.73 tonnes ha1) (see Figure 4). Themean percentage errors (MPEs) for these two models were 24.15% and 26.32%, respec-tively. Following the SCS+C process, the independent variable changed from S to M5by regression model implying that slope impacts on the estimation may be alleviated.

As expected, the estimated accuracy of the parametric method was, in all cases, quitelow due to limitations linked to geometric and radiometric deviations in the remote-sensing data. Moreover, the complexity of the relationships between forest features andrelevant TM reflectances or induced indices affected the performance of the parametricmodel (Maselli and Chiesi 2006). As with other regions of northwest China, due tobiophysical controlling factors (soil fertility, soil water, topography, and altitude), theforest features in the Qilian Mountains tend to vary significantly over a very small area.

As seen in Figure 4, the deviations of the model output from the field data were large.Although a radiometric terrain correction was performed and the results improved ascompared to the uncorrected model, routine monitoring of the forest AGB using thisparametric method is not an option.

4.2. The k-NN estimation

Considerable variation was found in the quality of the k-NN model output, depending onthe mathematical set-up and the selected feature space variables. Concerning the mathe-matical framework, the 5 5 window extraction clearly outperformed all of the othersampling methods (not shown). For this reason, presentation of the results here is limitedonly to the 5 5 window sampling size. The results of the testing using the LOOprocedure, for the 5 5 window sampling size, are shown in Figure 5. As mentioned,the accuracy of the various k-NN configurations varied a lot. Feature type 8 consistentlyoutperformed all the others, while the prediction accuracy of feature type 4 was alwaysinferior to that of the others.

7350 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

For each feature type listed in Table 2, the optimal k-NN configuration and the resultswere compared (Table 3). The values of k between 10 and 14 for features (3), (4), and (9);and the value of k neighbours between 4 and 7 for the other features were determinedusing the optimized process. The best configuration was as follows: k = 4; distancemeasure MD; feature (8) using the IRI; the atmospherically resistant vegetation index

(a)

(b)

160

R2 = 0.15RMSE = 29.73 (tonnes ha1)

120

80

40

00 40 80

Predicted forest AGB (tonnes ha1)

Mea

sure

d fo

rest

AG

B (

tonn

es h

a1 )

120 160

160

R2 = 0.33RMSE = 26.44 (tonnes ha1)

120

80

40

00 40 80


Mea

sure

d fo

rest

AG

B (

tonn

es h

a1 )

120 160

Figure 4. The fitted results of the measured and predicted value based on a stepwise multiple linearregression (the dashed line is a 1:1 fit: (a) without the SCS+C correction; (b) with the SCS+Ccorrection).


Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

(ARVI); ratios of band 2 to band 7 and band 4 to band 7; and the mean texture of band 5augmented with DEM information from the GDEM. This configuration generated the bestestimates, with R2 = 0.59, RMSE = 24.92 tonnes ha1, and MPE = 20.75%. In contrast,the optimized configuration with k = 12, the MD measure, and feature (4) composed ofthe IRI, the NDVI, and the first three PCs produced the poorest estimates for forest AGB,with R2 = 0.28 and RMSE = 33.40 tonnes ha1. The difference in the accuracy of the twoconfigurations above is high 0.31 for R2 and 8.48 tonnes ha1 for RMSE. All of thefeatures were mathematically optimized using the MD measure, with the exception offeature (3), which performed a little better than the worst optimal configuration.Additional configurations differed marginally with values of R2 between 0.42 to 0.45and the RMSE varying from 29.67 tonnes ha1 to 29.08 tonnes ha1.

The mathematical set-up was analysed further by comparing the results for thedifferent distance measures used and different values of k. In order to analyse theinfluence of the SCS+C correction on forest AGB estimations, the model performancewas plotted against the value of k for the different distance measures for both terrain-corrected and uncorrected feature (8). The results are shown in Figure 6, illustrated interms of both R2 and the RMSE. Clearly for the MD and FD, the terrain-corrected featurevector consistently outperformed the uncorrected feature. Interestingly, it was determined

48

44

40

36

32

28

24

R2

RM

SE

(to

nnes

ha

1 )

0.6

0.5

0.4

0.3

0.2

0.1

0.00 4 8 12

k k16 20 24

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

0 4 8 12 16 20 24

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Figure 5. The R2 and RMSE of the different k-NN configurations ((1) ... (10) are the feature typesin Table 2).

Table 3. The results of optimal k-NN estimations assessed using the LOO cross-validationprocedure for each feature vector (feature type).

Feature type k value Distance measure R2 RMSE (tonnes ha1)

(1) 5 MD 0.45 29.11(2) 5 MD 0.43 29.59(3) 10 FD 0.33 31.99(4) 12 MD 0.28 33.40(5) 4 MD 0.45 29.17(6) 7 MD 0.45 29.08(7) 4 MD 0.43 29.50(8) 4 MD 0.59 24.92(9) 14 MD 0.43 29.57(10) 7 MD 0.42 29.67

7352 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

that for one case, the terrain-corrected variant performed worse than the uncorrectedvariant (lower R2 and higher RMSE) for the case of the ED measure. The overallaccuracies when using the ED for both feature vectors were much lower than when theother two distance measures were used.

For all of the distance measures, R2 initially increased and the RMSE decreased withincreasing k, until a particular value of k was reached. For example, the MD performedbest at k = 5 for the uncorrected feature input, and at k = 4 for the corrected feature input.The performance of the ED changed for corrected and uncorrected feature vectors in thesimilar manner. FD was inferior to MD for the corrected feature vector, but slightly betterthan MD when k > 9.

The best optimized k-NN algorithm out of 3000 configurations was applied in order toderive forest AGB over the Qilian Mountains. As shown in Figure 7, AGB decreased withincreasing altitude and latitude, and increased with increasing longitude. The pattern issimilar to that reported by Peng et al. (2011).

Validated by the LOOmethod on the basis of 133 forest plot measurements (see Figure 8),the overall accuracy of forest AGB retrievals was satisfactory. The slope of the graph plottingthe field AGB against the remotely sensed AGB was approximately 1, and the offset was25.04 tonnes ha1. Differences between the two estimates were much smaller than thoseestimated by the other configurations.

(a) (b)

(c) (d)

0.6ED MD FD

ED MD FD ED MD FD

0.5

0.4

0.3R2

RM

SE

(to

nnes

ha

1 )

0.2

0.1

0.0

50

45

40

35

30

25

200 4 8 12

k

16 20 24

RM

SE

(to

nnes

ha

1 )

50

45

40

35

30

25

200 4 8 12

k

16 20 24

0 4 8 12

k

16 20 24

0.6ED MD FD

0.5

0.4

0.3R2

0.2

0.1

0.00 4 8 12

k

16 20 24

Figure 6. The R2 ((a) and (b)) and RMSE ((c) and (d)) of the three distance estimations based onuncorrected and corrected feature (8) in Table 2 (where ED is the Euclidean distance, MD is theMahalanobis distance, FD is the fuzzy distance; (a) and (c) are provided without the SCS+Ccorrection, and (b) and (d) with the SCS+C terrain correction).


Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

250

200

150

100

50

00 50 100


Mea

sure

d fo

rest

AG

B (

tonn

es h

a1 )

150 200

R2 = 0.59RMSE = 24.92 (tonnes ha1)

250

Figure 8. The cross-validation of the optimum k-NN estimation using the SCS+C process (thedashed line represents a 1:1 fit).

Figure 7. The grade distribution of forest above-ground biomass within the study area.

7354 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

To illustrate the influence of SCS+C correction on the montane forest AGB for the k-NN model, validation for the uncorrected feature (8) was also conducted (Figure 9).With alarger deviation (MPE = 21.67%), the overall accuracy of this feature vector (R2 = 0.52 andRMSE = 27.08 tonnes ha1) was a little lower than for the corrected one. This indicates

250R2 = 0.52RMSE = 27.08 (tonnes ha1)

200

150

100

50

00 50 100


Mea

sure

d fo

rest

AG

B (

tonn

es h

a1 )

150 200 250

Figure 9. The cross-validation of the optimum k-NN estimation without a SCS+C process (thedashed line represents a 1:1 fit).

R2 = 0.42RMSE = 29.74 (tonnes ha1)

0 50 100


Mea

sure

d fo

rest

AG

B (

tonn

es h

a1 )

150 200 250

200

150

100

50

0

250

Figure 10. The cross-validation results of the stepwise multiple linear regression (the dashed lineis a 1:1 fit).


Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

that the SCS+C process had a positive influence on forest AGB estimations using the TMdata.

4.3. A comparison of the performance of the two methods

Based on terrain compensatory feature variables, the results of applying the optimized k-NN configuration (with feature (8)) were also compared to those produced using theSMLR method. To compare the results of the two methods, the LOO was also applied tovalidate the SMLR results. As expected, the performance of the SMLR was worse thanthat of the k-NN, with R2 = 0.42, RMSE = 29.74 tonnes ha1, and MPE = 26.20% (Figure10). In particular, the k-NN method better reproduced the lowest and highest values of theforest AGB than the SMRL method. This was most obviously the case in highly hetero-geneous areas such as valleys and ridges where the forest was very sparse and usuallymixed with shrubs and grass. In these places, large divergence between the SMRL methodand field data occurred. The k-NN method performed better for these areas, although theproblem was not completely resolved.

5. Discussion

The most relevant issues to consider when estimating forest AGB over montane areas arestill practical problems (Dimitrov and Roumenina 2013; Sarker et al. 2012). One impor-tant practical issue for both parametric and non-parametric methods is the requirement ofhaving sufficient sample plots that represent forest AGB grades and the requirement forprecise geo-referencing within the study area (Gilichinsky et al. 2012; Hill et al. 2013). Asmentioned previously, making forest inventories for the cold and arid areas of the QilianMountains is extremely labour intensive.

The fragmented landscape pattern and the relatively small patch size of many moun-tain forests suggest the use of remote-sensing data as an appropriate alternative at finerspatial resolution. However, we concluded that reference data that span the entire variationrange of spectral space and forest variable space for a given geographical region are stillneeded.

In a previous study (Tian et al. 2012), we assessed AGB for a small portion of thesame study area using the high-resolution geometric (HRG) sensor on board the SPOT-5satellite (which has 10 m resolution). Comparing the results from that study with thepresent one indicated that SPOT-5 HRG data did not show an obvious advantage, with theexception that it had a slightly lower RMSE (R2 = 0.48 and RMSE = 20.70 tonnes ha1)than the TM images used in this study for the entire upper reach of the HRB. Besides that,other differences between the studies can possibly explain the reasons for this difference.For this study, more forest inventory data were used (75 plots were used in the 2012study) that spanned a wider range of AGB (18.85220.65 tonnes ha1 as compared to the18.85175.27 tonnes ha1 utilized in the earlier study), thanks to the larger study area.More feature variables, including textures, terrain types, and vegetation indices, were alsoderived than in the earlier study.

Apparently, a higher spatial resolution is not necessarily an advantage, as confirmedby the fact that the 5 5 window (5 5 pixel average) yielded better results than thepixel-by-pixel approach. The 5 5 window generally smoothes the heterogeneity of theforest landscape to some extent. But, from another side, the 5 5 window exaction mightenhance the texture information of the target plot. It is necessary to restate that there aretwo kinds of forest plots, rectangular forest plots with two sizes (20 m 20 m and

7356 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

25 m 25 m) and cycloidal forest plots (with diameters ranging from 10 m to 28 m),measured during the campaign. Generally, the plot geographic boundary may overlayseveral pixel grids or be mismatched with the target plot due to the unavoidable geometricoffset. This might explain why the 5 5 window extraction method performed better thanthe pixel-by-pixel or other methods.

A widely acknowledged problem hindering quantitative thematic information extrac-tion from remote sensing in mountainous areas is the difference in the illumination ofslopes with different steepness and azimuth angles (Kane et al. 2008; Reese and Olsson2011; Yang et al. 2013). To reduce this difference and to unmask the real reflectioncharacteristics of forest attributes, we introduced the SCS+C model. The correction hada clear positive effect on the quality of the results for both SMLR and k-NN method(with MD and FD). For SMLR, the model without SCS+C correction obviously under-estimated forest AGB and the deviations of its outputs were larger than those from thecorrected model. As it accounts for diffuse atmospheric and terrain radiance, thisphysically based correction can preserve the sun-canopy-sensor geometry and preventthe overcorrection (Ghasemi, Mohammadzadeh, and Sahebi 2013). It has been shown toprovide improved corrections over a wide range of terrain and forest structural condi-tions, particularly in steep terrains and for slopes facing away from the sun (Soenen,Peddle, and Coburn 2005; Soenen et al. 2008, 2010). Therefore, after correction, theSMLR compensated for the underestimation of forest AGB to some extent (see Figure4). For k-NN, one exception was the k-NN configuration that used the Euclideandistance, for which the results were poorer after the topographic correction. Closerinspection of this unexpected deterioration of the results after the topographic correctionrevealed that topographic correction using linear regression caused an autocorrelationbetween some variables (slope and reflectance). The simple ED distance measure doesnot take autocorrelation into account; hence, the information content of the feature spaceappears inferior following topographic correction. Introducing a variancecovariancematrix into the multispectral distance calculations solves this problem and explains whyresults using the MD and fuzzy distance measures improved after the topographiccorrection. As previous studies have indicated (Maselli 2001; Chirici et al. 2008), theresults are sensitive to the covariance matrix if the information content of the featurevector is heterogeneous. When the information content of the feature vectors is relativelyuniform, the improvement becomes marginal. In this study, the MD produced betterresults than the FD until k was greater than 8, suggesting that the information content ofthe feature vector was inhomogeneous.

The performance of the SMLR considerably improved following the SCS+C topo-graphic correction, indicating that topographic irregularities create non-linear relationshipsbetween feature variables and the AGB. The parametric method is unable to capture theserelationships well because it is based on the assumption of homogeneity between therelationships linking land surface and remote-sensing signals (Maselli and Chiesi 2006).Two significant advantages of the optimized k-NN method are its ability to derive areasonable forest AGB and its versatility, which preserves the spatial patterns of homo-geneity and heterogeneity (McRoberts et al. 2007), and explains why, in this work, thek-NN method performed better than the SMLR. Here, we would like to stress theadvantage of the LOO method over other calibrationvalidation procedures. The LOOapproach removes the need to separate forest measurements into a training and a valida-tion set. Since the mechanism for the separation of data into two categories may impactthe results, the LOO approach provides an advantage (Weiss 1991; Fuchs et al. 2009).


Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

Moreover, by not splitting the data into two categories, we made optimum use of theavailable data.

6. Conclusion

An optimum k-NN method determined using the LOO procedure was applied in order toderive forest AGB, by integrating forest measurements from TM data as well as GDEMand their relative indices. Satisfactory accuracy was accomplished yielding a coefficient ofdetermination (R2) of 0.59 and an RMSE of 24.92 tonnes ha1. Terrain influence hinderingthe prediction ability of remotely sensed data was mitigated using the SCS+C procedure,independent of the SMLR or the optimum k-NN method.

As evaluated by the LOO procedure, a comparison between the results obtained usingthe SMLR method and the optimized k-NN configuration indicated that the optimumk-NN configuration outperformed the SMLR (R2 = 0.42, RMSE = 29.74 tonnes ha1) witha higher R2, a lower RMSE, and fewer outliers. With a slightly higher RMSE but a muchhigher R2, the use of k-NN for estimating forest AGB using 30 m-resolution TM data anda GDEM was not less accurate than the results obtained using high-resolution (10 m)SPOT-5 HRG data and airborne lidar DEM (R2 = 0.48 and RMSE = 20.70 tonnes ha1), asapplied in a previous study (Tian et al. 2012).

In this study, the radiometric terrain compensation method and the k-NN optimizationstrategy amongst various k-NN configurations have been proved to have positive effectsto estimate of the montane forest AGB. Particularly, the two methods exhibited advan-tages when applied to montane forest areas with limited ground measurements andmiddle-to-high-resolution remote-sensing data.

AcknowledgementGround measurements used for this work were obtained from WATER. We also thank the jointexperimental team for providing the support needed to carry out the campaign. Sincerely, we aregrateful to the anonymous reviewers for the valuable comments.

FundingThis work was supported by the National Basic Research Programme of China (973 Programme) undergrant [2013CB733404]; National Natural Science Foundation under grant [41101379]; andNational HighTechnology Research and Development Programme (863 Programme) under grant [2011AA120405].

ReferenceAdler-Golden, S. M., M. W. Matthewa, L. S. Bernsteina, R. Y. Levinea, A. Berka, S. C. Richtsmeier,

P. K. Acharyaa, G. P. Andersonb, G. Feldeb, J. Gardnerb, M. Hoke, L. S. Jeong, B. Pukall, A.Ratkowski, and H.-H. Burkec. 1999. Atmospheric Correction for Short-Wave Spectral ImageryBased on MODTRAN4. SPIE Proceeding, Imaging Spectrometry 3753: 6169.

Arlot, S., and A. Celisse. 2010. A Survey of Cross-Validation Procedures for Model Selection.Statistics Surveys 4: 4079. doi:10.1214/09-SS054.

Baffetta, F., P. Corona, and L. Fattorini. 2012. A Matching Procedure to Improve K-NN Estimationof Forest Attribute Maps. Forest Ecology and Management 272: 3550. doi:10.1016/j.foreco.2011.06.037.

Bouchard, S., M. Landry, and Y. Gagnon. 2013. Methodology for the Large Scale Assessment ofthe Technical Power Potential of Forest Biomass: Application to the Province of NewBrunswick, Canada. Biomass & Bioenergy 54: 117. doi:10.1016/j.biombioe.2013.03.014.

7358 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

http://dx.doi.org/10.1214/09-SS054http://dx.doi.org/10.1016/j.foreco.2011.06.037http://dx.doi.org/10.1016/j.foreco.2011.06.037http://dx.doi.org/10.1016/j.biombioe.2013.03.014

Breidenbach, J., E. Naesset, and T. Gobakken. 2012. Improving K-Nearest Neighbor Predictions inForest Inventories by Combining High and Low Density Airborne Laser Scanning Data.Remote Sensing of Environment 117: 358365. doi:10.1016/j.rse.2011.10.010.

Brown, S. L., P. Schroeder, and J. S. Kern. 1999. Spatial Distribution of Biomass in Forests of the EasternUSA. Forest Ecology and Management 123: 8190. doi:10.1016/S0378-1127(99)00017-1.

Chirici, G., A. Barbati, P. Corona, M. Marchetti, D. Travaglini, F. Maselli, and R. Bertini. 2008.Non-Parametric and Parametric Methods Using Satellite Images for Estimating Growing StockVolume in Alpine and Mediterranean Forest Ecosystems. Remote Sensing of Environment 112:26862700. doi:10.1016/j.rse.2008.01.002.

Cohen, W. B., and T. A. Spies. 1992. Estimating Structural Attributes of Douglas-Fir/WesternHemlock Forest Stands from Landsat and SPOT Imagery. Remote Sensing of Environment 41:117. doi:10.1016/0034-4257(92)90056-P.

Dai, L. M., J. Jia, D. P. Yu, B. J. Lewis, L. Zhou, W. M. Zhou, W. Zhao, and L. H. Jiang. 2013.Effects of Climate Change on Biomass Carbon Sequestration in Old-Growth ForestEcosystems on Changbai Mountain in Northeast China. Forest Ecology and Management300: 106116. doi:10.1016/j.foreco.2012.06.046.

Dimitrov, P. K., and E. K. Roumenina. 2013. Combining SPOT 5 Imagery with Plotwise andStandwise Forest Data to Estimate Volume and Biomass in Mountainous Coniferous Site.Central European Journal of Geosciences 5: 208222. doi:10.2478/s13533-012-0124-9.

El-Masri, B., R. Barman, P. Meiyappan, Y. Song, M. L. Liang, and A. K. Jain. 2013. CarbonDynamics in the Amazonian Basin: Integration of Eddy Covariance and Ecophysiological Datawith a Land Surface Model. Agricultural and Forest Meteorology 182183: 156167.doi:10.1016/j.agrformet.2013.03.011.

Fehrmann, L., A. Lehtonen, C. Kleinn, and E. Tomppo. 2008. Comparison of Linear and Mixed-Effect Regression Models and a K-Nearest Neighbour Approach for Estimation of Single-TreeBiomass. Canadian Journal of Forest Research-Revue Canadienne de Recherche Forestiere38: 19. doi:10.1139/X07-119.

Franco-Lopez, H., A. R. Ek, and M. E. Bauer. 2001. Estimation and Mapping of Forest StandDensity, Volume, and Cover Type Using the K-Nearest Neighbors Method. Remote Sensing ofEnvironment 77: 251274. doi:10.1016/S0034-4257(01)00209-7.

Fuchs, H., P. Magdon, C. Kleinn, and H. Flessa. 2009. Estimating Aboveground Carbon in aCatchment of the Siberian Forest Tundra: Combining Satellite Imagery and Field Inventory.Remote Sensing of Environment 113: 518531. doi:10.1016/j.rse.2008.07.017.

Gallaun, H., G. Zanchi, G.-J. Nabuurs, G. Hengeveld, M. Schardt, and P. J. Verkerk. 2010. EU-Wide Maps of Growing Stock and Above-Ground Biomass in Forests Based on Remote Sensingand Field Measurements. Forest Ecology and Management 260: 252261. doi:10.1016/j.foreco.2009.10.011.

Ghasemi, N., A. Mohammadzadeh, and M. R. Sahebi. 2013. Assessment of Different TopographicCorrection Methods in ALOS AVNIR-2 Data over a Forest Area. International Journal ofDigital Earth 6: 504520. doi:10.1080/17538947.2011.625049.

Gilichinsky, M., J. Heiskanen, A. Barth, J. Wallerman, M. Egberth, and M. Nilsson. 2012.Histogram Matching for the Calibration of kNN Stem Volume Estimates. InternationalJournal of Remote Sensing 33: 71177131. doi:10.1080/01431161.2012.700134.

Gjertsen, A. K. 2007. Accuracy of Forest Mapping Based on Landsat TM Data and a kNN basedMethod. Remote Sensing of Environment 110: 420430. doi:10.1016/j.rse.2006.08.018.

Gu, D., and A. Gillespie. 1998. Topographic Normalization of Landsat TM Images of Forest Basedon Subpixel Sun-Canopy-Sensor Geometry. Remote Sensing of Environment 64: 166175.doi:10.1016/S0034-4257(97)00177-6.

Hauglin, M., R. Astrup, T. Gobakken, and E. Nsset. 2013. Estimating Single-Tree BranchBiomass of Norway Spruce with Terrestrial Laser Scanning Using Voxel-Based and CrownDimension Features. Scandinavian Journal of Forest Research 28: 456469. doi:10.1080/02827581.2013.777772.

Heumann, B. W. 2011. Satellite Remote Sensing of Mangrove Forests: Recent Advances and FutureOpportunities. Progress in Physical Geography 35: 87108. doi:10.1177/0309133310385371.

Hill, T. C., M. Williams, A. A. Bloom, E. T. A. Mitchard, C. M. Ryan, and B. Bond-Lamberty.2013. Are Inventory Based and Remotely Sensed Above-Ground Biomass EstimatesConsistent? PLoS One 8: e74170. doi:10.1371/journal.pone.0074170.


Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

http://dx.doi.org/10.1016/j.rse.2011.10.010http://dx.doi.org/10.1016/S0378-1127(99)00017-1http://dx.doi.org/10.1016/j.rse.2008.01.002http://dx.doi.org/10.1016/0034-4257(92)90056-Phttp://dx.doi.org/10.1016/j.foreco.2012.06.046http://dx.doi.org/10.2478/s13533-012-0124-9http://dx.doi.org/10.1016/j.agrformet.2013.03.011http://dx.doi.org/10.1139/X07-119http://dx.doi.org/10.1016/S0034-4257(01)00209-7http://dx.doi.org/10.1016/j.rse.2008.07.017http://dx.doi.org/10.1016/j.foreco.2009.10.011http://dx.doi.org/10.1016/j.foreco.2009.10.011http://dx.doi.org/10.1080/17538947.2011.625049http://dx.doi.org/10.1080/01431161.2012.700134http://dx.doi.org/10.1016/j.rse.2006.08.018http://dx.doi.org/10.1016/S0034-4257(97)00177-6http://dx.doi.org/10.1080/02827581.2013.777772http://dx.doi.org/10.1080/02827581.2013.777772http://dx.doi.org/10.1177/0309133310385371http://dx.doi.org/10.1371/journal.pone.0074170

Houghton, R. A., D. L. Skole, A. Carlos, C. A. Nobre, J. L. Hackler, K. T. Lawrence, and W. H.Chomentowski. 2000. Annual Fluxes of Carbon from Deforestation and Regrowth in theBrazilian Amazon. Nature 403: 301304. doi:10.1038/35002062.

Hu, Y. Q., Y. X. Gao, J. M. Wang, G. L. Ji, Z. B. Shen, L. S. Cheng, J. Y. Chen, and S. Q. Li. 1994.Some Achievements in Scientific Research during HEIFE. [In Chinese.] Plateau Meteorology13: 225236.

Kane, V. R., A. R. Gillespie, R. McGaughey, J. A. Lutz, K. Ceder, and J. F. Franklin. 2008.Interpretation and Topographic Compensation of Conifer Canopy Self-Shadowing. RemoteSensing of Environment 112: 38203832. doi:10.1016/j.rse.2008.06.001.

Keller, M., M. Palace, and G. Hurtt. 2001. Biomass Estimation in the Tapajos National Forest,Brazil: Examination of Sampling and Allometric Uncertainties. Forest Ecology andManagement 154: 371382. doi:10.1016/S0378-1127(01)00509-6.

Labrecque, S., R. A. Fournier, J. E. Luther, and D. Piercey. 2006. A Comparison of Four Methodsto Map Biomass from Landsat-TM and Inventory Data in Western Newfoundland. ForestEcology and Management 226: 129144. doi:10.1016/j.foreco.2006.01.030.

Lachenbruch, P. A. 1967. An Almost Unbiased Method of Obtaining Confidence Intervals for theProbability of Misclassification in Discriminant Analysis. Biometrics 23: 639645. doi:10.2307/2528418.

Lefsky, M. A., W. B. Cohen, G. G. Parker, and D. J. Harding. 2002. Lidar Remote Sensing forEcosystem Studies. BioScience 52: 1930. doi:10.1641/0006-3568(2002)052[0019:LRSFES]2.0.CO;2.

Li, X., G. D. Cheng, S. M. Liu, Q. Xiao, M. G. Ma, R. Jin, T. Che, Q. H. Liu, W. Z. Wang, Y. Qi, J.G. Wen, H. Y. Li, G. F. Zhu, J. W. Guo, Y. H. Ran, S. G. Wang, Z. L. Zhu, J. Zhou, X. L. Hu,and Z. W. Xu. 2013. Heihe Watershed Allied Telemetry Experimental Research (HiWATER):Scientific Objectives and Experimental Design. Bulletin of the American MeteorologicalSociety 94: 11451160. doi:10.1175/BAMS-D-12-00154.1.

Li, X., X.W. Li, Z. Y. Li, M. G.Ma, J.Wang, Q. Xiao, Q. Liu, T. Che, E. X. Chen, G. J. Yan, Z. Y. Hu, L.X. Zhang, R. Z. Chu, P. X. Su, Q. H. Liu, S. M. Liu, J. D.Wang, Z. Niu, Y. Chen, R. Jin,W. Z.Wang,Y. Ran, X. Z. Xin, and H. Z. Ren. 2009. Watershed Allied Telemetry Experimental Research.Journal of Geophysical Research-Atmospheres 114: D22103. doi:10.1029/2008JD011590.

Li, X., X. W. Li, K. Roth, M. Menenti, and W. Wagner. 2011. Preface Observing and Modeling theCatchment Scale Water Cycle. Hydrology and Earth System Sciences 15: 597601.doi:10.5194/hess-15-597-2011.

Lu, D., M. Batistella, and E. Moran. 2005. Satellite Estimation of Aboveground Biomass andImpacts of Forest Stand Structure. Photogrammetric Engineering and Remote Sensing 71:967974. doi:10.14358/PERS.71.8.967.

Magnussen, S., E. Tomppo, and R. McRoberts. 2010. A Model-Assisted K-Nearest NeighbourApproach to Remove Extrapolation Bias. Scandinavian Journal of Forest Research 25: 174184. doi:10.1080/02827581003667348.

Maselli, F. 2001. Extension of Environmental Parameters over the Land Surface by ImprovedFuzzy Classification of Remotely Sensed Data. International Journal of Remote Sensing 22:35973610. doi:10.1080/01431160010006458.

Maselli, F., and M. Chiesi. 2006. Evaluation of Statistical Methods to Estimate Forest Volume in aMediterranean Region. IEEE Transactions on Geoscience and Remote Sensing 44: 22392250.doi:10.1109/TGRS.2006.872074.

Maselli, F., G. Chirici, L. Bottai, P. Corona, and M. Marchetti. 2005. Estimation of MediterraneanForest Attributes by the Application of K-NNProcedures toMultitemporal Landsat ETM+ Images.International Journal of Remote Sensing 26: 37813796. doi:10.1080/01431160500166433.

Mattioli, W., V. Quatrini, S. Di Paolo, D. Di Santo, D. Giuliarelli, A. Angelini, L. Portoghesi, and P.Corona. 2012. Experimenting the Design-Based k-NN Approach for Mapping and Estimationunder Forest Management Planning. IForestBiogeosciences and Forestry 5: 2630.doi:10.3832/ifor0604-009.

McInerney, D. O., and M. Nieuwenhuis. 2009. A Comparative Analysis of kNN and Decision TreeMethods for the Irish National Forest Inventory. International Journal of Remote Sensing 30:49374955. doi:10.1080/01431160903022936.

McRoberts, R. 2009. Diagnostic Tools for Nearest Neighbors Techniques When Used with SatelliteImagery. Remote Sensing of Environment 113: 489499. doi:10.1016/j.rse.2008.06.015.

7360 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

http://dx.doi.org/10.1038/35002062http://dx.doi.org/10.1016/j.rse.2008.06.001http://dx.doi.org/10.1016/S0378-1127(01)00509-6http://dx.doi.org/10.1016/j.foreco.2006.01.030http://dx.doi.org/10.2307/2528418http://dx.doi.org/10.2307/2528418http://dx.doi.org/10.1641/0006-3568(2002)052[0019:LRSFES]2.0.CO;2http://dx.doi.org/10.1641/0006-3568(2002)052[0019:LRSFES]2.0.CO;2http://dx.doi.org/10.1175/BAMS-D-12-00154.1http://dx.doi.org/10.1029/2008JD011590http://dx.doi.org/10.5194/hess-15-597-2011http://dx.doi.org/10.14358/PERS.71.8.967http://dx.doi.org/10.1080/02827581003667348http://dx.doi.org/10.1080/01431160010006458http://dx.doi.org/10.1109/TGRS.2006.872074http://dx.doi.org/10.1080/01431160500166433http://dx.doi.org/10.3832/ifor0604-009http://dx.doi.org/10.1080/01431160903022936http://dx.doi.org/10.1016/j.rse.2008.06.015

McRoberts, R. E. 2012. Estimating Forest Attribute Parameters for Small Areas Using NearestNeighbors Techniques. Forest Ecology and Management 272 (SI): 312. doi:10.1016/j.foreco.2011.06.039.

McRoberts, R. E., E. O. Tomppo, A. O. Finley, and J. H. Heikkinen. 2007. Estimating Areal Meansand Variances of Forest Attributes Using the k-Nearest Neighbors Technique and SatelliteImagery. Remote Sensing of Environment 111: 466480. doi:10.1016/j.rse.2007.04.002.

Meijer, R. J., and J. J. Goeman. 2013. Efficient Approximate K-Fold and Leave-One-Out Cross-Validation for Ridge Regression. Biometrical Journal 55: 141155. doi:10.1002/bimj.201200088.

Morel, A. C., S. S. Saatchi, Y. Malhi, N. J. Berry, L. Banin, D. Burslem, R. Nilus, and R. C. Ong.2011. Estimating Aboveground Biomass in Forest and Oil Palm Plantation in Sabah,Malaysian Borneo Using ALOS PALSAR Data. Forest Ecology and Management 262:17861798. doi:10.1016/j.foreco.2011.07.008.

Palacios-Orueta, A., E. Chuvieco, A. Parra, and C. Carmona-Moreno. 2005. Biomass BurningEmissions: A Review of Models Using Remote Sensing Data. Environmental Monitoring andAssessment 104: 189209. doi:10.1007/s10661-005-1611-y.

Peng, S. Z., C. Y. Zhao, X. L. Zhen, Z. L. Xu, and L. He. 2011. Spatial Distribution Characteristicsof the Biomass and Carbon Storage of Qinghai Spruce (Picea crassifolia) Forests in QilianMountains. [In Chinese.] Chinese Journal of Applied Ecology 22: 16891694.

Peregon, A., and Y. Yamagata. 2013. The Use of ALOS/PALSAR Backscatter to Estimate Above-Ground Forest Biomass: A Case Study in Western Siberia. Remote Sensing of Environment137: 139146. doi:10.1016/j.rse.2013.06.012.

Popescu, S. C., and R. H. Wynne. 2004. Seeing the Trees in the Forest: Using Lidar andMultispectral Data Fusion with Local Filtering and Variable Window Size for Estimating TreeHeight. Photogrammetric Engineering and Remote Sensing 70: 589604. doi:10.14358/PERS.70.5.589.

Powell, S. L., W. B. Cohen, S. P. Healey, R. E. Kennedy, G. G. Moisen, K. B. Pierce, and J. L.Ohmann. 2010. Quantification of Live Aboveground Forest Biomass Dynamics with LandsatTime-Series and Field Inventory Data: A Comparison of Empirical Modeling Approaches.Remote Sensing of Environment 114: 10531068. doi:10.1016/j.rse.2009.12.018.

Reese, H., M. Nilsson, T. G. Pahlen, Q. Hagner, S. Joyce, U. Tingelof, M. Egberth, and H. Olsson.2003. Countrywide Estimates of Forest Variables Using Satellite Data and Field Data from theNational Forest Inventory. AMBIO 32: 542548.

Reese, H., and H. Olsson. 2011. C-Correction of Optical Satellite Data over Alpine VegetationAreas: A Comparison of Sampling Strategies for Determining the Empirical C-Parameter.Remote Sensing of Environment 115: 13871400. doi:10.1016/j.rse.2011.01.019.

Riegel, J. B., E. Bernhardt, J. Swenson, and M. Convertino. 2013. Estimating Above-GroundCarbon Biomass in a Newly Restored Coastal Plain Wetland Using Remote Sensing. PLoSOne 8: e68251. doi:10.1371/journal.pone.0068251.

Routa, J., S. Kellomki, and H. Strandman. 2012. Effects of Forest Management on Total BiomassProduction and CO2 Emissions from use of Energy Biomass of Norway Spruce and Scots Pine.Bioenergy Research 5: 733747. doi:10.1007/s12155-012-9183-5.

Saatchi, S., L. Ulander, M. Williams, S. Quegan, T. LeToan, H. Shugart, and J. Chave. 2012. ForestBiomass and the Science of Inventory from Space. Nature Climate Change 2: 826827.doi:10.1038/nclimate1759.

Salas, C., L. Ene, T. G. Gregoire, E. Nsset, and T. Gobakken. 2010. Modelling Tree Diameterfrom Airborne Laser Scanning Derived Variables: A Comparison of Spatial Statistical Models.Remote Sensing of Environment 114: 12771285. doi:10.1016/j.rse.2010.01.020.

Sarker, M. L. R., J. Nichol, B. Ahmad, I. Busu, and A. A. Rahman. 2012. Potential of TextureMeasurements of Two-Date Dual Polarization PALSAR Data for the Improvement of ForestBiomass Estimation. ISPRS Journal of Photogrammetry and Remote Sensing 69: 146166.doi:10.1016/j.isprsjprs.2012.03.002.

Sivrikaya, F., S. Kele, and G. akir. 2007. Spatial Distribution and Temporal Change of CarbonStorage in Timber Biomass of Two Different Forest Management Units. EnvironmentalMonitoring and Assessment 132: 429438. doi:10.1007/s10661-006-9545-6.

Soenen, S. A., D. R. Peddle, and C. A. Coburn. 2005. SCS+C: A Modified Sun-Canopy-SensorTopographic Correction in Forested Terrain. IEEE Transactions on Geoscience and RemoteSensing 43: 21482159. doi:10.1109/TGRS.2005.852480.


Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

http://dx.doi.org/10.1016/j.foreco.2011.06.039http://dx.doi.org/10.1016/j.foreco.2011.06.039http://dx.doi.org/10.1016/j.rse.2007.04.002http://dx.doi.org/10.1002/bimj.201200088http://dx.doi.org/10.1016/j.foreco.2011.07.008http://dx.doi.org/10.1007/s10661-005-1611-yhttp://dx.doi.org/10.1016/j.rse.2013.06.012http://dx.doi.org/10.14358/PERS.70.5.589http://dx.doi.org/10.14358/PERS.70.5.589http://dx.doi.org/10.1016/j.rse.2009.12.018http://dx.doi.org/10.1016/j.rse.2011.01.019http://dx.doi.org/10.1371/journal.pone.0068251http://dx.doi.org/10.1007/s12155-012-9183-5http://dx.doi.org/10.1038/nclimate1759http://dx.doi.org/10.1016/j.rse.2010.01.020http://dx.doi.org/10.1016/j.isprsjprs.2012.03.002http://dx.doi.org/10.1007/s10661-006-9545-6http://dx.doi.org/10.1109/TGRS.2005.852480

Soenen, S. A., D. R. Peddle, C. A. Coburn, R. J. Hall, and F. G. Hall. 2008. Improved TopographicCorrection of Forest Image Data Using a 3-D Canopy Reflectance Model in Multiple ForwardMode. International Journal of Remote Sensing 29: 10071027. doi:10.1080/01431160701311291.

Soenen, S. A., D. R. Peddle, R. J. Hall, C. A. Coburn, and F. G. Hall. 2010. EstimatingAboveground Forest Biomass from Canopy Reflectance Model Inversion in MountainousTerrain. Remote Sensing of Environment 114: 13251337. doi:10.1016/j.rse.2009.12.012.

Stephens, B. B., K. R. Gurney, P. P. Tans, C. Sweeney, W. Peters, L. Bruhwiler, P. Ciais, M.Ramonet, P. Bousquet, T. Nakazawa, S. Aoki, T. Machida, G. Inoue, N. Vinnichenko, J. Lloyd,A. Jordan, M. Heimann, O. Shibistova, R. L. Langenfelds, L. P. Steele, R. J. Francey, and A. S.Denning. 2007. Weak Northern and Strong Tropical Land Carbon Uptake from VerticalProfiles of Atmospheric CO2. Science 316: 17321735. doi:10.1126/science.1137004.

Tian, X., Z. B. Su, E. X. Chen, Z. Y. Li, C. van der Tol, J. P. Guo, and Q. S. He. 2012. Estimationof Forest Above-Ground Biomass Using Multi-Parameter Remote Sensing Data over a Cold andArid Area. International Journal of Applied Earth Observation and Geoinformation 14: 160168. doi:10.1016/j.jag.2011.09.010.

Tomppo, E. O., and M. Halme. 2004. Using Coarse Scale Forest Variables as Ancillary Informationand Weighting of Variables in kNN Estimation: A Genetic Algorithm Approach. RemoteSensing of Environment 92: 120. doi:10.1016/j.rse.2004.04.003.

Vanderwel, M. C., D. A. Coomes, and D. W. Purves. 2013. Quantifying Variation in ForestDisturbance, and Its Effects on Aboveground Biomass Dynamics, across the Eastern UnitedStates. Global Change Biology 19: 15041517. doi:10.1111/gcb.12152.

Wang, J. Y., K. J. Ju, H. E. Fu, X. X. Chang, and H. Y. He. 1998. Study on Biomass of WaterConservation Forest on North Slope of Qilian Mountains. [In Chinese.] Journal of FujianCollege of Forestry 18: 319323.

Wang, X. C., G. F. Shao, H. Chen, B. J. Lewis, G. Qi, D. P. Yu, L. Zhou, and L. M. Dai. 2013. AnApplication of Remote Sensing Data in Mapping Landscape-Level Forest Biomass forMonitoring the Effectiveness of Forest Policies in Northeastern China. EnvironmentalManagement 52: 612620. doi:10.1007/s00267-013-0089-6.

Weiss, S. M. 1991. Small Sample Error Rate Estimation for k-NN Classifiers. IEEE Transactionson Pattern Analysis and Machine Intelligence 13: 285289. doi:10.1109/34.75516.

Ximenes, F. A., W. D. Gardner, and A. Kathuria. 2008. Proportion of Above-Ground Biomass inCommercial Logs and Residues following the Harvest of Five Commercial Forest Species inAustralia. Forest Ecology and Management 256: 335346. doi:10.1016/j.foreco.2008.04.037.

Yang, G. J., R. L. Pu, J. X. Zhang, C. J. Zhao, H. K. Feng, and J. H. Wang. 2013. Remote Sensingof Seasonal Variability of Fractional Vegetation Cover and Its Object-Based Spatial PatternAnalysis over Mountain Areas. ISPRS Journal of Photogrammetry and Remote Sensing 77:7993. doi:10.1016/j.isprsjprs.2012.11.008.

7362 X. Tian et al.

Dow

nloa

ded

by [

Mon

ash

Uni

vers

ity L

ibra

ry]

at 0

6:53

06

Dec

embe

r 20

14

http://dx.doi.org/10.1080/01431160701311291http://dx.doi.org/10.1080/01431160701311291http://dx.doi.org/10.1016/j.rse.2009.12.012http://dx.doi.org/10.1126/science.1137004http://dx.doi.org/10.1016/j.jag.2011.09.010http://dx.doi.org/10.1016/j.rse.2004.04.003http://dx.doi.org/10.1111/gcb.12152http://dx.doi.org/10.1007/s00267-013-0089-6http://dx.doi.org/10.1109/34.75516http://dx.doi.org/10.1016/j.foreco.2008.04.037http://dx.doi.org/10.1016/j.isprsjprs.2012.11.008

Abstract1. Introduction2. Study area and data2.1. Study area2.2. Ground and remote-sensing data

3. Methodology3.1. The SCS+C radiometric terrain correction3.2. Forest discrimination3.3. Regression method3.4. The k-NN method3.5. Construction of an optimal k-NN

4. Results4.1. The regression estimation4.2. The k-NN estimation4.3. A comparison of the performance of the two methods

5. Discussion6. ConclusionAcknowledgementFundingReference

Estimating montane forest above-ground biomass in the upper reaches of the Heihe River Basin using...

Documents

Transcript of Estimating montane forest above-ground biomass in the upper reaches of the Heihe River Basin using...