EstimatingtheAbovegroundBiomassofanEvergreenBroadleaf...

Research ArticleEstimating the Aboveground Biomass of an Evergreen BroadleafForest in Xuan Lien Nature Reserve Thanh Hoa Vietnam UsingSPOT-6 Data and the Random Forest Algorithm

The Dung Nguyen 12 and Martin Kappas1

1Department of Cartography GIS and Remote Sensing Georg-August-Universitat Gottingen Gottingen 37077 Germany2Department of Forest Inventory and Planning Vietnam National University of Forestry Xuan Mai Ha Noi 156200 Vietnam

Correspondence should be addressed to e Dung Nguyen tnguyen4gwdgde

Received 28 February 2020 Revised 7 August 2020 Accepted 12 August 2020 Published 27 August 2020

Academic Editor omas Campagnaro

Copyright copy 2020e Dung Nguyen and Martin Kappas is is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in anymedium provided the original work isproperly cited

Forest biomass is an important ecological indicator for the sustainable management of forests e aim of this study was toestimate forest aboveground biomass (AGB) by integrating SPOT-6 data with field-based measurements using the random forest(RF) algorithm In total 52 remote sensing variables including spectral bands vegetation indices topography data and textureswere extracted from SPOT-6 images to predict the forest AGB of Xuan Lien Nature Reserve Vietnam To determine the optimalpredictor variables for AGB estimation 10 different RF models were built To evaluate these models 10-fold cross-validation wasapplied We found that a combination of spectral and vegetation indices and topography variables offer the highest predictionresults (R2

adj 074 and RMSE 6124Mghaminus1) Adding texture features into the predictor variables did not improve the modelperformance In addition the SPOT-6 sensor has the potential to predict forest AGB using the RF algorithm

1 Introduction

e forest ecosystem is one of the primary sources of carbonstorage in the terrestrial ecosystem and constitutes ap-proximately 80 of all living terrestrial biomass [1] With amassive carbon pool the forest ecosystem plays an im-portant role in reducing global warming [2 3] Most ac-tivities related to forest biomass assessments focus on theaboveground biomass (AGB) of living trees because AGBrepresents the largest amount of total biomass in forestseaccurate assessment and evaluation of forest AGB stores andtheir spatiotemporal patterns are important for the sus-tainable management of forests [4 5] Estimating AGB is oneof the most important steps in measuring and evaluating thecarbon stocks and carbon sequestration of forests [6]

In general field measurements (including destructivesampling or using allometric equationsconversion factors)and remote sensing (RS) are the main methods used toestimate forest AGB [7] e traditional method based on

field measurements is the most accurate but is difficult tocover large areas due to it being expensive labor-intensiveimpractical harmful to nature and time-consuming at alarge scale [8ndash10] Compared to the traditional approach theRS technique has advantages in its ability to obtain effectiveand repeatable vegetation information in large areas es-pecially for remote regions [11] e forest AGB can beestimated from different RS sensor types including syntheticaperture radars [11ndash13] light detection and ranging (Li-DAR) and optical sensors e radar and LiDAR datasetshave the advantage of penetrability through the forestcanopy to obtain more information in the following such astrunks and branches which contain more than 60 of theAGB [11] this information will help achieve a higher ac-curacy e limitations of these types of datasets are theirhigh costs and large data volume requirements to capture theinformation in large-scale areas [14ndash17] Besides radar andLiDAR very high-resolution (VHR) optical images such asIKONOSWorldView-2 GeoEye-1 and SPOT-6 or SPOT-7

HindawiInternational Journal of Forestry ResearchVolume 2020 Article ID 4216160 13 pageshttpsdoiorg10115520204216160

also allow one to estimate AGB by using the empirical re-lationships between the AGB and RS spectral bands vege-tation indices (VIs) and texture and topographicinformation with acceptable accuracy For example Motlaghet al [18] Hirata et al [19] Hussin et al [20] Karna et al[21] Li et al [22] and Gara et al [23] successfully predictedforest biomass based on those LiDAR and VHR sensorscombined with sufficient field data

Regardless of the RS data sources there are no RStechniques that are capable of providing a direct measure-ment of biomass As a result biomass prediction accuracyincreases when combined with field-sampled data especiallywhen using machine learning approaches to build biomassmodels [4 24ndash26] Machine learning algorithms allow one toanalyze a large number of predictor variables from remotesensing data thereby filling in the missing data and reducingthe error of the prediction models [27ndash29] A wide variety ofmachine learning algorithms have been employed to esti-mate AGB including an artificial neural network (ANN)K-nearest neighbor (KNN) support vector machine (SVM)and random forest (RF) In recent years RF has been widelyused to develop predictive models for AGB at the localregional and global areas because it can run efficiently onlarge datasets with a high accuracy Furthermore RF has theability to determine the importance of variables [30 31]

Evergreen broadleaf (EB) forests are estimated to covermore than 57 of national forests [32ndash35] and harbourapproximately 44 of the total forest carbon stock inVietnam [35] is forest type plays an important key role inecosystem carbon sequestration in Vietnam However inVietnam there have only been a few studies on forest carbonestimation with the integration of RS techniques especiallyusing VHR sensors For instance Dang et al [36] usedSentinel-2 satellite images combined with field-measureddata to estimate the AGB in Yok Don National Park eresearch of Pham and Brabyn [8] successfully proved theaccuracy of predicting the AGB of mangrove forests in CanGio (73) by integrating spectral information vegetationtype texture features and vegetation indices from SPOT-4and SPOT-5 images

e main objective of this study is (i) to test the ability ofthe spectral and vegetation indices and topographic andtexture features derived from SPOT-6 images to predictAGB in combination with field data using the RF algorithmand (ii) to identify the most desirable predictors for AGBestimation

2 Materials and Methods

21 Study Site is study was conducted at Xuan LienNature Reserve anh Hoa Vietnam located at 19∘52prime-20∘02primeN 104∘58prime-105∘15primeE which covers 23404 ha of twoforest types in the southwest of anh Hoa province isreserve is bordered by Cao River in the north the Nghe Anprovince in the south and west and the Ta Leo and BuKhongmountains and the confluence of Cao and Chu Riversto the east (Figure 1) e study area is situated in a belt ofmountains from Sam Neua in Laos to the uong Xuan andNhu Xuan districts in anh Hoa province which contain

some high peaks (eg Ta Leo (1400m) Bu Cho (1563m) BuHon Han (1208m) and an unnamed 1605m peak) emean temperature is about 23ndash24degC and the mean annualrainfall is approximately 1700ndash1900mm which occursmainly from May to October and accounts for 90 of thetotal annual rainfall [37] e main soil in the nature reserveis feralite soil feralite humus soil in the medium-highmountains (FH) feralite soil in the lowlands (F) and alluvialsoil (P) associated with streams or rivers and the valleybottom [37]

e vegetation in the study area was mainly closedevergreen broadleaf forest which was classified into threeforest types [38] based on the classification of ai VanTrung [39] e first forest type is distributed from mediumto high montane consisting of mixed coniferous andbroadleaf evergreen forests (MCBEV) between 800m and1605m (asl) is forest type is generally undisturbed anddominated by upper storeyed broadleaf tree species from thefamilies of Fagaceae Lauraceae Euphorbiaceae FabaceaeMagnoliaceae Dipterocarpaceae and Sapotaceae [37 38]e second forest type is located in the low montanebroadleaf evergreen forests (BEV) which are distributedunder 800m asl and have been weakly impacted by humanactivities Common species include Leguminosae Euphor-biaeceae Lauraceae Rutaceae Rosaceae and Meliaceae[37 38] e final forest type is secondary forests (SF) whichare mainly a mix of Neohouzeaua dullooa Dendrocalamuspatellaris Bambusa sp and broadleaf evergreen forest[37 38]

22 Field Data Collection Field surveys were conducted inFebruary 2015 In total 180 plots (20mtimes 25m) were sampledand inventoried ese plots were randomly generated inArcGIS 104 and then located in the field using a GPS devicewith errors up to 5m Within the plots the diameter at breastheight (dbh) and the total height (h) of each living tree withdbh greater than 5 cm were measured using a diameter tapeand a Vertex Hypsometer respectively Tree species were alsorecorded for each measured tree

23 Aboveground Biomass Estimation We considered onlythe aboveground living tree biomass for carbon estimationsAboveground biomass (AGB) was estimated as the sum ofthe individual components (stumps stems bark branchesseeds and foliage) of the individual living trees that werepredicted using appropriate allometric equations [6] eseallometric equations were carefully chosen depending on theforest types and the tree or bamboo species available in theinput dataset For the evergreen broadleaf forests we usedthe biomass equation developed by Huy et al [40] whichwas specifically developed for evergreen broadleaf forests inthe North Central region of Vietnam (1) For bambooforests we opted for the equation from Vu et al [41] whichwas developed for bamboo forests at a national scale (2) Formixed forests of bamboo and evergreen broadleaf forestsboth (1) and (2) were used to estimate the total biomass Allof the selected equations above are based on the treebamboodiameters (dbh) and total heights (h)

2 International Journal of Forestry Research

AGBt 2532449 times db h2h1113872 1113873

095102R2

09555 (1)

AGBb 019431 times db h16922

times h02778

R2

07810 (2)

Finally to synchronize the estimated AGB for eachsample plot to the remotely sensed data the AGB valueswere prorated and scaled to obtain the per-hectare values

24 Remotely Sensed (RS) Data Due to the availabilitySPOT-6 dataset was opted as RS data in this paper SPOT-6is an optical satellite that was developed by Astrium with thecapacity to obtain panchromatic and multispectral imageryat spectral resolutions of 15m and 6m respectively [42]Two orthorectified scenes of SPOT-6 images taken on 20May and 05 December 2013 were obtained for this researchBoth image scenes consist of four multispectral bands (blue450ndash520 nm green 530ndash590 nm red 625ndash695 nm andnear-infrared (NIR) 760ndash890 nm) each with a 6m spatialresolution and one panchromatic band (450ndash745 nm) with a15m spatial resolution [42] e digital number (DN) of theSPOT-6 images was first used to calculate the radiance data

and then convert those data to the reflectance value usingatmospheric correction in ENVI 54 We applied theFLAASH (Fast Line-of-Sight Atmospheric Analysis ofSpectral Hypercube) radiative transfer model to correct theatmospheric interference in each image [43]

e 6m spatial resolution digital elevationmodel (DEM)was first created from a topographic map with 5m contourlines [44] using the ldquoTopo to Rasterrdquo interpolationmethod inArcGIS 104 e topographic data (elevation slope andaspect) were then generated from a 6m DEM

25 Variables for AGB Prediction To explore the effective-ness of the SPOT-6 sensor for estimating forest AGB dif-ferent types of RS features were considered ese featuresincluded raw spectral bands topographic data vegetationindices (VIs) and texture (Table 1) Based on the coordi-nates size and shape of each sample plot we created apolygon shapefile using the ldquorectangles ovals and dia-mondsrdquo plugin in QGIS 180 [52] which we then overlaidonto the RS datae values of all pixels within each polygonplot were derived for the four different spectral bands andthen averaged for each plot e extracted values were then

0 2 4 6 81Km

1cm = 1km

Very rich EB forest

Rich EB forestMedium EB forestPoor EB forest

Bamboo

PlantationBarelandBotanic garden

Agriculture

SettlementWaterSample points

20deg5primeN

104deg50primeE

100degE 104degE 108degE

104deg55primeE 105deg0primeE 105deg5primeE 105deg10primeE 105deg15primeE 105deg20primeE

104deg50primeE 104deg55primeE 105deg0primeE 105deg5primeE 105deg10primeE 105deg15primeE 105deg20primeE

20deg0primeN

19deg5

5primeN

22degN

18degN

14degN

10degN

19deg5

0primeN

20deg5primeN

20deg0primeN

19deg5

5primeN

19deg5

0primeN

N

Cua Dat Lake

Service Layer Credits National Geographic Esri GarminHERE UNEP-WCMC USGS NASA ESA METI NRCANGEBCO NOAA increment P Corp

Figure 1 Location of the sample plots in Xuan Lien Nature Reserve

International Journal of Forestry Research 3

used to calculate the 9 VIs We used the following vegetationindices most often used in remote sensing-based studies onforest biomass and its properties [4 45 53 54] NDVI(normalized difference vegetation index) RDVI (renor-malized difference vegetation index) RVI (ratio vegetationindex) DVI (difference vegetation index) MSR (modifiedsimple ratio) and EVI (enhanced vegetation index) Sincesome locations in the study area have low vegetation cover(Figure 1) we additionally used SAVI (soil-adjusted vege-tation index) OSAVI (optimized soil-adjusted vegetationindex) and GEMI (global environment monitoring index)to minimize the effect of soil background reflectance [47]e topographical conditions including elevation slopeand aspect were also considered as factors affecting theforestrsquos structure composition and distribution [55ndash57]e texture feature calculations were carried out using PCIGeomatica 2013 ese calculations were performed on allimages using a 5times 5 (900m2) 6m-pixel window [50] Foreach spectral band eight texture parameters as per Haralicket al [51] were calculated In total 52 independent variableswere used

26 Correlation between the AGB and RS Data e analysisof the relationship between the AGB and RS data was carriedout using the RF algorithm that was integrated into therandomForest package in R software [58] RF is an ensemblemachine learning algorithm that has been widely used inbiomass modeling with the advantages of being able tohandle a large number of input variables and identify themost significant variables as well as to reduce or evenovercome the overfitting problem and thereby improvemodel accuracy [8 59 60] e RF algorithm (RF) was firstdeveloped by Breiman [30] is ensemble learning methodgenerates many decision trees from a randomly selectedsample via bootstrapping known as a training dataset efeatures for modeling at each node of the decision trees arealso randomly selected e results are then obtained byaveraging the predictions from all decision trees To estimatethe model errors a subset of samples comprising theremaining data from the original dataset (called out-of-bagdata or OOB data) is used as validation samples ese OOBdata are not only used to calculate prediction errors bycomparing the predictions from the training dataset with the

Table 1 Variables used in this study for estimating biomass

Categories Variables Algorithm References

Raw spectral features

Blue B1 (mean)Green B2 (mean)Red B3 (mean)NIR B4 (mean)

TopographyDEM

6 metersSlopeAspect

Vegetation indices

NDVI NDVI B4 minus B3B4 + B3 [45]RVI RVI B4B3 [45]DVI DVI B4 minus B3 [45]RDVI RDVI (B4 minus B3)

B4 + B3

1113968[46]

MSR MSR (B4B3 minus 1)(B4B3

1113968+ 1) [14]

SAVI SAVI (B4 minus B3B4 + B3 + 05) times (1 + 05) [47]OSAVI OSAVI (1 + 16)(B4 minus B3B4 + B3 + 016) [48]

GEMI GEMI n(1 minus 025n) minus B3 minus 01251 minus B3 [49]n 2(B2

4 minus B23) + 15B4 + 05B3B4 + B3 + 05

EVI EVI 25( B4 minus B3B4 + 60B3 minus 75B1 + 1) [47]

Texture (derived from each spectralband)

GLCM mean (mean) Meani 1113936Nminus1ij0ilowastPij [50 51]Meanj 1113936

Nminus1ij0jlowastPij

GLCM variance (Var) Var 1113936Nminus1ij0Pij lowast (i minus Mean)2 [50 51]

Homogeneity (Hom) Hom 1113936Nminus1ij0(Pij1 + (i minus j)2) [50 51]

Contrast (Con) Con 1113936Nminus1ij0Pij lowast (i minus j)2 [50 51]

Dissimilarity (Dis) Dis 1113936Nminus1ij0Pij lowast |i minus j| [50 51]

Entropy (Ent) Ent 1113936Nminus1ij0(minusPij lowastLn(Pij)) [50 51]

Angular second moment(ASM) ASM 1113936

Nminus1ij0(P2

ij) [50 51]

Correlation (Cor) Cor 1113936Nminus1ij0Pij[(i minus Meani)(j minus Meanj)

Vari lowastVarj

1113969] [50 51]

Inverse difference (InvD) InvD 1113936Nminus1ij0

Pij

|iminus j|2for i j [50 51]

Note NIR near infrared DEM digital elevation model NDVI normalized difference vegetation index RVI ratio vegetation index DVI differencevegetation index RDVI renormalized difference vegetation index MSR modified simple ratio SAVI soil-adjusted vegetation index OSAVI optimized soil-adjusted vegetation index GEMI global environmental monitoring index EVI enhanced vegetation index GLCM grey-level co-occurrencematrix Pij is theprobability of values i and j occurring in adjacent pixels in the original image within the window defining the neighborhood i refers to the digital number(DN) value of a target pixel j is the DN value of its immediate neighbor and N is the number of grey levels


OOB data but are also used to measure the importance of thevariables [30]

In RF modeling there are two important training pa-rameters that need specification ntree is the number of treesto grow in the forest and mtry is the number of randomlyselected variables used in each node of the tree A good RFmodel which is built from the desirable values of ntree andmtry will have a low root mean square error (RMSE) To findthe ntree value that corresponds to a desirable predictordifferent ntree values varying from 50 to 1000 with an in-terval of 50 were tested e final ntree value was selectedbased on the stability of the RMSE (see Figure 2) To identifythe optimal mtry values we used the tuneRF function in therandomForest package

To evaluate the importance of each variable RF definestwo measures which are computed from the OOB data efirst measure is the percent increase in the mean square error(IncMSE) that was calculated for the prediction of eachtree [31] HigherIncMSE values indicate a more importantpredictor e second measure is the total decrease in nodeimpurities (IncNodePurity) which is the average of theresidual sum of squares over all trees when splitting thevariables at each node [31] Higher IncNodePurity valuesindicate a more important variable According to Strobl et al[61] the IncNodePurity method is biased and not recom-mended for use erefore in this study we only use the IncMSE measure to identify the importance of variables

Overall 10 RF models were built to determine the mostdesirable predictor for forest AGB estimation (Table 2)

27Model Validation For validation the original data wererandomly divided into two separate parts a training dataset(70) and a testing dataset (30) Each RF modelrsquos

performance was validated through a 10-fold cross-valida-tion e validation measures include the adjusted coeffi-cient of determination (R2

adj) and the root mean square error(RMSE)

3 Results

31 Tree AGB Estimation from Field Data Table 3 andFigure 3 show the results of tree AGB calculations for eachforest type at the plot level from field data measurementse results show that the forest AGB ranges from1832Mg haminus1 to 54386Mg haminus1 e average AGB esti-mated for Xuan Lien Nature Reserve was 15823Mg haminus1 forthe four forest types e MCBEV forests had the highestAGB followed by the BEV and SF forests Secondary forestshad the lowest AGB and were mostly mixed bamboo andevergreen forests or developed on abandoned agricultureland

In total 189 species from 55 families were recorded inthe field e five most dominant species were Castanopsisindica Engelhardia roxburghiana Ormosia sp Fokieniahodginsii and Archidendron balansae

32 Variable Importance and Variable Selection for the FinalRF Models Because models RF2 RF3 RF4 RF5 and RF6are a combination of spectral features vegetation indicestopographic data and texture features only models RF1 (allvariables) RF7 (spectral variables) RF8 (vegetation indicesrsquovariables) RF9 (topographic variables) and RF10 (texturevariables) were used to investigate the importance of thepredictor variables Each RF model was run 100 times todetermine the variation of each variablersquos importance

RF3RF9RF5RF1RF7RF6

RF4RF10RF2

RF8

60

70

80

90

100

110

120

130

140

0 50 100 150 200 250 300 350 400 450 500ntree

RMSE

(Mg

handash1

)

Figure 2 e RMSE is stable after ntree 300 for all 10 RF models


Table 2 Different RF models and their settings

Model Variable combination Number of variables ntree mtryRF1 Spectral topography vegetation indices and texture 52 300 26RF2 Spectral vegetation indices and texture 49 300 6RF3 Spectral topography and vegetation indices 16 300 12RF4 Spectral and vegetation indices 13 300 9RF5 Spectral and topography 7 300 3RF6 Spectral and texture 40 300 29RF7 Spectral 4 300 3RF8 Vegetation indices 9 300 7RF9 Topography 3 300 2RF10 Texture 36 300 12

Table 3 Summary statistics for the forest aboveground biomass (AGB) at the plot level

No Forest type No of plots Min AGB (Mg haminus1) Max AGB (Mg haminus1) Mean AGB (Mg haminus1) Standard deviation (Mg haminus1)1 MCBEV 64 4088 54388 25181 125432 BEV 29 6456 30376 15336 61613 SF 87 1832 25201 9101 5378

Totalaverage 180 1832 54388 15823 11336

0

50

100

150

200

250

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140dbh (cm)

Biom

ass (

Mg

handash1

)

Forest typesMCBEVBEV

SF

Figure 3 Distribution of the AGB of the field data by dbh in three forest types medium to high montane mixed coniferous and broadleafevergreen forests (MCBEVs) low montane broadleaf evergreen forests (BEVs) and secondary forests (SFs)


Con_RedDis_Red

Dis_GreenCon_GreenMean_Red

AspectMean_NIRHom_NIRInvD_NIRInvD_Red

GreenHom_Red

SD_NIRASM_NIR

BlueEnt_NIR

EVINIRDVI

GEMISAVI

RDVIOSAVINDVI

MSRRVI

Mean_BlueMean_Green

RedElevation

RF1

0 5 10 15 20 25 25 30 35IncMSE

(a)

Cor_BlueCor_RedSD_RedSD_Blue

Cor_NIREnt_Blue

ASM_BlueASM_Green

Ent_GreenDis_Blue

Hom_BlueInvD_BlueDis_Green

Hom_GreenCor_Green

InvD_GreenCon_NIR

Con_GreenCon_BlueCon_Red

Hom_NIRInvD_NIR

Dis_NIRSD_Green

SD_NIREnt_RedDis_Red

ASM_RedMean_NIRMean_Red

Ent_NIRASM_NIRInvD_RedHom_Red

Mean_BlueMean_Green

RF10

ndash5 0 5 10 15 20IncMSE

(b)

Blue

Green

NIR

Red

RF7

0 5 10 15 20 25 25 30 35IncMSE

(c)

Slope

Aspect

Elevation

RF9

0 10 20 30 40 50 60 70IncMSE

(d)

RDVI

NDVI

GEMI

MSR

SAVI

OSAVI

RVI

EVI

DVI

RF8

0 2 4 6 8 10 1412 16IncMSE

(e)

Figure 4 Ranking of each variablersquos importance measured in IncMSE by running the RF algorithm 100 times for different types ofpredictors all variables (model RF1mdashtop 30 variables) spectral reflectance (model RF7) vegetation indices (model RF8) topography(model RF9) and texture features (model RF10)


Among all the variables (model RF1) the most 10important variables are elevation spectral band 3 (red)the GLCM mean of the green and blue band RVI MSRNDVI OSAVI RDVI and SAVI (Figure 4 RF1) Whenusing spectral features as predictors the most importantband is band 3 (red) following by band 4 (NIR) band 2(green) and band 1 (blue) (Figure 4 RF7) Among thenine VIs used for AGB estimation there is no large

difference in the IncMSE values (Figure 4 RF8) forwhich DVI and EVI have highest values If we use onlytopographic data to predict AGB elevation has the largestinfluence followed by aspect and slope (Figure 4 RF9) Fortexture features the result from model RF10 reveals thatthe texture means of the green and blue bands are the twomost important variables for AGB estimation (Figure 4RF10)

Number of variables

RF1

65

70

75

80

85

1 10 20 30 40 51

RMSE

(Mg

ha)

(a)

Number of variables

RF2

80

82

84

86

88

90

1 10 20 30 40 51

RMSE

(Mg

ha)

(b)

Number of variables

RF3

1 4 8 12 1660

64

68

72

76

80

RMSE

(Mg

ha)

(c)

Number of variables

RF4

1 3 5 7 9 11 1375

77

79

81

83

85

RMSE

(Mg

ha)

(d)

Number of variables

RMSE

(Mg

ha)

RF5

65

67

69

71

73

75

1 2 3 4 5 76

(e)

Number of variables

RMSE

(Mg

ha)

RF6

70

74

78

82

86

90

1 10 155 20 25 30 35 40

(f )

Number of variables

RF7

70

72

74

76

78

80

1 2 3 4

RMSE

(Mg

ha)

(g)

Number of variables

RF8

95

97

99

101

103

105

1 2 3 54 6 7 8 9

RMSE

(Mg

ha)

(h)

Number of variables

RMSE

(Mg

ha)

RF9

65

66

67

68

69

70

1 2 3

(i)

Number of variables

RMSE

(Mg

ha)

80

83

86

89

92

95 RF10

1 11 166 21 26 31 36

(j)

Figure 5e number of variables (x-axis) used vs the RMSE (y-axis) of 10 RF models based on 100-times 10-fold cross-validatione redline indicates the mean RMSE the green and blue lines indicate the mean RMSE+ SD (standard deviation) or the mean RMSEminusSDrespectively


(RF1)0

50

100

150

200

250

300

350

400

450

50 100 150 200 250 300 350Predicted biomass (Mg handash1)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)y = 012 + 084middotxR2

adj = 069RMSE = 5253 (Mg handash1)

(a)

50 100 150 200 250 300Predicted biomass (Mg handash1)

(RF2)0

50

100

150

200

250

300

350

400

450

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

y = ndash66 + 14middotxR2


(b)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF3)50

100

150

200

250

300

350

400

450

500 y = ndash66 + 14middotxR2


(c)


(RF4)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500y = ndash66 + 14middotxR2


(d)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF5)50

100

150

200

250

300

350

400

450y = 15 + 09middotxR2


(e)


(RF6)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450y = 38 + 063middotxR2


(f )

Figure 6 Continued


To select variables for the final models knowing theimportance of each variable is not enough Each modelmust have an optimal number of variables which willimprove the modelrsquos accuracy To obtain the optimalvariables we used a 10-fold cross-validation method(running 100 times) e optimal variables for the 10models are shown in Figure 5

33 Performance of the Ten RF Models Figure 6 shows theresults of the ten models which are consistent with theprocessing done in Section 32 Model RF6 shows thelowest result followed by models RF10 RF8 and RF7Models RF4 and RF9 have similar accuracy and present aslightly lower result than model RF2 Model RF1 presentsthe highest result but requires 52 variables and resulted inoverfitting Although models RF3 and RF5 have a result

lower than model RF1 they require only 7 and 16 variablesrespectively and do not show over- or underfitting in theirresults

4 Discussion

e main objective of this study was to test the possibility ofusing SPOT-6 images for estimating the AGB of evergreenbroadleaf forests in Xuan Lien Nature Reserve using therandom forest algorithm Figure 4 (RF1) shows that ele-vation was the most important variable for predicting AGBis is mainly because vegetation types strongly vary alongthe altitudinal gradient within the study area [38] Similarlyother studies have proven that forest biomass has a sig-nificant relationship with vegetation types and elevation[3 8 62ndash64] e next three most important variables were


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF7)0

50

100

150

200

250

300

350

400

450y = 37 + 072middotxR2


(g)


(RF8)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

0

50

100

150

200

250

300

350

400

450y = 14 + 098middotxR2


(h)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF9)50

100

150

200

250

300

350

400

450y = 34 + 072middotxR2


(i)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500

(RF10)

y = 77 + 1middotxR2


(j)

Figure 6 Comparison of the predicted and observed AGB using the 10 RF models (RF1ndashRF10)e red line indicates the 1 1 linee blueline is the regression line


red reflectance and the two textures derived from the greenand blue bands

Among the four spectral bands the red band has thestrongest correlation with forest AGB (Figure 4 RF7) fol-lowing by NIR and the green and blue bands e possiblereason for this result is that red reflectance and NIR re-flectance are more sensitive to vegetation characteristics(eg tree species or stem volume) than other visible types ofreflectance [65] e performance of the regression wasimproved when we combined vegetation indices andortopographical features with spectral band reflectance Somesimilar findings were presented by Pandit et al [4] andAdam et al [66] who stated that using VIs improves resultsbecause VIs diminish the influence of environmental con-ditions and shadow effects on reflectance

In this study we have shown that the AGB and differentVIs have a significant correlation with each other e mostuseful VI for predicting forest AGB was DVI followed byEVI RVI OSAVI SAVI MSR GEMI NDVI and RDVIhowever the difference between these VIs was not veryhigh

Figure 5 clearly shows that the texture features (modelsRF10 RF6 and RF2mdashthose that included texture as apredictor) are less important than other features in AGBestimation In other words the accuracy of the model wasnot improved when using texture as an additional predictorfor AGB estimation is result is similar to that of Phamand Brabyn [8]

5 Conclusions

is study used the RF algorithm for modeling and pre-dicting the forest AGB in Xuan Lien Nature ReserveVietnam using VHR SPOT-6 data combined with field-based data e results showed a significant statistical re-lationship between the AGB and the SPOT-6 data eSPOT-6 data effectively predicted the AGB of the EB forestwith R2

adj 074 and RMSE 6124Mg haminus1 e accuracy ofAGB estimation was affected by many factors among whichelevation was indicated to be the most important for AGBmodels e random forest model selection of importantvariables showed that using elevation and vegetation indicesand spectral reflectance could significantly improve biomassestimations in evergreen broadleaf forests e RF algorithmis also suitable for estimating the AGB of evergreen broadleafforests

Based on the results of the study above some future workshould be considered For example when applying themodel from this study to different types of forests in differentecoregions the topography spectral reflectance and textureshould be taken into consideration Although the methodintroduced in this study is applicable to other forest eco-systems in Vietnam evaluating forest types is one of thenecessary impact predictors

It is also possible to use other RS data sources or machinelearning algorithms which may have better fit estimationserefore the next step of this study is to compare differentmachine learning techniques to predict forest AGB usingtwo optical sensor types (SPOT-6 and Sentinel-2 MSI)

In this study we found that elevation is one of the mostimportant predictors for forest AGB estimation ereforewhen designing a field survey for forest biomass estimationthe elevation should always be recorded

Data Availability

e data that support the findings of this study are availablefrom the corresponding author upon request

Conflicts of Interest

e authors declare no conflicts of interest

Acknowledgments

e authors are thankful to the support of various foun-dations e authors are also grateful to all the forest rangersand local members for their kind support and assistanceduring the fieldwork e authors would like to thank theVietnamese Government for their financial support andXuan Lien Nature Reserve for providing permission tocollect their field data e publication of this paper wassupported financially by the Open Access Grant Program ofthe German Research Foundation (DFG) and the OpenAccess Publication Fund of the Georg-August University ofGottingen

References

[1] F S Chapin P A Matson and H A Mooney Principles ofTerrestrial Ecosystem Ecology Springer New York NY USA2002

[2] R A Houghton ldquoAboveground forest biomass and the globalcarbon balancerdquo Global Change Biology vol 11 no 6pp 945ndash958 2005

[3] P Vicharnakorn R Shrestha M Nagai A Salam andS Kiratiprayoon ldquoCarbon stock assessment using remotesensing and forest inventory data in savannakhet Lao PDRrdquoRemote Sensing vol 6 no 6 pp 5452ndash5479 2014

[4] S Pandit S Tsuyuki and T Dube ldquoEstimating above-groundbiomass in sub-tropical buffer zone community forestsNepal using Sentinel 2 datardquo Remote Sensing vol 10 no 4p 601 2018

[5] M Zhang H Du G Zhou et al ldquoEstimating forest above-ground carbon storage in hang-jia-hu using landsat TMOLIdata and random forest modelrdquo Forests vol 10 no 11p 1004 2019

[6] S Eggleston L Buendia and K Miwa IPCC Guidelines forNational Greenhouse Gas Inventories IntergovernmentalPanel on Climate Change e Institute for Global Environ-mental Strategies Kanagawa Japan 2006

[7] D Lu ldquoe potential and challenge of remote sensing-basedbiomass estimationrdquo International Journal of Remote Sensingvol 27 no 7 pp 1297ndash1328 2007

[8] L T H Pham and L Brabyn ldquoMonitoring mangrove biomasschange in Vietnam using SPOT images and an object-basedapproach combined with machine learning algorithmsrdquoISPRS Journal of Photogrammetry and Remote Sensingvol 128 pp 86ndash97 2017

[9] L Du T Zhou Z Zou X Zhao K Huang and H WuldquoMapping forest biomass using remote sensing and national


forest inventory in Chinardquo Forests vol 5 no 6pp 1267ndash1283 2014

[10] D Deb J P Singh S Deb D Datta A Ghosh andR S Chaurasia ldquoAn alternative approach for estimating aboveground biomass using resourcesat-2 satellite data and artificialneural network in Bundelkhand region of Indiardquo Environ-mental Monitoring and Assessment vol 189 no 11 p 5762017

[11] K M Bergen and M C Dobson ldquoIntegration of remotelysensed radar imagery in modeling and mapping of forestbiomass and net primary productionrdquo Ecological Modellingvol 122 no 3 pp 257ndash274 1999

[12] O Hamdan I M Hasmadi and H K Aziz ldquoCombination ofSPOT-5 and ALOS PALSAR images in estimating above-ground biomass of lowland Dipterocarp forestrdquo IOP Con-ference Series Earth and Environmental Science vol 18p 012016 2014

[13] H Huang C Liu X Wang X Zhou and P Gong ldquoInte-gration of multi-resource remotely sensed data and allometricmodels for forest aboveground biomass estimation in ChinardquoRemote Sensing of Environment vol 221 pp 225ndash234 2019

[14] J M Chen ldquoEvaluation of vegetation indices and a modifiedSimple Ratio for boreal applicationsrdquo Canadian Journal ofRemote Sensing vol 22 no 3 pp 229ndash242 2014

[15] H-E Andersen J Strunk and H Temesgen ldquoUsing airbornelight detection and ranging as a sampling tool for estimatingforest biomass resources in the upper tanana valley of interiorAlaskardquo Western Journal of Applied Forestry vol 26 no 4pp 157ndash164 2011

[16] E Naeligsset ldquoPredicting forest stand characteristics with air-borne scanning laser using a practical two-stage procedureand field datardquo Remote Sensing of Environment vol 80pp 88ndash99 2002

[17] L I Duncanson K O Niemann and M A Wulder ldquoEsti-mating forest canopy height and terrain relief from GLASwaveform metricsrdquo Remote Sensing of Environment vol 114no 1 pp 138ndash154 2010

[18] M G Motlagh S B Kafaky A Mataji and R AkhavanldquoEstimating and mapping forest biomass using regressionmodels and Spot-6 images (case study hyrcanian forests ofnorth of Iran)rdquo Environmental Monitoring and Assessmentvol 190 no 6 p 352 2018

[19] Y Hirata N Furuya H Saito et al ldquoObject-based mapping ofaboveground biomass in tropical forests using LiDAR andvery-high-spatial-resolution satellite datardquo Remote Sensingvol 10 no 3 p 438 2018

[20] Y A Hussin H Gilani L van Leeuwen et al ldquoEvaluation ofobject-based image analysis techniques on very high-reso-lution satellite image for biomass estimation in a watershed ofhilly forest of Nepalrdquo Applied Geomatics vol 6 no 1pp 59ndash68 2014

[21] Y K Karna Y A Hussin H Gilani et al ldquoIntegration ofWorldView-2 and airborne LiDAR data for tree species levelcarbon stock mapping in Kayar Khola watershed NepalrdquoInternational Journal of Applied Earth Observation andGeoinformation vol 38 pp 280ndash291 2015

[22] W Li Z Niu X Liang et al ldquoGeostatistical modeling usingLiDAR-derived prior knowledge with SPOT-6 data to esti-mate temperate forest canopy cover and above-ground bio-mass via stratified random samplingrdquo International Journal ofApplied Earth Observation and Geoinformation vol 41pp 88ndash98 2015

[23] T W Gara A Murwira and H Ndaimani ldquoPredicting forestcarbon stocks from high resolution satellite data in dry forests

of Zimbabwe exploring the effect of the red-edge band inforest carbon stocks estimationrdquo Geocarto Internationalvol 31 no 2 pp 176ndash192 2015

[24] A Rosenqvist A Milne R Lucas M Imhoff and C DobsonldquoA review of remote sensing technology in support of thekyoto protocolrdquo 2019 httpswwwsciencedirectcomsciencearticleabspiiS1462901103000704

[25] B Kong H Yu R Du andQWang ldquoQuantitative estimationof biomass of alpine grasslands using hyperspectral remotesensingrdquo Rangeland Ecology amp Management vol 72 no 2pp 336ndash346 2019

[26] V Patil A Singh N Naik and S Unnikrishnan ldquoEstimationof mangrove carbon stocks by applying remote sensing andGIS techniquesrdquo Wetlands vol 35 no 4 pp 695ndash707 2015

[27] X Dou Y Yang and J Luo ldquoEstimating forest carbon fluxesusing machine learning techniques based on eddy covariancemeasurementsrdquo Sustainability vol 10 no 1 p 203 2018

[28] P M Lopez-Serrano C A Lopez-Sanchez J G Alvarez-Gonzalez and J Garcıa-Gutierrez ldquoA comparison of machinelearning techniques applied to landsat-5 TM spectral data forbiomass estimationrdquo Canadian Journal of Remote Sensingvol 42 no 6 pp 690ndash705 2016

[29] I Ali F Greifeneder J Stamenkovic M Neumann andC Notarnicola ldquoReview of machine learning approaches forbiomass and soil moisture retrievals from remote sensingdatardquo Remote Sensing vol 7 no 12 pp 16398ndash16421 2015

[30] L Breiman ldquoRandom forestsrdquo Machine Learning vol 45no 1 pp 5ndash32 2001

[31] D R Cutler T C Edwards K H Beard et al ldquoRandomforests for classification in ecologyrdquo Ecology vol 88 no 11pp 2783ndash2792 2007

[32] V T Phuong ldquoForest environment of Vietnam features offorest vegetation and soilsrdquo in Forest Environments In theMekong River Basins 1 H Sawada N Chappell J VLaFrankie A Shimizu and M Araki Eds pp 189ndash200Springer Tokyo Japan 2007

[33] Vietnam Administration of Forestry ldquoVietnam forest datasharing systemrdquo 2020 httpmapsvnforestgovvn

[34] C T Ha Vietnam National Forest Status of 2012 AnnualReport of Ministry of Agriculture and Rural DevelopmentHanoi Vietnam 2013

[35] Ministry of Agriculture and Rural Development ldquoVietnamrsquosmodified submission on reference levels for REDD+ resultsbased payments under UNFCCCrdquo 2016 httpsreddunfcccintfilesvietnam_frl_modified__submission_final_for_postingpdf

[36] A T N Dang S Nandy R Srinet N V Luong S Ghosh andA Senthil Kumar ldquoForest aboveground biomass estimationusing machine learning regression algorithm in Yok DonNational Park Vietnamrdquo Ecological Informatics vol 50pp 24ndash32 2019

[37] T T Le V C Le D T Bui et al A Feasibility Study for theEstablishment of Xuan Lien Nature Reserve Banh HoaProvince Vietnam BirdLife International Vietnam Pro-gramme Hanoi Vietnam 1999

[38] L T Trai L V Cham B D Tuyen et alA Feasibility Study forthe Establishment of Xuan Lien Nature Reserve anh HoaProvince Vietnam BirdLife International Vietnam Pro-gramme Hanoi Vietnamm 1999

[39] V T ai ldquoảm thực vật rừng Việt Nam (tren quan Ciểmhệ sinh thai)rdquo in Lần Bứ 2 Co Sửa ChữaKhoa học va k~ythuật Hanoi Vietnam 1978

[40] B Huy K Kralicek K P Poudel et al ldquoAllometric equationsfor estimating tree aboveground biomass in evergreen


broadleaf forests of Viet Namrdquo Forest Ecology and Man-agement vol 382 pp 193ndash205 2016

[41] T P Vu V X Nguyen T M L Nguyen and D T PhungAllometric Equations at National Scale for Tree Biomass As-sessment in Vietnam Part B4-Allometric Equations forBamboo Forests Vietnam Academy of Forest Sciences andUN-REDD Vietnam Hanoi Vietnam 2014

[42] Astrium Services SPOT 67 Imagery-User Guide[43] L H S Rotta E H Alcantara F S Y Watanabe

T W P Rodrigues and N N Imai ldquoAtmospheric correctionassessment of SPOT-6 image and its influence on models toestimate water column transparency in tropical reservoirrdquoRemote Sensing Applications Society and Environment vol 4pp 158ndash166 2016

[44] ldquoCartographic maps (5-m contour lines) rdquo Center of Serveyand Mapping Data Vietnam Department of Servey Mappingand Geographic Information Hanoi Vietnam 2013 httpmapsvnforestgovvn

[45] C J Tucker ldquoRed and photographic infrared linear combi-nations for monitoring vegetationrdquo Remote Sensing of En-vironment vol 8 no 2 pp 127ndash150 1979

[46] J-L Roujean and F-M Breon ldquoEstimating PAR absorbed byvegetation from bidirectional reflectance measurementsrdquoRemote Sensing of Environment vol 51 no 3 pp 375ndash3841995

[47] A R Huete ldquoA soil-adjusted vegetation index (SAVI)rdquo Re-mote Sensing of Environment vol 25 no 3 pp 295ndash309 1988

[48] G Rondeaux M Steven and F Baret ldquoOptimization of soil-adjusted vegetation indicesrdquo Remote Sensing of Environmentvol 55 no 2 pp 95ndash107 1996

[49] B Pinty and M M Verstraete ldquoGEMI a non-linear index tomonitor global vegetation from satellitesrdquo Vegetatio vol 101no 1 pp 15ndash20 1992

[50] M Hall-Beyer ldquoPractical guidelines for choosing GLCMtextures to use in landscape classification tasks over a range ofmoderate spatial scalesrdquo International Journal of RemoteSensing vol 38 no 5 pp 1312ndash1338 2017

[51] R M Haralick K Shanmugam and I H Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on Sys-tems Man and Cybernetics vol 6 pp 610ndash621 1973

[52] P Kapusta B R Bohard and D Ovals Plugin for QGIS (ver112) 2012 httpspluginsqgisorgpluginsrectovalDigitversion112

[53] L Cortes ldquoEstimation of above-ground forest biomass usinglandsat ETM+ aster GDEM and lidarrdquo Forest Research OpenAccess vol 03 no 2 2014

[54] M L Clark D A Roberts J J Ewel and D B Clark ldquoEs-timation of tropical rain forest aboveground biomass withsmall-footprint lidar and hyperspectral sensorsrdquo RemoteSensing of Environment vol 115 no 11 pp 2931ndash2942 2011

[55] R Muscarella S Kolyaie D C Morton J K ZimmermanM Uriarte and T Jucker ldquoEffects of topography on tropicalforest structure depend on climate contextrdquo Journal ofEcology vol 108 no 1 p 145 2019

[56] Q Wang R Punchi-Manage Z Lu et al ldquoEffects of to-pography on structuring species assemblages in a subtropicalforestrdquo Journal of Plant Ecology vol 280 p 47 2016

[57] S Ediriweera T Danaher and S Pathirana ldquoe influence oftopographic variation on forest structure in two woody plantcommunities a remote sensing approachrdquo Forest Systemsvol 25 no 1 p 049 2016

[58] R Core Team A Language and Environment for StatisticalComputing R Foundation for Statistical Computing ViennaAustria 2013

[59] S Pandit S Tsuyuki and T Dube ldquoLandscape-scaleAboveground biomass estimation in buffer zone communityforests of Central Nepal coupling in situ measurements withlandsat 8 satellite datardquo Remote Sensing vol 10 no 11p 1848 2018

[60] K Were D T Bui O B Dick and B R Singh ldquoA com-parative assessment of support vector regression artificialneural networks and random forests for predicting andmapping soil organic carbon stocks across an Afromontanelandscaperdquo Ecological Indicators vol 52 pp 394ndash403 2015

[61] C Strobl A-L Boulesteix A Zeileis and T Hothorn ldquoBias inrandom forest variable importance measures illustrationssources and a solutionrdquo BMC Bioinformatics vol 8 no 1p 25 2007

[62] Y Zhu K Liu L Liu S Wang and H Liu ldquoRetrieval ofmangrove aboveground biomass at the individual species levelwith WorldView-2 imagesrdquo Remote Sensing vol 7 no 9pp 12192ndash12214 2015

[63] A Ensslin G Rutten U Pommer R ZimmermannA Hemp andM Fischer ldquoEffects of elevation and land use onthe biomass of trees shrubs and herbs at Mount KilimanjarordquoEcosphere vol 6 no 3 p 45 2015

[64] C V de Castilho W E Magnusson R N O de Araujo et alldquoVariation in aboveground tree live biomass in a centralAmazonian Forest effects of soil and topographyrdquo ForestEcology and Management vol 234 pp 85ndash96 2006

[65] A R Huete ldquoRemote sensing for environmental monitoringrdquoin Environmental Monitoring and Characterization J FArtiola I L Pepper and M L Brusseau Eds pp 183ndash206Elsevier Science Amsterdam Netherland 2004

[66] E Adam O Mutanga E M Abdel-Rahman and R IsmailldquoEstimating standing biomass in papyrus (Cyperus papyrus L)swamp exploratory of in situ hyperspectral indices andrandom forest regressionrdquo International Journal of RemoteSensing vol 35 no 2 pp 693ndash714 2014


also allow one to estimate AGB by using the empirical re-lationships between the AGB and RS spectral bands vege-tation indices (VIs) and texture and topographicinformation with acceptable accuracy For example Motlaghet al [18] Hirata et al [19] Hussin et al [20] Karna et al[21] Li et al [22] and Gara et al [23] successfully predictedforest biomass based on those LiDAR and VHR sensorscombined with sufficient field data

Regardless of the RS data sources there are no RStechniques that are capable of providing a direct measure-ment of biomass As a result biomass prediction accuracyincreases when combined with field-sampled data especiallywhen using machine learning approaches to build biomassmodels [4 24ndash26] Machine learning algorithms allow one toanalyze a large number of predictor variables from remotesensing data thereby filling in the missing data and reducingthe error of the prediction models [27ndash29] A wide variety ofmachine learning algorithms have been employed to esti-mate AGB including an artificial neural network (ANN)K-nearest neighbor (KNN) support vector machine (SVM)and random forest (RF) In recent years RF has been widelyused to develop predictive models for AGB at the localregional and global areas because it can run efficiently onlarge datasets with a high accuracy Furthermore RF has theability to determine the importance of variables [30 31]

Evergreen broadleaf (EB) forests are estimated to covermore than 57 of national forests [32ndash35] and harbourapproximately 44 of the total forest carbon stock inVietnam [35] is forest type plays an important key role inecosystem carbon sequestration in Vietnam However inVietnam there have only been a few studies on forest carbonestimation with the integration of RS techniques especiallyusing VHR sensors For instance Dang et al [36] usedSentinel-2 satellite images combined with field-measureddata to estimate the AGB in Yok Don National Park eresearch of Pham and Brabyn [8] successfully proved theaccuracy of predicting the AGB of mangrove forests in CanGio (73) by integrating spectral information vegetationtype texture features and vegetation indices from SPOT-4and SPOT-5 images

e main objective of this study is (i) to test the ability ofthe spectral and vegetation indices and topographic andtexture features derived from SPOT-6 images to predictAGB in combination with field data using the RF algorithmand (ii) to identify the most desirable predictors for AGBestimation

2 Materials and Methods

21 Study Site is study was conducted at Xuan LienNature Reserve anh Hoa Vietnam located at 19∘52prime-20∘02primeN 104∘58prime-105∘15primeE which covers 23404 ha of twoforest types in the southwest of anh Hoa province isreserve is bordered by Cao River in the north the Nghe Anprovince in the south and west and the Ta Leo and BuKhongmountains and the confluence of Cao and Chu Riversto the east (Figure 1) e study area is situated in a belt ofmountains from Sam Neua in Laos to the uong Xuan andNhu Xuan districts in anh Hoa province which contain

some high peaks (eg Ta Leo (1400m) Bu Cho (1563m) BuHon Han (1208m) and an unnamed 1605m peak) emean temperature is about 23ndash24degC and the mean annualrainfall is approximately 1700ndash1900mm which occursmainly from May to October and accounts for 90 of thetotal annual rainfall [37] e main soil in the nature reserveis feralite soil feralite humus soil in the medium-highmountains (FH) feralite soil in the lowlands (F) and alluvialsoil (P) associated with streams or rivers and the valleybottom [37]

e vegetation in the study area was mainly closedevergreen broadleaf forest which was classified into threeforest types [38] based on the classification of ai VanTrung [39] e first forest type is distributed from mediumto high montane consisting of mixed coniferous andbroadleaf evergreen forests (MCBEV) between 800m and1605m (asl) is forest type is generally undisturbed anddominated by upper storeyed broadleaf tree species from thefamilies of Fagaceae Lauraceae Euphorbiaceae FabaceaeMagnoliaceae Dipterocarpaceae and Sapotaceae [37 38]e second forest type is located in the low montanebroadleaf evergreen forests (BEV) which are distributedunder 800m asl and have been weakly impacted by humanactivities Common species include Leguminosae Euphor-biaeceae Lauraceae Rutaceae Rosaceae and Meliaceae[37 38] e final forest type is secondary forests (SF) whichare mainly a mix of Neohouzeaua dullooa Dendrocalamuspatellaris Bambusa sp and broadleaf evergreen forest[37 38]

22 Field Data Collection Field surveys were conducted inFebruary 2015 In total 180 plots (20mtimes 25m) were sampledand inventoried ese plots were randomly generated inArcGIS 104 and then located in the field using a GPS devicewith errors up to 5m Within the plots the diameter at breastheight (dbh) and the total height (h) of each living tree withdbh greater than 5 cm were measured using a diameter tapeand a Vertex Hypsometer respectively Tree species were alsorecorded for each measured tree

23 Aboveground Biomass Estimation We considered onlythe aboveground living tree biomass for carbon estimationsAboveground biomass (AGB) was estimated as the sum ofthe individual components (stumps stems bark branchesseeds and foliage) of the individual living trees that werepredicted using appropriate allometric equations [6] eseallometric equations were carefully chosen depending on theforest types and the tree or bamboo species available in theinput dataset For the evergreen broadleaf forests we usedthe biomass equation developed by Huy et al [40] whichwas specifically developed for evergreen broadleaf forests inthe North Central region of Vietnam (1) For bambooforests we opted for the equation from Vu et al [41] whichwas developed for bamboo forests at a national scale (2) Formixed forests of bamboo and evergreen broadleaf forestsboth (1) and (2) were used to estimate the total biomass Allof the selected equations above are based on the treebamboodiameters (dbh) and total heights (h)


AGBt 2532449 times db h2h1113872 1113873

095102R2

09555 (1)


times h02778

R2

07810 (2)






0 2 4 6 81Km

1cm = 1km

Very rich EB forest


Bamboo


Agriculture


20deg5primeN

104deg50primeE




20deg0primeN

19deg5

5primeN

22degN

18degN

14degN

10degN

19deg5

0primeN

20deg5primeN

20deg0primeN

19deg5

5primeN

19deg5

0primeN

N

Cua Dat Lake










TopographyDEM

6 metersSlopeAspect

Vegetation indices


B4 + B3

1113968[46]


1113968+ 1) [14]



4 minus B23) + 15B4 + 05B3B4 + B3 + 05











Nminus1ij0(P2

ij) [50 51]


Vari lowastVarj

1113969] [50 51]


Pij











3 Results




RF3RF9RF5RF1RF7RF6

RF4RF10RF2

RF8

60

70

80

90

100

110

120

130

140

0 50 100 150 200 250 300 350 400 450 500ntree

RMSE

(Mg

handash1

)







Totalaverage 180 1832 54388 15823 11336

0

50

100

150

200

250

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140dbh (cm)

Biom

ass (

Mg

handash1

)


SF



Con_RedDis_Red



GreenHom_Red

SD_NIRASM_NIR

BlueEnt_NIR

EVINIRDVI

GEMISAVI

RDVIOSAVINDVI

MSRRVI

Mean_BlueMean_Green

RedElevation

RF1

0 5 10 15 20 25 25 30 35IncMSE

(a)


Cor_NIREnt_Blue

ASM_BlueASM_Green

Ent_GreenDis_Blue


Hom_GreenCor_Green

InvD_GreenCon_NIR


Hom_NIRInvD_NIR

Dis_NIRSD_Green




Mean_BlueMean_Green

RF10


(b)

Blue

Green

NIR

Red

RF7

0 5 10 15 20 25 25 30 35IncMSE

(c)

Slope

Aspect

Elevation

RF9

0 10 20 30 40 50 60 70IncMSE

(d)

RDVI

NDVI

GEMI

MSR

SAVI

OSAVI

RVI

EVI

DVI

RF8

0 2 4 6 8 10 1412 16IncMSE

(e)





Number of variables

RF1

65

70

75

80

85

1 10 20 30 40 51

RMSE

(Mg

ha)

(a)

Number of variables

RF2

80

82

84

86

88

90

1 10 20 30 40 51

RMSE

(Mg

ha)

(b)

Number of variables

RF3

1 4 8 12 1660

64

68

72

76

80

RMSE

(Mg

ha)

(c)

Number of variables

RF4

1 3 5 7 9 11 1375

77

79

81

83

85

RMSE

(Mg

ha)

(d)

Number of variables

RMSE

(Mg

ha)

RF5

65

67

69

71

73

75

1 2 3 4 5 76

(e)

Number of variables

RMSE

(Mg

ha)

RF6

70

74

78

82

86

90

1 10 155 20 25 30 35 40

(f )

Number of variables

RF7

70

72

74

76

78

80

1 2 3 4

RMSE

(Mg

ha)

(g)

Number of variables

RF8

95

97

99

101

103

105

1 2 3 54 6 7 8 9

RMSE

(Mg

ha)

(h)

Number of variables

RMSE

(Mg

ha)

RF9

65

66

67

68

69

70

1 2 3

(i)

Number of variables

RMSE

(Mg

ha)

80

83

86

89

92

95 RF10

1 11 166 21 26 31 36

(j)



(RF1)0

50

100

150

200

250

300

350

400

450


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)y = 012 + 084middotxR2


(a)


(RF2)0

50

100

150

200

250

300

350

400

450

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)



(b)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF3)50

100

150

200

250

300

350

400

450



(c)


(RF4)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550



(d)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF5)50

100

150

200

250

300

350

400

450y = 15 + 09middotxR2


(e)


(RF6)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450y = 38 + 063middotxR2


(f )

Figure 6 Continued





4 Discussion



Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF7)0

50

100

150

200

250

300

350

400

450y = 37 + 072middotxR2


(g)


(RF8)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

0

50

100

150

200

250

300

350

400

450y = 14 + 098middotxR2


(h)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF9)50

100

150

200

250

300

350

400

450y = 34 + 072middotxR2


(i)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500

(RF10)

y = 77 + 1middotxR2


(j)







5 Conclusions






Data Availability




Acknowledgments


References









































































AGBt 2532449 times db h2h1113872 1113873

095102R2

09555 (1)


times h02778

R2

07810 (2)






0 2 4 6 81Km

1cm = 1km

Very rich EB forest


Bamboo


Agriculture


20deg5primeN

104deg50primeE




20deg0primeN

19deg5

5primeN

22degN

18degN

14degN

10degN

19deg5

0primeN

20deg5primeN

20deg0primeN

19deg5

5primeN

19deg5

0primeN

N

Cua Dat Lake










TopographyDEM

6 metersSlopeAspect

Vegetation indices


B4 + B3

1113968[46]


1113968+ 1) [14]



4 minus B23) + 15B4 + 05B3B4 + B3 + 05











Nminus1ij0(P2

ij) [50 51]


Vari lowastVarj

1113969] [50 51]


Pij











3 Results




RF3RF9RF5RF1RF7RF6

RF4RF10RF2

RF8

60

70

80

90

100

110

120

130

140

0 50 100 150 200 250 300 350 400 450 500ntree

RMSE

(Mg

handash1

)







Totalaverage 180 1832 54388 15823 11336

0

50

100

150

200

250

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140dbh (cm)

Biom

ass (

Mg

handash1

)


SF



Con_RedDis_Red



GreenHom_Red

SD_NIRASM_NIR

BlueEnt_NIR

EVINIRDVI

GEMISAVI

RDVIOSAVINDVI

MSRRVI

Mean_BlueMean_Green

RedElevation

RF1

0 5 10 15 20 25 25 30 35IncMSE

(a)


Cor_NIREnt_Blue

ASM_BlueASM_Green

Ent_GreenDis_Blue


Hom_GreenCor_Green

InvD_GreenCon_NIR


Hom_NIRInvD_NIR

Dis_NIRSD_Green




Mean_BlueMean_Green

RF10


(b)

Blue

Green

NIR

Red

RF7

0 5 10 15 20 25 25 30 35IncMSE

(c)

Slope

Aspect

Elevation

RF9

0 10 20 30 40 50 60 70IncMSE

(d)

RDVI

NDVI

GEMI

MSR

SAVI

OSAVI

RVI

EVI

DVI

RF8

0 2 4 6 8 10 1412 16IncMSE

(e)





Number of variables

RF1

65

70

75

80

85

1 10 20 30 40 51

RMSE

(Mg

ha)

(a)

Number of variables

RF2

80

82

84

86

88

90

1 10 20 30 40 51

RMSE

(Mg

ha)

(b)

Number of variables

RF3

1 4 8 12 1660

64

68

72

76

80

RMSE

(Mg

ha)

(c)

Number of variables

RF4

1 3 5 7 9 11 1375

77

79

81

83

85

RMSE

(Mg

ha)

(d)

Number of variables

RMSE

(Mg

ha)

RF5

65

67

69

71

73

75

1 2 3 4 5 76

(e)

Number of variables

RMSE

(Mg

ha)

RF6

70

74

78

82

86

90

1 10 155 20 25 30 35 40

(f )

Number of variables

RF7

70

72

74

76

78

80

1 2 3 4

RMSE

(Mg

ha)

(g)

Number of variables

RF8

95

97

99

101

103

105

1 2 3 54 6 7 8 9

RMSE

(Mg

ha)

(h)

Number of variables

RMSE

(Mg

ha)

RF9

65

66

67

68

69

70

1 2 3

(i)

Number of variables

RMSE

(Mg

ha)

80

83

86

89

92

95 RF10

1 11 166 21 26 31 36

(j)



(RF1)0

50

100

150

200

250

300

350

400

450


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)y = 012 + 084middotxR2


(a)


(RF2)0

50

100

150

200

250

300

350

400

450

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)



(b)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF3)50

100

150

200

250

300

350

400

450



(c)


(RF4)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550



(d)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF5)50

100

150

200

250

300

350

400

450y = 15 + 09middotxR2


(e)


(RF6)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450y = 38 + 063middotxR2


(f )

Figure 6 Continued





4 Discussion



Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF7)0

50

100

150

200

250

300

350

400

450y = 37 + 072middotxR2


(g)


(RF8)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

0

50

100

150

200

250

300

350

400

450y = 14 + 098middotxR2


(h)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF9)50

100

150

200

250

300

350

400

450y = 34 + 072middotxR2


(i)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500

(RF10)

y = 77 + 1middotxR2


(j)







5 Conclusions






Data Availability




Acknowledgments


References















































































TopographyDEM

6 metersSlopeAspect

Vegetation indices


B4 + B3

1113968[46]


1113968+ 1) [14]



4 minus B23) + 15B4 + 05B3B4 + B3 + 05











Nminus1ij0(P2

ij) [50 51]


Vari lowastVarj

1113969] [50 51]


Pij











3 Results




RF3RF9RF5RF1RF7RF6

RF4RF10RF2

RF8

60

70

80

90

100

110

120

130

140

0 50 100 150 200 250 300 350 400 450 500ntree

RMSE

(Mg

handash1

)







Totalaverage 180 1832 54388 15823 11336

0

50

100

150

200

250

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140dbh (cm)

Biom

ass (

Mg

handash1

)


SF



Con_RedDis_Red



GreenHom_Red

SD_NIRASM_NIR

BlueEnt_NIR

EVINIRDVI

GEMISAVI

RDVIOSAVINDVI

MSRRVI

Mean_BlueMean_Green

RedElevation

RF1

0 5 10 15 20 25 25 30 35IncMSE

(a)


Cor_NIREnt_Blue

ASM_BlueASM_Green

Ent_GreenDis_Blue


Hom_GreenCor_Green

InvD_GreenCon_NIR


Hom_NIRInvD_NIR

Dis_NIRSD_Green




Mean_BlueMean_Green

RF10


(b)

Blue

Green

NIR

Red

RF7

0 5 10 15 20 25 25 30 35IncMSE

(c)

Slope

Aspect

Elevation

RF9

0 10 20 30 40 50 60 70IncMSE

(d)

RDVI

NDVI

GEMI

MSR

SAVI

OSAVI

RVI

EVI

DVI

RF8

0 2 4 6 8 10 1412 16IncMSE

(e)





Number of variables

RF1

65

70

75

80

85

1 10 20 30 40 51

RMSE

(Mg

ha)

(a)

Number of variables

RF2

80

82

84

86

88

90

1 10 20 30 40 51

RMSE

(Mg

ha)

(b)

Number of variables

RF3

1 4 8 12 1660

64

68

72

76

80

RMSE

(Mg

ha)

(c)

Number of variables

RF4

1 3 5 7 9 11 1375

77

79

81

83

85

RMSE

(Mg

ha)

(d)

Number of variables

RMSE

(Mg

ha)

RF5

65

67

69

71

73

75

1 2 3 4 5 76

(e)

Number of variables

RMSE

(Mg

ha)

RF6

70

74

78

82

86

90

1 10 155 20 25 30 35 40

(f )

Number of variables

RF7

70

72

74

76

78

80

1 2 3 4

RMSE

(Mg

ha)

(g)

Number of variables

RF8

95

97

99

101

103

105

1 2 3 54 6 7 8 9

RMSE

(Mg

ha)

(h)

Number of variables

RMSE

(Mg

ha)

RF9

65

66

67

68

69

70

1 2 3

(i)

Number of variables

RMSE

(Mg

ha)

80

83

86

89

92

95 RF10

1 11 166 21 26 31 36

(j)



(RF1)0

50

100

150

200

250

300

350

400

450


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)y = 012 + 084middotxR2


(a)


(RF2)0

50

100

150

200

250

300

350

400

450

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)



(b)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF3)50

100

150

200

250

300

350

400

450



(c)


(RF4)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550



(d)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF5)50

100

150

200

250

300

350

400

450y = 15 + 09middotxR2


(e)


(RF6)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450y = 38 + 063middotxR2


(f )

Figure 6 Continued





4 Discussion



Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF7)0

50

100

150

200

250

300

350

400

450y = 37 + 072middotxR2


(g)


(RF8)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

0

50

100

150

200

250

300

350

400

450y = 14 + 098middotxR2


(h)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF9)50

100

150

200

250

300

350

400

450y = 34 + 072middotxR2


(i)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500

(RF10)

y = 77 + 1middotxR2


(j)







5 Conclusions






Data Availability




Acknowledgments


References
















































































3 Results




RF3RF9RF5RF1RF7RF6

RF4RF10RF2

RF8

60

70

80

90

100

110

120

130

140

0 50 100 150 200 250 300 350 400 450 500ntree

RMSE

(Mg

handash1

)







Totalaverage 180 1832 54388 15823 11336

0

50

100

150

200

250

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140dbh (cm)

Biom

ass (

Mg

handash1

)


SF



Con_RedDis_Red



GreenHom_Red

SD_NIRASM_NIR

BlueEnt_NIR

EVINIRDVI

GEMISAVI

RDVIOSAVINDVI

MSRRVI

Mean_BlueMean_Green

RedElevation

RF1

0 5 10 15 20 25 25 30 35IncMSE

(a)


Cor_NIREnt_Blue

ASM_BlueASM_Green

Ent_GreenDis_Blue


Hom_GreenCor_Green

InvD_GreenCon_NIR


Hom_NIRInvD_NIR

Dis_NIRSD_Green




Mean_BlueMean_Green

RF10


(b)

Blue

Green

NIR

Red

RF7

0 5 10 15 20 25 25 30 35IncMSE

(c)

Slope

Aspect

Elevation

RF9

0 10 20 30 40 50 60 70IncMSE

(d)

RDVI

NDVI

GEMI

MSR

SAVI

OSAVI

RVI

EVI

DVI

RF8

0 2 4 6 8 10 1412 16IncMSE

(e)





Number of variables

RF1

65

70

75

80

85

1 10 20 30 40 51

RMSE

(Mg

ha)

(a)

Number of variables

RF2

80

82

84

86

88

90

1 10 20 30 40 51

RMSE

(Mg

ha)

(b)

Number of variables

RF3

1 4 8 12 1660

64

68

72

76

80

RMSE

(Mg

ha)

(c)

Number of variables

RF4

1 3 5 7 9 11 1375

77

79

81

83

85

RMSE

(Mg

ha)

(d)

Number of variables

RMSE

(Mg

ha)

RF5

65

67

69

71

73

75

1 2 3 4 5 76

(e)

Number of variables

RMSE

(Mg

ha)

RF6

70

74

78

82

86

90

1 10 155 20 25 30 35 40

(f )

Number of variables

RF7

70

72

74

76

78

80

1 2 3 4

RMSE

(Mg

ha)

(g)

Number of variables

RF8

95

97

99

101

103

105

1 2 3 54 6 7 8 9

RMSE

(Mg

ha)

(h)

Number of variables

RMSE

(Mg

ha)

RF9

65

66

67

68

69

70

1 2 3

(i)

Number of variables

RMSE

(Mg

ha)

80

83

86

89

92

95 RF10

1 11 166 21 26 31 36

(j)



(RF1)0

50

100

150

200

250

300

350

400

450


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)y = 012 + 084middotxR2


(a)


(RF2)0

50

100

150

200

250

300

350

400

450

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)



(b)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF3)50

100

150

200

250

300

350

400

450



(c)


(RF4)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550



(d)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF5)50

100

150

200

250

300

350

400

450y = 15 + 09middotxR2


(e)


(RF6)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450y = 38 + 063middotxR2


(f )

Figure 6 Continued





4 Discussion



Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF7)0

50

100

150

200

250

300

350

400

450y = 37 + 072middotxR2


(g)


(RF8)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

0

50

100

150

200

250

300

350

400

450y = 14 + 098middotxR2


(h)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF9)50

100

150

200

250

300

350

400

450y = 34 + 072middotxR2


(i)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500

(RF10)

y = 77 + 1middotxR2


(j)







5 Conclusions






Data Availability




Acknowledgments


References













































































Totalaverage 180 1832 54388 15823 11336

0

50

100

150

200

250

0 10 20 30 40 50 60 70 80 90 100 110 120 130 140dbh (cm)

Biom

ass (

Mg

handash1

)


SF



Con_RedDis_Red



GreenHom_Red

SD_NIRASM_NIR

BlueEnt_NIR

EVINIRDVI

GEMISAVI

RDVIOSAVINDVI

MSRRVI

Mean_BlueMean_Green

RedElevation

RF1

0 5 10 15 20 25 25 30 35IncMSE

(a)


Cor_NIREnt_Blue

ASM_BlueASM_Green

Ent_GreenDis_Blue


Hom_GreenCor_Green

InvD_GreenCon_NIR


Hom_NIRInvD_NIR

Dis_NIRSD_Green




Mean_BlueMean_Green

RF10


(b)

Blue

Green

NIR

Red

RF7

0 5 10 15 20 25 25 30 35IncMSE

(c)

Slope

Aspect

Elevation

RF9

0 10 20 30 40 50 60 70IncMSE

(d)

RDVI

NDVI

GEMI

MSR

SAVI

OSAVI

RVI

EVI

DVI

RF8

0 2 4 6 8 10 1412 16IncMSE

(e)





Number of variables

RF1

65

70

75

80

85

1 10 20 30 40 51

RMSE

(Mg

ha)

(a)

Number of variables

RF2

80

82

84

86

88

90

1 10 20 30 40 51

RMSE

(Mg

ha)

(b)

Number of variables

RF3

1 4 8 12 1660

64

68

72

76

80

RMSE

(Mg

ha)

(c)

Number of variables

RF4

1 3 5 7 9 11 1375

77

79

81

83

85

RMSE

(Mg

ha)

(d)

Number of variables

RMSE

(Mg

ha)

RF5

65

67

69

71

73

75

1 2 3 4 5 76

(e)

Number of variables

RMSE

(Mg

ha)

RF6

70

74

78

82

86

90

1 10 155 20 25 30 35 40

(f )

Number of variables

RF7

70

72

74

76

78

80

1 2 3 4

RMSE

(Mg

ha)

(g)

Number of variables

RF8

95

97

99

101

103

105

1 2 3 54 6 7 8 9

RMSE

(Mg

ha)

(h)

Number of variables

RMSE

(Mg

ha)

RF9

65

66

67

68

69

70

1 2 3

(i)

Number of variables

RMSE

(Mg

ha)

80

83

86

89

92

95 RF10

1 11 166 21 26 31 36

(j)



(RF1)0

50

100

150

200

250

300

350

400

450


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)y = 012 + 084middotxR2


(a)


(RF2)0

50

100

150

200

250

300

350

400

450

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)



(b)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF3)50

100

150

200

250

300

350

400

450



(c)


(RF4)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550



(d)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF5)50

100

150

200

250

300

350

400

450y = 15 + 09middotxR2


(e)


(RF6)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450y = 38 + 063middotxR2


(f )

Figure 6 Continued





4 Discussion



Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF7)0

50

100

150

200

250

300

350

400

450y = 37 + 072middotxR2


(g)


(RF8)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

0

50

100

150

200

250

300

350

400

450y = 14 + 098middotxR2


(h)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF9)50

100

150

200

250

300

350

400

450y = 34 + 072middotxR2


(i)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500

(RF10)

y = 77 + 1middotxR2


(j)







5 Conclusions






Data Availability




Acknowledgments


References









































































Con_RedDis_Red



GreenHom_Red

SD_NIRASM_NIR

BlueEnt_NIR

EVINIRDVI

GEMISAVI

RDVIOSAVINDVI

MSRRVI

Mean_BlueMean_Green

RedElevation

RF1

0 5 10 15 20 25 25 30 35IncMSE

(a)


Cor_NIREnt_Blue

ASM_BlueASM_Green

Ent_GreenDis_Blue


Hom_GreenCor_Green

InvD_GreenCon_NIR


Hom_NIRInvD_NIR

Dis_NIRSD_Green




Mean_BlueMean_Green

RF10


(b)

Blue

Green

NIR

Red

RF7

0 5 10 15 20 25 25 30 35IncMSE

(c)

Slope

Aspect

Elevation

RF9

0 10 20 30 40 50 60 70IncMSE

(d)

RDVI

NDVI

GEMI

MSR

SAVI

OSAVI

RVI

EVI

DVI

RF8

0 2 4 6 8 10 1412 16IncMSE

(e)





Number of variables

RF1

65

70

75

80

85

1 10 20 30 40 51

RMSE

(Mg

ha)

(a)

Number of variables

RF2

80

82

84

86

88

90

1 10 20 30 40 51

RMSE

(Mg

ha)

(b)

Number of variables

RF3

1 4 8 12 1660

64

68

72

76

80

RMSE

(Mg

ha)

(c)

Number of variables

RF4

1 3 5 7 9 11 1375

77

79

81

83

85

RMSE

(Mg

ha)

(d)

Number of variables

RMSE

(Mg

ha)

RF5

65

67

69

71

73

75

1 2 3 4 5 76

(e)

Number of variables

RMSE

(Mg

ha)

RF6

70

74

78

82

86

90

1 10 155 20 25 30 35 40

(f )

Number of variables

RF7

70

72

74

76

78

80

1 2 3 4

RMSE

(Mg

ha)

(g)

Number of variables

RF8

95

97

99

101

103

105

1 2 3 54 6 7 8 9

RMSE

(Mg

ha)

(h)

Number of variables

RMSE

(Mg

ha)

RF9

65

66

67

68

69

70

1 2 3

(i)

Number of variables

RMSE

(Mg

ha)

80

83

86

89

92

95 RF10

1 11 166 21 26 31 36

(j)



(RF1)0

50

100

150

200

250

300

350

400

450


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)y = 012 + 084middotxR2


(a)


(RF2)0

50

100

150

200

250

300

350

400

450

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)



(b)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF3)50

100

150

200

250

300

350

400

450



(c)


(RF4)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550



(d)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF5)50

100

150

200

250

300

350

400

450y = 15 + 09middotxR2


(e)


(RF6)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450y = 38 + 063middotxR2


(f )

Figure 6 Continued





4 Discussion



Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF7)0

50

100

150

200

250

300

350

400

450y = 37 + 072middotxR2


(g)


(RF8)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

0

50

100

150

200

250

300

350

400

450y = 14 + 098middotxR2


(h)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF9)50

100

150

200

250

300

350

400

450y = 34 + 072middotxR2


(i)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500

(RF10)

y = 77 + 1middotxR2


(j)







5 Conclusions






Data Availability




Acknowledgments


References











































































Number of variables

RF1

65

70

75

80

85

1 10 20 30 40 51

RMSE

(Mg

ha)

(a)

Number of variables

RF2

80

82

84

86

88

90

1 10 20 30 40 51

RMSE

(Mg

ha)

(b)

Number of variables

RF3

1 4 8 12 1660

64

68

72

76

80

RMSE

(Mg

ha)

(c)

Number of variables

RF4

1 3 5 7 9 11 1375

77

79

81

83

85

RMSE

(Mg

ha)

(d)

Number of variables

RMSE

(Mg

ha)

RF5

65

67

69

71

73

75

1 2 3 4 5 76

(e)

Number of variables

RMSE

(Mg

ha)

RF6

70

74

78

82

86

90

1 10 155 20 25 30 35 40

(f )

Number of variables

RF7

70

72

74

76

78

80

1 2 3 4

RMSE

(Mg

ha)

(g)

Number of variables

RF8

95

97

99

101

103

105

1 2 3 54 6 7 8 9

RMSE

(Mg

ha)

(h)

Number of variables

RMSE

(Mg

ha)

RF9

65

66

67

68

69

70

1 2 3

(i)

Number of variables

RMSE

(Mg

ha)

80

83

86

89

92

95 RF10

1 11 166 21 26 31 36

(j)



(RF1)0

50

100

150

200

250

300

350

400

450


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)y = 012 + 084middotxR2


(a)


(RF2)0

50

100

150

200

250

300

350

400

450

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)



(b)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF3)50

100

150

200

250

300

350

400

450



(c)


(RF4)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550



(d)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF5)50

100

150

200

250

300

350

400

450y = 15 + 09middotxR2


(e)


(RF6)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450y = 38 + 063middotxR2


(f )

Figure 6 Continued





4 Discussion



Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF7)0

50

100

150

200

250

300

350

400

450y = 37 + 072middotxR2


(g)


(RF8)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

0

50

100

150

200

250

300

350

400

450y = 14 + 098middotxR2


(h)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF9)50

100

150

200

250

300

350

400

450y = 34 + 072middotxR2


(i)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500

(RF10)

y = 77 + 1middotxR2


(j)







5 Conclusions






Data Availability




Acknowledgments


References









































































(RF1)0

50

100

150

200

250

300

350

400

450


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)y = 012 + 084middotxR2


(a)


(RF2)0

50

100

150

200

250

300

350

400

450

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)



(b)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF3)50

100

150

200

250

300

350

400

450



(c)


(RF4)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550



(d)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF5)50

100

150

200

250

300

350

400

450y = 15 + 09middotxR2


(e)


(RF6)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450y = 38 + 063middotxR2


(f )

Figure 6 Continued





4 Discussion



Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF7)0

50

100

150

200

250

300

350

400

450y = 37 + 072middotxR2


(g)


(RF8)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

0

50

100

150

200

250

300

350

400

450y = 14 + 098middotxR2


(h)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF9)50

100

150

200

250

300

350

400

450y = 34 + 072middotxR2


(i)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500

(RF10)

y = 77 + 1middotxR2


(j)







5 Conclusions






Data Availability




Acknowledgments


References












































































4 Discussion



Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF7)0

50

100

150

200

250

300

350

400

450y = 37 + 072middotxR2


(g)


(RF8)

Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

0

50

100

150

200

250

300

350

400

450y = 14 + 098middotxR2


(h)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

(RF9)50

100

150

200

250

300

350

400

450y = 34 + 072middotxR2


(i)


Fiel

dminusm

easu

red

biom

ass (

Mg

handash1

)

50

100

150

200

250

300

350

400

450

550

500

(RF10)

y = 77 + 1middotxR2


(j)







5 Conclusions






Data Availability




Acknowledgments


References













































































5 Conclusions






Data Availability




Acknowledgments


References









































































EstimatingtheAbovegroundBiomassofanEvergreenBroadleaf...

Documents

Transcript of EstimatingtheAbovegroundBiomassofanEvergreenBroadleaf...