Use of ALS, Airborne CIR and ALOS AVNIR-2 data for estimating tropical forest attributes in Lao PDR

ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786

Contents lists available at SciVerse ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

journal homepage: www.elsevier .com/ locate/ isprs jprs

Use of ALS, Airborne CIR and ALOS AVNIR-2 data for estimating tropicalforest attributes in Lao PDR

Zhengyang Hou a,b,⇑, Qing Xu b, Timo Tokola b

a European Forest Institute (EFI), Torikatu 34, FI-80100 Joensuu, Finlandb University of Eastern Finland, Faculty of Science and Forestry, School of Forest Sciences, P.O. Box 111, FI-80101 Joensuu, Finland

a r t i c l e i n f o

Article history:Received 7 February 2011Received in revised form 2 September 2011Accepted 3 September 2011Available online 1 October 2011

Keywords:ALSAirborne CIRALOS AVNIR-2Tropical forestForest monitoring

0924-2716/$ - see front matter � 2011 Internationaldoi:10.1016/j.isprsjprs.2011.09.005

⇑ Corresponding author at: European Forest InstitutJoensuu, Finland. Tel.: +35 8417239419; fax: +35 810

E-mail addresses: [email protected], houzhen

a b s t r a c t

In this study, the potential of remote sensing in tropical forests is examined in relation to thediversification of sensors. We report here on the comparison of alternative methods that use multisourcedata from Airborne Laser Scanning (ALS), Airborne CIR and ALOS AVNIR-2 to estimate stem volume andbasal area, in Laos. Multivariate linear regression analyses with stepwise selection of predictors wereimplemented for modelling. The predictors of ALS metrics were calculated by means of the canopy heightdistribution approach, while predictors from both spectral and textual features were respectively gener-ated for Airborne CIR and ALOS AVNIR-2 data. With respect to the estimation capacity from individualdata sources after leave-one-out cross-validation, the ALS data proved superior, with the lowest RMSEof 36.92% for stem volume and 47.35% for basal area, whereas Airborne CIR and ALOS AVNIR-2 remainedat similar accuracy levels, but fell well behind the ALS data. By integrating ALS metrics with otherpredictors from Airborne CIR or ALOS AVNIR-2, hybrid modelling was further tested respectively. Theresults showed that only the hybrid model for stem volume involving ALS and Airborne CIR improvedthe accuracy of 1.9% in terms of relative RMSE than that of using ALS alone.� 2011 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS) Published by Elsevier

B.V. All rights reserved.

1. Introduction there are two established approaches commonly adopted in prac-

Remote sensing techniques have been explicitly advocated bythe Copenhagen Accord and its guidelines on methodologies facil-itating REDD (Reducing Emissions from Deforestation and ForestDegradation in Developing Countries) for estimating forest carbonstocks in global context (UNFCCC, 2010). By means of remotesensing, many of the forest attributes of interest are retrievableat varying accuracy levels with due cost-effectiveness. Comparedwith traditional field inventory work, forest inventories assistedby remote sensing reap the benefits not only of lower cost and lesstime consumed, but also with respect to the feasibility of conduct-ing inventories in unreachable forests located in remote or evensometimes life-threatening areas, such as in Laos, where 12 outof the 18 provinces are peppered with unexploded bombs orlandmines as a legacy of past wars (Tansubhapol, 1998).

In the light of the Scandinavian experience of assessing borealforest attributes, Airborne Laser Scanning (ALS) providing 3D struc-tural information on the vegetation shows the potential forestimating biophysical and structural properties of forests, suchas tree height, stem volume and basal area. In processing ALS data,

Society for Photogrammetry and R

e (EFI), Torikatu 34, [email protected] (Z. Hou).

tice, known as the canopy height distribution (CHD) approachand the individual tree delineation (ITD) (Hyyppä and Inkinen,1999; Holmgren, 2004; Næsset, 2004; Vauhkonen et al., 2009).

For optical data from passive sensors, both spectral and texturalfeatures are usually employed to explore the relationships betweenthe forest attributes of interest and remote sensing data at thestand, micro-stand or more commonly plot level. Since the individ-ual digital number (DN value) of a high or moderate-resolution im-age reflects less information than spatially grouped DN values,statistics such as mean or aggregate figures extracted at the re-quired level are often tested as potential predictors rather than val-ues for single pixels (Hyvönen et al., 2005; Packalén, 2009). Beingone of the most canonical algorithms of texture, the Haralick meth-od is based on the principle of a spatial dependence matrix of DNvalues (Haralick et al., 1973), and multiple derivative versionsbased on this are also popular in forestry (Tuominen and Pekkari-nen, 2005; Packalén and Maltamo, 2007). The advantage of texturelies in the independence of the spectral features with respect tospatial variations, which has hastened the use of a combination ofspectral and textural features for estimating forest attributes,providing better accuracy than the use of any feature alone.

Along with the growing attention to REDD issues, carbon inven-tory on estimating tropical forest biomass and carbon stock hasdrawn particular interests. Among ALS studies on tropical forests,

emote Sensing, Inc. (ISPRS) Published by Elsevier B.V. All rights reserved.

Z. Hou et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786 777

Drake et al. (2002) predicted basal area and aboveground biomassin Costa Rica with R2 values up to 72% and 93%, respectively. Hurttet al. (2004) combined ALS data and a height-structured model toestimate carbon of tropical forests also in Costa Rica with R2 valueof 78%. Amazon forests, representing certain types of tropicalforests, have been studied as hotspots for carbon assessment ordetection of deforestation with mostly satellite imagery (Souzaet al., 2003; Olander et al., 2008; Broich et al., 2009; Asner et al.,2009; Asner, 2009).

However, there is still scarcity in literatures on estimating trop-ical forest attributes, especially for stem volume, an essentialvariable also in carbon inventory by means of indirect methodusing biomass expansion factor. The reason was early pointedout by Nelson et al. (1997) that it has been difficult to establish sta-tistical relationships between ALS metrics and field measurementsof forest attributes in tropical forests due to the complex foreststructure and species composition. According to GOFC-GOLD(2009), studies on testing the performance of ALS and comparingthe applicability of ALS with optical sensors in tropical forestmapping over large and widely different areas still remainedinsufficient.

This study in response aims at estimating two basic biologicalindicators in tropical forests in Laos, stem volume and basal area,with alternative ALS, ALOS AVNIR-2 and Airborne CIR materials.The performance of each technique and their combinations wereevaluated and compared in order to distinguish the optimum onesat given accuracy level, especially for ALOS AVNIR-2 data, theapplicability of which is not much studied in forestry applications(Sarker and Nichol, 2011).

2. Materials

2.1. Study area

The forest area concerned is the Dongsithouane production for-ests, area c. 25 000 ha, in the province of Savanakhet (latitude 16�

Fig. 1. Study area and loca

330 N, longitude 104� 450 E) in Laos (Fig. 1). It is situated in hilly ter-rain and has a tropical climate.

2.2. Field data

The field data were based on a field campaign conducted in Feb-ruary 2009 for purpose of collecting transects and informationfrom different forest characteristics for modelling. The tally recordsshowed the existence of 268 tree species in this area, with onlydeciduous species dominant, principally represented by membersof the Dipterocarpaceae family such as Dipterocarpus tuberculatus,Pterospermum magalocarpum, Dipterocarpus alatus, Pentacme siam-ensis and Shorea obtuse. However, tree species differences wereignored in that detailed discrimination between species wasconsidered to have a negligible influence on the objectives of thepresent work. The canopy closure was measured based on circularfisheye photography. It differed greatly, ranging from 10% to 70%,but there could be errors since the survey was conducted in thedry season when dry dipterocarp had their leaves shed.

A total of 233 sample plots following either the permanent plotsor the ALS scan line in the Dongsithouane forests were surveyed,and the coordinates of each plot centre were recorded by handholdGPS. These plots were located in linear clusters (or inventory lines),each cluster being a strip of size 100 by 20 m, divided into twoplots at both ends each 20 by 20 m. The varying sized field plotswere defined by following rules: large trees with a Diameter atBreast Height (DBH) of 20 cm or larger were measured on plotsof 20 by 20 m (400 m2), stems with a DBH of 10 cm or larger butless than 20 cm on subplots of 10 by 10 m (100 m2), saplings ofheight 1.5 m or more and a DBH of less than 10 cm on subplotsof 4 by 4 m (16 m2) and seedlings of height 0.3 m or more but lessthan 1.5 m on subplots of 2 by 2 m (4 m2). Subplot informationensured that data were collected for trees of DBH below 20 cm.Each tree, whether alive or not, within a plot was tallied. Slopecorrection was also considered and was implemented at the plotlevel. Table 1 provides the characteristics of the sample plots.

tion of the field plots.

Table 1The field measured characteristics of the sample plots.

Characteristics Mean Max. Min. S.D. Model

Basal area (m2/ha) 16.6 148.3 1.1 14.1Number of stems per ha 1405.9 20575.0 25.0 2372.8Basal area median diameter (cm) 31.1 109.7 14.3 10.1Height of the basal area median tree (m) 12.1 23.5 5.0 3.9Stem volume (m3/ha) 110.1 1512.4 6.4 124.6 Pukkala, 2005

Table 2Origins of the multisource data.

Data type Scenes Resolution Acquisition time Supplier

ALS 1 pulse/m2 February 6–8, 2009 Finnmap and ArbonautAirborne CIR 989 0.25 � 0.25 m February 6–8, 2009 FinnmapALOS AVNIR-2 1 10 � 10 m September 13, 2006 Savannakhet

778 Z. Hou et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786

2.3. Multisource remote sensing data

The multisource remote sensing data procured for this area in-cluded ALS data, airborne high resolution digital colour infrared(CIR) orthophotos and optical data from the ALOS AVNIR-2 satel-lite. Spatial registration of ALS and imagery data was done byemploying ground control points in ortho-rectification. All the re-mote sensing data and field measurements were projected andregistered to WGS84/UTM zone N48 geo-reference system. Basicinformation regarding each data source is provided in Table 2.

The ALS data and the Airborne CIR orthophotos were obtainedsynchronously from a Piper PA-31 Navajo aircraft during the day-time in the course of three days, February 6–8, 2009. The flyingaltitude was 2000 m and the speed 120 knots, with 19 optimal tra-jectories planned in Leica Flight Planning and Evaluation Software(Leica FPES, 2011) plus an extra trajectory perpendicular to theothers used only for the purpose of calibration.

The ALS data were collected with a Leica ALS 40 LiDAR scannerwith parameters set at a field of view (FOV) of 30� and a sidelap of20% between flight lines. This resulted in a nominal sampling den-sity of 1 pulse/m2. Although the Leica ALS 40 LiDAR was capable ofcapturing multi-range measurements for each pulse, the ALS dataemployed in this paper were the first and last echoes. Terrasolidtools (Terrasolid, 2011) were utilised for ALS data preprocessing.Terramatch was employed to equalise the ground level betweeneach flight line and ground control points, while Terrascan wasused to process and classify the point data.

The aerial photos were taken with a Leica MP 39 digital camera.On account of their simultaneous recording, the Leica MP 39 andLeica ALS 40 datasets share some parameters in common, mostnotably flight altitude, speed and trajectory. In order to cover thewhole area, a total of 989 images were photographed with a reso-lution of 0.25 m. Nevertheless, considering the nature of forestryapplications, only the green, red and near-infrared bands wereused, giving an image type known as colour infrared (CIR). Allimages were ortho-rectified and geo-referred according to com-mercial photogrammetric mapping standards of Finnmap (2011).Although the weather conditions were ideal for aerial photogra-phy, an obvious colour difference between the days was found tobe caused by a bi-directional reflectance effect and the clicking ofthe shutter, so that radiometric correction was assumed to beneeded.

ALOS is a Japanese satellite designed for earth observation. Ithas onboard sensors of three types, a Panchromatic Remote-sens-ing Instrument for Stereo Mapping (PRISM), an Advanced Visibleand Near-Infrared Radiometer-2 (AVNIR-2), and a Phased ArrayL-band Synthetic Aperture Radar (PALSAR). Only the AVNIR-2

radiometric data were available for the present purposes. AVNIR-2 provides multispectral optical imagery of four bands (blue, green,red and near-infrared) with the spatial resolution of 10 m. Due toits sun-synchronous orbit at a height of 692 km above the ground,the swath width of image covers 70 km at the nadir, so that onescene was enough to cover the whole area on a scale of 1:25,000.The quality of the image was satisfactory for direct applicationwithout any need for further radiometric calibration.

3. Methods

3.1. Radiometric calibration of Airborne CIR

Once the anomalies in the 989 airborne CIR images caused bythe bi-directional reflectance effect had been identified, radiomet-ric calibration was needed before any further processing could bedone.

As it was not feasible to perform absolute radiometric calibra-tion, a relative technique, local radiometric correction, by themethod introduced by Tuominen and Pekkarinen (2004), wasemployed instead. The main idea of this lies in correction of theanomalies in Airborne CIR by means of local adjustment withthird-party imagery as a reference, given that the latter is lessprone to the bi-directional reflectance effect. In general, the refer-ence imagery should be independent of the bi-directional reflec-tance distribution function (BRDF), or at least the effect of thelatter should be insignificant. A satellite image can therefore qual-ify as a reference for aerial photographs. Ideally, the only imagefrom ALOS AVNIR-2 fulfils this purpose, but due to the differencein pixel size between Airborne CIR and ALOS AVNIR-2, localadjustment of the correction has to be based on a unit of the focalwindow representing one satellite pixel and several pixels in theaerial photo. In this instance the focal window was set at 10 by10 m so as to represent one ALOS AVNIR-2 pixel and 400 AirborneCIR pixels whose resolution were resampled to 0.5 m. Further-more, local adjustment had to be carried out separately by Eq.(1) for each corresponding band, green to green, red to red andNIR to NIR.

f CIRBandiðx; yÞ ¼

�f ALOSi

�f CIRBandi

� fCIRBandiðx; yÞ; ðx; yÞ�I; i ¼ 1;2;3 ð1Þ

wheref CIRBandi

ðx; yÞ: adjusted value at pixel (x,y) of Bandi in AirborneCIR I;�f ALOSi

: mean pixel value within the focal window of ALOSAVNIR-2;

Table 3Textural features used as predictors.

Layers Texturalfeature names

Equations

Texture 1 Contrast Pi;jði� jÞ2pði; jÞ; i–j

Texture 2 CorrelationP

i

Pjði�lx Þðj�lyÞpði;jÞ

rxry

Texture 3 Energy (Angular Second Moment) PiP

jpði; jÞ2

Texture 4 Entropy �P

iP

jpði; jÞ log½pði; jÞ�Texture 5 Homogeneity P

iP

jpði;jÞ

1þði�jÞ2; i–j

Texture 6 Inverse difference moment PiP

jpði;jÞði�jÞ2

; i–j

Texture 7 Max. probability maxij½pði; jÞ�Texture 8 Variance P

ið1� liÞ2P

jpði; jÞh iP

jð1� liÞ2P

ipði; jÞh i

Z. Hou et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786 779

�f CIRBandi: mean pixel value within the focal window of Airborne

CIR;fCIRBandi

ðx; yÞ: original pixel value in the corresponding AirborneCIR.

The advantage of this method is that it keeps the shape of thehistogram unchanged within the focal window and only shifts itslocation. Another benefit is that no priori data are needed, so thatthe method is always applicable as long as the reference satelliteimagery is reliable.

3.2. Statistical modelling

Plot-level multivariate linear least-squares regression analysisby stepwise selection of predictors according to the Akaike infor-mation criterion (AIC) was applied as a statistical modellingapproach in the R (version 2.10.1) environment. Other popularalternative estimation approaches would be classification andnon-parametric methods, but these were rejected here due to dif-ficulties in defining the classification categories and the insufficientnumber of sample plots for non-parametric methods.

Being the basic biological indicators required for evaluatingpresent and future conditions in a forest, stem volume and basalarea are predicted variables of interest in modelling. The stem vol-ume is also a key variable relevant to biomass and thus representsone step further on the way to estimating the carbon stock by anindirect method employing biomass expansion factors, while thebasal area is another crucial variable representing the structuralcondition of forests, which provides empirical understanding ofvariations in density.

As multisource data had been procured, the predictors neededto be selected for respective data type. It should be emphasised,however, that a large variation was identified between the fielddata from the 233 sample plots and the remote sensing data. Anal-yses attributed this to synergy between positioning errors causedby practical difficulties in some dense forests using handholdGPS and the deviation in measurements when aggregating sub-plots to that of standard-sized unit. For these reasons, the plot datawere combined by averaging their field measurements and remotesensing data to establish training areas.

The training areas used for modelling were based on 78 com-bined plots. In details, the tallied plots were sorted according tothe stem volume, and then every three plots were combined usingmean characteristics from the field measurements and theextracted remote sensing metrics. Therefore, the plots used for atraining area were unnecessarily spatially next to each other, butstill similar according to forest characteristics.

3.3. Predictors of ALS

The canopy height distribution approach (or the area-based sta-tistical method) was used for calculating ALS metrics. The last echodata were classified into ground and aboveground hits. The groundhits were then employed to produce a triangulated irregular net-work (TIN) that was later linearly interpolated for each point in or-der to produce the digital elevation model (DEM) for the areastudied (Axelsson, 1999). The canopy height model (CHM) wasgenerated by subtracting the DEM height from that of each firstand last echo hit. A threshold of 2 m was then applied to the can-opy height model in order to further exclude ground hits and hitsfrom stones, shrubs, etc. For establishing the relationship betweenfield measurements and ALS data, height distribution of the firstand last echoes was created for each 5 by 5 m grid from the canopyheight model, and various statistics or metrics were computed.

The derived metrics were height percentiles, proportional den-sity and other statistics. The height percentiles included quantiles

corresponding to the 5%, 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%,90% and 95% of the height distribution. The proportional canopydensities were also calculated for these quantiles. Other statisticsreferred to the mean, standard deviation, coefficient of variationof elevation and density. Ground hits were excluded for elevationmean, standard deviation and coefficient of variation. Coefficientof variation was set to 0 where mean elevation was 0. Densitywas calculated as the ratio of number of vegetation hits to thenumber of all hits.

3.4. Predictors of Airborne CIR and ALOS AVNIR-2

As texture features are inevitably related to spectral features inmuch the same way as particles and waves in light transmissiontheory (Haralick et al., 1973), both the spectral and the texture fea-tures were extracted as predictors for modelling. The extraction ofboth features took the size of the plot and that of an image pixelinto account, so that the extracted image values could properlyrepresent the forest information at the plot level. The spectral pre-dictors were derived from each band of optical data by taking themean DN values in 20 by 20 m neighbourhood of the plot centre.

Haralick’s textural features have been widely adopted for qual-itative and quantitative analyses in forestry applications (Packalénand Maltamo, 2007). Each element of the grey-tone spatial depen-dence matrix represents the frequency of two pixels having valuesdenoted as i and j and being defined by two parameters such asdistance and direction. In Haralick’s assumption, all the textureinformation of the image is contained in the grey-tone spatialdependence matrix, so that calculations based on the matrix canmeasure different aspects of the texture features. Out of 14 Hara-lick’s textural features, seven were used in the present workaccording to their relatively higher correlations with forest attri-butes than others shown in previous studies (Holopainen andWang, 1998; Tuominen and Pekkarinen, 2005), and the maximumprobability of the matrix was also extracted as the eighth texturalfeature, as shown in Table 3.

Since vegetation characteristics are of more concern than otherland cover types in forestry application, the Normalised DifferenceVegetation Index (NDVI) and the first principal component wereused instead of the original spectral bands for extracting the tex-tural features. NDVI serves well to distinguish vegetation fromother land cover types, and the first principal component accountsfor the most information contained in the original spectral bandsand achieves the purpose of dimension reduction.

NDVI is commonly used index for detecting live green vegeta-tion in multispectral remote sensing images, and has been widelyemployed for quantitative assessment of vegetation propertiessuch as productivity, biomass, Leaf Area Index, etc. (Baldi et al.,2008). As Eq. (2) indicates, the NDVI of a densely vegetated area

Fig. 2. Subset of textural features in the raster: (a) original Airborne CIR image; (b) PCA-derived entropy of Airborne CIR; (c) PCA-derived homogeneity of Airborne CIR; (d)original ALOS AVNIR-2; (e) NDVI-derived contrast of ALOS AVNIR-2; (f) NDVI-derived energy of ALOS AVNIR-2.

780 Z. Hou et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786

tends to be positive with values below 1, that of cloud and snowtends to be negative with values above �1 and that of water andsoil tends to be positive with low values, or sometimes evennegative.

NDVI ¼ NIR � REDNIR þ RED

ð2Þ

whereNIR: near-infrared bandRED: red band

Principal component analysis is an analytical method for pro-cessing data with p dimensions which takes a linear combinationof the original variables Xi, i = 1,2,. . .,p, and transforms them intoZi as shown in Eq. (3). The first principal component makes thelargest contribution to the variance and has the greatest powerto explain the original variables, while the subsequent principalcomponents exhibit gradually less explanatory power. The princi-pal components can be resolved by means of a correlation matrixor a covariance matrix, the latter approach being adopted herebecause the original variables all had the same measurement unitand were of similar magnitude.

Z ¼ Q T X ¼ ða1; a2; :::; apÞT ð3Þ

wherea1, a2,. . .,ap: eigenvectors corresponding to the eigenvalue of the

covariance matrix.The coding algorithm was produced and the output generated in aMatlab (64bit version, R2009b) environment, using a moving win-dow technique for extracting the textural features in order to pro-duce a wall-to-wall raster for detailed analysis (Fig. 2). The rasterpredictors were produced based on NDVI and the first principalcomponents of the original respective spectral bands.

There are several important parameters to consider in the pro-cess of extracting Haralick textural features: re-quantificationclass, texture unit size, direction and distance. These parameterswere determined by means of practical tests and observationsmade in previous studies, e.g. Holopainen and Wang (1998) andTuominen and Pekkarinen (2005). The final features for the ALOSAVNIR-2 images were calculated with 16 re-quantification classes,a 30 � 30 m neighbourhood around each sample plot as the tex-ture unit and one pixel as the distance. Mean values for four direc-tions were calculated and stretched linearly to an unsigned 8-bitform. For the Airborne CIR photographs 16 re-quantification clas-ses and a 20 � 20 m neighbourhood were used, following thesuggestions of Holopainen and Wang (1998), who maintained thata 20 � 20 m neighbourhood results in a near-optimal texture unitfor extracting Haralick textural features from aerial photographsin forestry applications. One pixel was used as the distance andmean values for four directions were again stretched linearly toan unsigned 8-bit form.

3.5. Accuracy assessment

The Root Mean Square Error (RMSE) and its relative form havebeen widely used as an established way of validating the accuracyof forest inventories and estimating the goodness of fit of models.This provides a measure of variability that includes the effects ofboth random error and bias. Both absolute and relative forms ofthe RMSE are defined in Eq. (4). Although the regression model isin most cases unbiased, bias may exist if transformation of thepredicted variable is implemented. Correspondingly, the bias alsohas absolute and relative forms, as shown in Eq. (5). In the end,the statistics based on leave-one-out cross-validation were recal-culated to assist in interpreting the stability of the models andcontrolling instances of over-fitting.

Table 4Comparison of models.

Data source Model S.E. Predictors Before Cross-validation Cross-validation

R2R2

adjRMSE RMSE% Bias Bias% RMSE RMSE% Bias Bias%

ALS Stem volume (m3/ha) 0.390 2 0.911 0.908 39.561 34.636 4.705 4.119 42.174 36.923 0.141 0.124Basal area (m2/ha) 6.318 7 0.822 0.804 5.990 35.300 0 0 8.032 47.347 0.012 0.072

ALOS AVNIR-2 Stem volume (m3/ha) 0.445 7 0.779 0.756 62.219 54.472 0.005 0.004 78.314 68.563 0.306 0.268Basal area (m2/ha) 9.182 7 0.623 0.586 8.699 51.277 0 0 11.218 66.127 0.049 0.289

Airborne CIR Stem volume (m3/ha) 0.435 6 0.809 0.793 57.733 50.545 0.256 0.224 75.169 65.810 0.212 0.186Basal area (m2/ha) 7.994 6 0.710 0.686 7.009 41.285 0 0 9.692 57.086 0.070 0.410

ALS + ALOS AVNIR-2 Stem volume (m3/ha) 0.362 5 0.872 0.863 47.314 41.423 0.472 0.413 54.990 48.140 �0.241 �0.211Basal area (m2/ha) 7.123 4 0.764 0.751 6.891 40.621 0 0 8.478 49.979 0.047 0.277

ALS + Airborne CIR Stem volume (m3/ha) 0.374 3 0.918 0.915 37.761 33.065 4.499 3.939 39.954 34.980 �0.013 �0.011Basal area (m2/ha) 6.967 4 0.774 0.761 6.740 39.698 0 0 8.843 52.085 0.023 0.137

Z. Hou et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786 781

RMSE ¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPni¼1ðyi � yiÞ2

n

s; RMSE% ¼

RMSEPni¼1

yin

� 100; ð4Þ

wheren: number of sample plots;i: observation index;yi: observed value for sample plot i;yi: predicted value for sample plot i.

Bias ¼Xn

i¼1

yi � yi

n; Bias% ¼

BiasPni¼1

yin

� 100 ð5Þ

wheren: number of sample plots;i: observation index;yi: observed value for sample plot i;yi: predicted value for sample plot i.

4. Results

4.1. Form of the models

Natural logarithmic transformation to the predicted variable ofstem volume was applied for correcting detected problems on non-normality and non-constant variances, while all predictors remainunchanged. Before transforming back to the original scale of anti-logarithm, half of the error variance of the model was thus addedto the model prediction to attain unbiased model predictions(Baskerville, 1972; Flewelling and Pienaar, 1981). R2 (coefficientof determination) and the adjusted R2 were recalculated after backtransformation. As to modelling basal area, no transformation ofeither the predictors or the predicted variable was found neces-sary. The overall statistics for each model are detailed in Table 4and several scatter plots are presented in Figs. 3 and 4 to illustratethe residual plots and plots comparing the predictions with fieldmeasurements.

It can be seen from Table 4 that the ALS data provide the bestperformance for non-hybrid models of stem volume and basal areain terms of higher R2 values, with correspondingly lower RMSEvalues than the others. The Airborne CIR models perform slightlybetter than those based on ALOS AVNIR-2, though both remain ata similar accuracy level and fall considerably behind ALS in termsof performance.

Although the hybrid stem volume model based on ALS and Air-borne CIR improved the RMSE%, this improvement of 1.9% was notmuch. As for hybrid models predicting basal area, any opticalvariable was failed to be involved with analyses attributing thisto interferences brought by extra predictors in course of stepwiseselection. Statistical outliers were identified in each residual plotand at least one exerted a strong influence during modelling, but

these were not removed as an attempt was being made to toleratepossible errors arising from training areas.

4.2. Performance of predictors

Details of the predictors involved in each model are provided inthe Appendix, where the textures follow the order presented inTable 3. Analyses of the predictors in each model pointed to somegeneral tendencies that were found to be informative.

The first pulse proportional density metrics at 10% and 60%constituted the best model for stem volume, while the height met-rics at 50%, 60% and mean height have a high correlation with basalarea in addition to the first pulse density metrics. It is true that thelast pulse entails some uncertainties due to the complex itinerariesof echo reflectance, but auxiliary field data suggest that one possi-ble reason could be the complicated tropical forest structures,which hinder accurate estimation based on first echoes reflectedmainly by the upper canopies (Fig. 5). In order to verify this issuefurther more clearly stratified sample plots would be needed, butat least it can be said that last pulse metrics proved statisticallysignificant for raising the estimation capacity in this study.

As for the optical data for estimating stem volume and basalarea, the near-infrared (NIR) band, which reveals the most vegeta-tion information of all the original spectral bands, was selected inalmost every pure optical model, except the Airborne CIR model forbasal area, which instead selected the green band. Likewise, NDVIand its textural form also displayed high correlations and wereincluded in each optical model. NDVI textures accounted for allthe textural predictors in the ALOS AVNIR-2 models, whereasPCA textures contributed only to the Airborne CIR models. Lookedat in more detail, the textural features that contributed most tothese models were ones referring to features of contrast, energy(angular second moment), entropy and homogeneity. The contrastfeature mainly reveals the difference between the top and bottomvalues of neighbouring pixels, so that the lower the contrast is, thelower the spatial frequencies. Energy reflects the texture unifor-mity by calculating pixel pair repetitions, so that the higher the en-ergy is the more constant or periodic is the distribution of pixelvalues. Entropy measures the degree of disorder in an image, sothat the higher the entropy is, the lower the textural uniformity,while homogeneity, as the word suggests, measures image homo-geneity, a feature that is very sensitive to the values of near-diag-onal cells in a grey-level co-occurrence matrix (GLCM).

5. Discussion

Among the remote sensing approaches employed here it wasALS that provided the most promising performance when estimat-ing stem volume and basal area in a mixed-species tropical forest.ALOS AVNIR-2 and Airborne CIR data performed less well. Though,

0 200 400 600 800 1000 1200

−50

050

100

Back−Transformed Residual Plot: ALS Volume (m3 ha)

Fit

Res

idua

l

0 200 400 600 800 1000

020

040

060

080

010

0012

00

Measurement vs Prediction with 1:1 Line: ALS Volume (m3 ha)

Measurement

Pred

ictio

n

0 100 200 300 400 500 600

−400

−300

−200

−100

010

0

Back−Transformed Residual Plot: ALOS Volume (m3 ha)

Fit

Res

idua

l

0 200 400 600 800 1000

010

020

030

040

050

060

0Measurement vs Prediction with 1:1 Line: ALOS Volume (m3 ha)

Measurement

Pred

ictio

n

Fig. 3. ALS and ALOS AVNIR-2 plots of stem volume.

782 Z. Hou et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786

the ALS estimates of basal area did not attain the same accuracylevel as that of stem volume after cross-validation. By contrast,optical data were found to be more accurate at estimating basalarea than stem volume, although still with a less appealing perfor-mance than ALS.

When comparing the results with other ALS-based studies ofestimation accuracy, it was noted that better results were ob-tained for boreal forests in Scandinavian countries, where relativeRMSE values between 20.51% and 23.86% were reported for stemvolume and 17.15% for basal area by Packalén and Maltamo(2006, 2007) using hybrid models composed of ALS and aerialphotographs. Yu et al. (2010) predicted stem volume with rela-tive RMSE of 20.9% by CHD approach. For temperate deciduousforest, a study carried out in Japan reported a RMSE of41.9 m3/ha for stem volume (Keiko et al., 2010), which comesclose to our result, 42.2 m3/ha. For optical data, by comparisonwith the results of Tuominen and Haakana (2005) who usedLandsat TM and aerial photographs to estimate stem volume, this

study derived smaller relative RMSE values than theirs, in therange 74.5–83.3%.

Tropical forests have a more complicated vertical structure anda greater variety of tree species, and this presents challenges whenthe aim is to obtain as good results as under boreal conditions.Especially, tropical forest may pose a challenge to the successfuldetection of tree crowns due to the density and the overlay ofneighbouring tree crowns. A systematic underestimation of treenumbers and forest attributes mentioned by Gonzalez et al.(2010) would be even worse with tropical forests. Despite this,the greatest limitation on using optical data to estimate tropicalforest attributes arises from the fact that optical sensors recordmainly the tree crown surface, while the suppressed understoryoften remains undetectable. This also helps to explain the rela-tively large estimation error and the low saturation level affectingpassive optical sensors.

A general consensus has been reached that passive sensors areprone to saturation at high stem volume or biomass levels (Frank-

0 20 40 60 80

−15

−10

−50

510

15Residual Plot: ALS Basal Area (m2 ha)

Fit

Res

idua

l

0 20 40 60 80 100

020

4060

80

Measurement vs Prediction with 1:1 Line: ALS Basal Area (m2 ha)

Measurement

Pred

ictio

n

0 10 20 30 40 50 60

−20

−10

010

2030

40

Residual Plot: ALOS Basal Area (m2 ha)

Fit

Res

idua

l

0 20 40 60 80 100

010

2030

4050

60

Measurement vs Prediction with 1:1 Line: ALOS Basal Area (m2 ha)

Measurement

Pred

ictio

n

Fig. 4. ALS and ALOS AVNIR-2 plots of basal area.

Fig. 5. ALS profile of multi-storeyed forest in the study area with raw echoes. Blue colour stands for only echoes, green for first of many echoes and red for last of manyechoes.

Z. Hou et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786 783

lin, 1986; Horler and Ahern, 1986; Cohen and Spies, 1992; Asneret al., 2003; Garcia et al., 2010), while active sensors do not sufferfrom this risk until the levels become somewhat higher (Lefskyet al., 1999; Melon et al., 2001; Drake et al., 2002; Nelson et al.,2007). ALOS AVNIR-2 and Airborne CIR data provided estimatesof similar accuracies without any obvious saturation level being

identified. The reason for this was ascertained through analysesof tallied records as that the figures for the true timber stock avail-able in the forests were too low, in relative terms, to allow satura-tion to be detected. Nevertheless, it may be deduced from thescatter plots that the estimates up to 300 m3/ha for stem volumeand 50 m2/ha for basal area should be reliable.

784 Z. Hou et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786

Methodology wise, it was feasible to transplant popular Scandi-navian approaches from estimating boreal forest attributes to thatof tropical context. Among the data sources tested, it was ALS thatproved to be the most accurate and competent in Laos, SoutheastAsia, thus complement other investigations focused on differenttropical areas (Drake et al., 2002; Hurtt et al., 2004; Chamberset al., 2007). However, inaccurate ALS measurements are still proneto occur in hilly or mountainous regions with errors extending upto several metres (McKean and Roering, 2004). Furthermore, theawareness concerning the effectiveness of hybrid models coincideswith another study conducted by Nelson et al. (2007) who com-pared the ALS-only and joint ALS–RaDAR models and concludedthat there was little gain brought by combination.

With respect to the cost of data procurement, ALS was the mostcostly despite its low pulse density at 1 pulse/m2, and the secondmost expensive was Airborne CIR. ALOS AVNIR-2 was in relativeterms ten times cheaper than ALS. If taking the cost-effectivenessunder tropical context into consideration, ALOS AVNIR-2 data is

Models Predictors

ALS (Intercept)Stem volume First pulse proportional density 10%

First pulse proportional density 60%ALS (Intercept)Basal area First pulse height percentile 50%

Last pulse height percentile 50%Last pulse height percentile 60%First pulse proportional density 50%First pulse proportional density 90%First pulse proportional density 95%Last pulse mean elevation

ALOS AVNIR-2 (Intercept)Stem volume ALOS red band

ALOS NIR bandALOS NDVIALOS NDVI texture 1ALOS NDVI texture 3ALOS NDVI texture 4ALOS NDVI texture 5

ALOS AVNIR-2 (Intercept)Basal area ALOS red band

ALOS NIR bandALOS NDVIALOS NDVI texture 1ALOS NDVI texture 3ALOS NDVI texture 5ALOS NDVI texture 8

Airborne CIR (Intercept)Stem volume CIR NIR band

CIR PCA texture 5CIR PCA texture 6CIR PCA texture 7CIR NDVICIR NDVI texture 4

Airborne CIR (Intercept)Basal area CIR green band

CIR PCA texture 1CIR PCA texture 4CIR PCA texture 5CIR PCA texture 6

of potential to be used for obtaining rough but economic estimates,while ALS data is an alternative to satisfy needs demanding betteraccuracy.

Acknowledgements

The authors thank the SUFORD project and the Finnish Ministryof Foreign Affairs for their support for the data acquisition. We arealso very grateful to the staff of the University of Eastern Finlandand the European Forest Institute (EFI), especially Matti Maltamoand Tuula Nuutinen, for their invaluable suggestions and construc-tive comments. Funding for this research was received from aPonsse Grant 2010 managed by the Foundation for European ForestResearch (FEFR). Last but not least, the comments provided by thetwo anonymous reviewers are gratefully acknowledged.

Appendix A. Predictors in each model

Estimate S.E. t Value Pr(>|t|)

2.338 0.184 12.736 ⁄⁄⁄0.084 0.034 2.461 ⁄0.119 0.032 3.730 ⁄⁄⁄�8.732 3.565 �2.449 ⁄�3.912 1.727 �2.266 ⁄�4.658 1.527 �3.051 ⁄⁄7.496 1.519 4.936 ⁄⁄⁄6.720 2.127 3.160 ⁄⁄26.801 7.567 3.542 ⁄⁄⁄�26.803 7.519 �3.565 ⁄⁄⁄�3.742 1.150 �3.254 ⁄⁄�43.047 11.352 �3.792 ⁄⁄⁄0.171 0.073 2.333 ⁄�0.179 0.052 �3.446 ⁄⁄⁄0.164 0.045 3.645 ⁄⁄⁄0.551 0.136 4.053 ⁄⁄⁄�0.014 0.004 �3.536 ⁄⁄⁄0.054 0.023 2.348 ⁄0.047 0.011 4.430 ⁄⁄⁄�709.517 198.389 �3.576 ⁄⁄⁄5.447 1.516 3.592 ⁄⁄⁄�5.340 1.067 �5.004 ⁄⁄⁄3.944 0.959 4.113 ⁄⁄⁄7.354 2.698 2.726 ⁄⁄�0.265 0.071 �3.717 ⁄⁄⁄0.630 0.208 3.030 ⁄⁄0.306 0.111 2.769 ⁄⁄168.070 59.812 2.810 ⁄⁄�0.046 0.018 �2.601 ⁄0.070 0.024 2.855 ⁄⁄�0.699 0.246 �2.837 ⁄⁄�0.018 0.006 �2.878 ⁄⁄0.064 0.009 7.376 ⁄⁄⁄�0.023 0.006 �3.511 ⁄⁄⁄5008.570 975.487 5.134 ⁄⁄⁄�1.816 0.254 �7.160 ⁄⁄⁄9.693 3.186 3.042 ⁄⁄�0.787 0.237 �3.325 ⁄⁄1.588 0.483 3.290 ⁄⁄�19.674 3.967 �4.960 ⁄⁄⁄

Appendix A (continued)

Models Predictors Estimate S.E. t Value Pr(>|t|)

CIR NDVI texture 4 �0.316 0.121 �2.611 ⁄ALS + ALOS AVNIR-2 (Intercept) 0.196 0.574 0.342Stem volume First pulse proportional density 30% 0.190 0.031 6.084 ⁄⁄⁄

Last pulse proportional density 50% 0.165 0.052 3.182 ⁄⁄Last pulse proportional density 90% �0.168 0.055 �3.042 ⁄⁄ALOS NDVI texture 2 0.033 0.007 4.383 ⁄⁄⁄ALOS NDVI texture 7 0.007 0.002 3.050 ⁄⁄

ALS + ALOS AVNIR-2 (Intercept) �15.970 3.525 �4.531 ⁄⁄⁄Basal area Last pulse height percentile 40% �4.537 0.792 �5.727 ⁄⁄⁄

Last pulse height percentile 60% 4.354 0.867 5.021 ⁄⁄⁄First pulse proportional density 90% 30.926 8.260 3.744 ⁄⁄⁄First pulse proportional density 95% �28.579 8.245 �3.466 ⁄⁄⁄

ALS + Airborne CIR (Intercept) 5.251 1.023 5.135 ⁄⁄⁄Stem volume First pulse height percentile 40% 0.091 0.029 3.142 ⁄⁄

First pulse proportional density 10% 0.109 0.028 3.853 ⁄⁄⁄CIR NDVI texture 4 �0.016 0.006 �2.818 ⁄⁄

ALS + Airborne CIR (Intercept) �10.933 3.495 �3.128 ⁄⁄Basal area First pulse height percentile 50% �6.472 1.754 �3.689 ⁄⁄⁄

Last pulse height percentile 50% �6.188 1.554 �3.983 ⁄⁄⁄Last pulse height percentile 60% 5.977 1.617 3.696 ⁄⁄⁄First pulse proportional density 50% 8.895 1.649 5.393 ⁄⁄⁄

Significance codes: 0‘⁄⁄⁄’; 0.001‘⁄⁄’; 0.01‘⁄’; 0.05‘.’; 0.1‘ ’ 1.

Z. Hou et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786 785

References

Asner, G.P., Bustamante, M.M.C., Townsend, A.R., 2003. Scale dependence ofbiophysical structure in deforested areas bordering the Tapajos NationalForest, Central Amazon. Remote Sensing of Environment 87 (4), 507–520.

Asner, G.P., 2009. Tropical forest carbon assessment: integrating satellite andairborne mapping approaches. Environmental Research Letters 4 (3), 034009.

Asner, G.P., Knapp, D.E., Balaji, A., Paez-Acosta, G., 2009. Automated mapping oftropical deforestation and forest degradation: CLASlite. Journal of AppliedRemote Sensing 3 (1), 033543.

Axelsson, P.E., 1999. Processing of laser scanner data – algorithms and applications.ISPRS Journal of Photogrammetry and Remote Sensing 54 (2-3), 138–147.

Baldi, G., Nosetto, M.D., Aragón, R., Aversa, F., Paruelo, J.M., Jobbágy, E.G., 2008.Long-term satellite NDVI data sets: evaluating their ability to detect ecosystemfunctional changes in South America. Sensors 8 (9), 5397–5425.

Baskerville, G.L., 1972. Use of logarithmic regression in the estimation of plantbiomass. Canadian Journal of Forest Research 2 (2), 49–53.

Broich, M., Stehman, S.V., Hansen, M.C., Potapov, P., Shimabukuro, Y.E., 2009. Acomparison of sampling designs for estimating deforestation from Landsatimagery: a case study of the Brazilian Legal Amazon. Remote Sensing ofEnvironment 113 (11), 2448–2454.

Chambers, J.Q., Asner, G.P., Morton, D.C., Anderson, L.O., Saatchi, S.S., Espirito-Santo,F.D., Palace, M., Souza, C.Jr., 2007. Regional ecosystem structure and function:ecological insights from remote sensing of tropical forests. Trends in Ecologyand Evolution 22 (8), 414–423.

Cohen, W.B., Spies, T.A., 1992. Estimating structural attributes of Douglas-fir/western hemlock forest stands from Landsat and SPOT imagery. Remote Sensingof Environment 41 (1), 1–17.

Drake, J.B., Dubayah, R.O., Clark, D.B., Knox, R.G., Blair, J.B., Hofton, M.A., Chazdon,R.L., Weishampel, J.F., Prince, S., 2002. Estimation of tropical forest structuralcharacteristics using large-footprint Lidar. Remote Sensing of Environment 79(2), 305–319.

Flewelling, J.W., Pienaar, L.V., 1981. Multiplicative regression with lognormal errors.Forest Sciences 27 (2), 281–289.

Finnmap, 2011. FM-International Oy Finnmap. Available from: <http://www.finnmap.com> (accessed September 2, 2011).

Franklin, J., 1986. Thematic mapper analysis of coniferous forest structure andcomposition. International Journal of Remote Sensing 7 (10), 1287–1301.

Garcia, M., Riano, D., Chuvieco, E., Danson, F.M., 2010. Estimating biomass carbonstocks for a Mediterranean forest in central Spain using LiDAR height andintensity data. Remote Sensing of Environment 114 (4), 816–830.

GOFC-GOLD, 2009. A sourcebook of methods and procedures for monitoring andreporting anthropogenic greenhouse gas emissions and removals caused bydeforestation, gains and losses of carbon stocks in forests remaining forestsandforestation. GOFC-GOLD Report version COP15–1. GOFC-GOLD Project Office,Natural Resources Canada, Alberta, Canada.

Gonzalez, P., Asner, G.P., Battles, J.J., Lefsky, M.A., Waring, K.M., Palace, M., 2010.Forest carbon densities and uncertainties from Lidar, Quickbird, and field

measurements in California. Remote Sensing of Environment 114 (7), 1561–1575.

Haralick, R.M., Shanmugam, K., Dinsteein, J., 1973. Textural features for imageclassification. IEEE Transactions on Systems Man and Cybernetics 3 (6), 610–621.

Holmgren, J., 2004. Prediction of tree height, basal area and stem volume in foreststands using airborne laser scanning. Scandinavian Journal of Forest Research19 (6), 543–553.

Holopainen, M., Wang, G., 1998. The calibration of digitized aerial photographs forforest stratification. International Journal of Remote Sensing 19 (4), 677–696.

Horler, D.N.H., Ahern, F.J., 1986. Forestry information content of thematic mapperdata. International Journal of Remote Sensing 7 (3), 405–428.

Hurtt, G.C., Dubayah, R., Drake, J., Moorcroft, P., Pacala, S.W., Blair, J.B., Fearon, M.G.,2004. Beyond potential vegetation: combining LiDAR data and a heightstructured model for carbon studies. Ecological Applications 14 (3), 873–883.

Hyvönen, P., Pekkarinen, A., Tuominen, S., 2005. Segment-level stand inventory forforest management. Scandinavian Journal of Forest Research 20 (1), 75–84.

Hyyppä, J., Inkinen, M., 1999. Detecting and estimating attributes for single treesusing laser scanner. The Photogrammetric Journal of Finland 16 (2), 27–42.

Keiko, I., Junichi, I., Takeshi, S., Yukihiro, M., Katsunori, K., 2010. Estimating standvolume in broad-leaved forest using discrete-return LiDAR: plot-basedapproach. Landscape and Ecological Engineering 6 (1), 29–36.

Lefsky, M.A., Harding, D., Cohen, W.B., Parker, G., Shugart, H.H., 1999. Surface Lidarremote sensing of basal area and biomass in deciduous forests of EasternMaryland, USA. Remote Sensing of Environment 67 (1), 83–98.

Leica FPES, 2011. Leica Flight Planning and Evaluation Software. Available from:<http://www.leica-geosystems.com/en/Leica-FPES_86824.htm> (accessedSeptember 2, 2011).

McKean, J., Roering, J., 2004. Operational accuracy of LiDAR in mountainous terrain.In: 10th Biennial Forest Service Remote Sensing Applications Conference, April5–9, 2004, Salt Lake City.

Melon, P., Martinez, J.M., Le Toan, T., Ulander, L.M.H., Beaudoin, A., 2001. On theretrieving of forest stem volume from VHF SAR data: observation andmodelling. IEEE Transactions on Geoscience and Remote Sensing 39 (11),2364–2372.

Næsset, E., 2004. Effects of different flying altitudes on biophysical standproperties estimated from canopy height and density measured with a small-footprint airborne scanning laser. Remote Sensing of Environment 91 (2),243–255.

Nelson, R., Oderwald, R., Gregoire, T.G., 1997. Separating the ground and airbornelaser sampling phases to estimate tropical forest basal area, volume, andbiomass. Remote Sensing of Environment 60 (3), 311–326.

Nelson, R.F., Hyde, P., Johnson, P., Emessiene, B., Imhoff, M.L., Campbell, R., 2007.Investigating RaDAR-LiDAR synergy in a North Carolina pine forest. RemoteSensing of Environment 110 (1), 98–108.

Olander, L.P., Gibbs, H.K., Sterininger, M., Swenson, J.J., Murray, B.C., 2008. Referencescenarios for deforestation and forest degradation in support of REDD: a reviewof data and methods. Environmental Research Letters 3 (2), 025011.

786 Z. Hou et al. / ISPRS Journal of Photogrammetry and Remote Sensing 66 (2011) 776–786

Packalén, P., Maltamo, M., 2006. Predicting the plot volume by tree species usingairborne laser scanning and aerial photographs. Forest Science 52 (6), 611–622.

Packalén, P., Maltamo, M., 2007. The k-MSN method for the prediction of species-specific stand attributes using airborne laser scanning and aerial photographs.Remote Sensing of Environment 109 (3), 328–341.

Packalén, P., 2009. Using airborne laser scanning data and digital aerial photographsto estimate growing stock by tree species. Dissertationes Forestales 77. Doctoralthesis, University of Joensuu, Faculty of Forest Sciences, Joensuu, Finland.

Pukkala, T., 2005. Volume and taper models for forest trees in Lao PDR. SustainableForestry and Rural Development Project - Lao PDR. Technical Report.

Sarker, L.R., Nichol, J.E., 2011. Improved forest biomass estimates using ALOSAVNIR-2 texture indices. Remote Sensing of Environment 115 (4), 968–977.

Souza, C., Firestone, L., Silva, L.M., Roberts, D., 2003. Mapping forest degradation inthe Eastern Amazon from SPOT 4 through spectral mixture models. RemoteSensing of Environment 87 (4), 494–506.

Tansubhapol, B., 1998. Logistics prevent Laos from signing mine pact. Bangkok Post,Thailand. Available from:<http://www.hartford-hwp.com/archives/27a/075.html> (accessed September 2, 2011).

Terrasolid, 2011. Terrasolid - The standard for airborne and mobile LiDAR andimage processing. Available from: <http://www.terrasolid.fi/en> (accessedSeptember 2, 2011).

Tuominen, S., Pekkarinen, A., 2004. Local radiometric correction of digital aerialphotographs for multi source forest inventory. Remote Sensing of Environment89 (1), 72–82.

Tuominen, S., Pekkarinen, A., 2005. Performance of different spectral and texturalaerial photograph features in multi-source forest inventory. Remote Sensing ofEnvironment 94 (2), 256–268.

Tuominen, S., Haakana, M., 2005. Landsat TM imagery and high altitude aerialphotographs in estimation of forest characteristics. Silva Fennica 39 (4), 573–584.

UNFCCC, 2010. Report of the Conference of the Parties on its fifteenth session, heldin Copenhagen from 7 to 19 December 2009, Addendum, Part Two: Action takenby the Conference of the Parties at its fifteenth session (Copenhagen Accord).United Nations Office at Geneva, Geneva (Switzerland). Available from: <http://unfccc.int/resource/docs/2009/cop15/eng/11a01.pdf> (accessed September 2,2011).

Vauhkonen, J., Tokola, T., Packalen, P., Maltamo, M., 2009. Identification ofScandinavian commercial species of individual trees from airborne laserscanning data using alpha shape metrics. Forest Science 55 (1), 37–47.

Yu, X., Hyyppä, J., Holopainen, M., Vastaranta, M., 2010. Comparison of area-basedand individual tree-based methods for predicting plot-level forest attributes.Remote Sensing 2 (6), 1481–1495.