isprs journal of photogrammetry and remote...

ISPRS Journal of Photogrammetry and Remote Sensing 139 (2018) 228–240

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

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

Automated method for measuring the extent of selective logging damagewith airborne LiDAR data

https://doi.org/10.1016/j.isprsjprs.2018.02.0220924-2716/� 2018 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.

Abbreviations: ELCH, contiguous areas of Extremely Low Canopy Height; NU,contiguous areas with No Understory; RCH, contiguous areas of Reduced CanopyHeight; AA, features identified by the Automated Algorithm; FM, features identifiedas Field Measurements or through aerial photos.⇑ Corresponding author at: Applied GeoSolutions, 55 Main St., Newmarket, NH

03857, USA.E-mail address: [email protected] (S.C. Hagen).

L. Melendy a, S.C. Hagen a,⇑, F.B. Sullivan b, T.R.H. Pearson c, S.M. Walker c, P. Ellis d, Kustiyo e,Ari Katmoko Sambodo e, O. Roswintiarti e, M.A. Hanson f, A.W. Klassen g, M.W. Palace b, B.H. Braswell a,G.M. Delgado a

aApplied GeoSolutions, Newmarket, NH, USAb Earth Systems Research Center, Institute for the Study of Earth, Oceans, and Space, University of New Hampshire, USAcEcosystem Services Unit, Winrock International, Arlington, USAd The Nature Conservancy, Arlington, VA, USAe Indonesia National Institute of Aeronautics and Space (LAPAN), Jl. Lapan No. 70, Pekayon Pasar Rebo, Jakarta, IndonesiafDevelopment Seed, USAg Tropical Forest Foundation, Bogor, West Java, Indonesia

a r t i c l e i n f o a b s t r a c t

Article history:Received 13 October 2017Received in revised form 7 February 2018Accepted 26 February 2018Available online 26 March 2018

Keywords:LidarSelective loggingTropical forest monitoringREDD+Automated logging algorithmKalimantanIndonesiaReduced impact logging (RIL)

Selective logging has an impact on the global carbon cycle, as well as on the forest micro-climate, andlonger-term changes in erosion, soil and nutrient cycling, and fire susceptibility. Our ability to quantifythese impacts is dependent on methods and tools that accurately identify the extent and features oflogging activity. LiDAR-based measurements of these features offers significant promise. Here, we presenta set of algorithms for automated detection and mapping of critical features associated with logging –roads/decks, skid trails, and gaps – using commercial airborne LiDAR data as input. The automatedalgorithm was applied to commercial LiDAR data collected over two logging concessions in Kalimantan,Indonesia in 2014. The algorithm results were compared to measurements of the logging features col-lected in the field soon after logging was complete. The automated algorithm-mapped road/deck and skidtrail featuresmatch closely with features measured in the field, with agreement levels ranging from 69% to99% when adjusting for GPS location error. The algorithm performed most poorly with gaps, which, bytheir nature, are variable due to the unpredictable impact of tree fall versus the linear and regular featuresdirectly created by mechanical means. Overall, the automated algorithm performs well and offerssignificant promise as a generalizable tool useful to efficiently and accurately capture the effects of selec-tive logging, including the potential to distinguish reduced impact logging from conventional logging.� 2018 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier

B.V. All rights reserved.

1. Introduction

As the global human population continues to rise, there is anincreasing impact on the declining tropical forest regions (Bushet al., 2016; Ryan et al., 2017). These impacts include deforestation,degradation of carbon reserves, and the reduction of biodiversity

(Bustamante et al., 2016). With 20% of tropical forests designatedas commercial logging concessions (Blaser et al., 2011), manage-ment decisions regarding timber harvest play a crucial role in mit-igating these impacts (Putz et al., 2012). Efforts to improvemanagement practices in tropical forests have been extensive overthe past three decades. Most directly, the primary improvement isthe widespread replacement of conventional practices, repletewith poor planning on inventory, felling and extraction, withreduced impact practices for the selection and removal of a subsetof individual trees within a stand. Practice improvements thatreduce the impact of logging on carbon emissions and biodiversitycenter around planning: the implementation of a detailed skiddingplan, a carefully scheduled logging cycle, and a plan to minimize

http://crossmark.crossref.org/dialog/?doi=10.1016/j.isprsjprs.2018.02.022&domain=pdfhttps://doi.org/10.1016/j.isprsjprs.2018.02.022mailto:[email protected]://doi.org/10.1016/j.isprsjprs.2018.02.022http://www.sciencedirect.com/science/journal/09242716http://www.elsevier.com/locate/isprsjprs

L. Melendy et al. / ISPRS Journal of Photogrammetry and Remote Sensing 139 (2018) 228–240 229

damage associated with the felling of a tree. The resilience of forestcarbon and diversity in these stands is heavily dependent on theimplementation of these practices. While logging practices varysubstantially across the world’s tropical forests, these reducedimpact logging techniques are now practiced in many areas(Pearson et al., 2017). But detailed information on the adoptionof these practices is still limited.

Tropical forest regions are vast and often difficult to access – soremote sensing tools are likely to be an important component ofcost effective logging impact monitoring systems (Chamberset al., 2007; DeFries et al, 2005; Hansen et al., 2008). Even withtoday’s advances in remote sensing technology, the extent of log-ging impacts across the tropics remains poorly quantified (Asneret al., 2005). Estimates from satellite imagery have shown thatthe land area affected by logging in the Brazilian Amazon oftenexceeded the area of annual deforestation in the early 2000s(Asner et al., 2005). Furthermore, reliance on traditional methodssuch as mill surveys and field reports for estimating logging areaand damage is susceptible to under-reporting due to extensivebut frequently under-counted illegal logging activity. Research inthe last decade indicates as much as 72, 61, and 65% of logging isillegal in the Brazilian Amazon, Indonesia, and in Ghana, respec-tively (Lawson and MacFaul, 2010), but the spatial location andintensity of illegal logging is entirely unknown and virtuallyimpossible to track without expensive, highly focused studies.

In addition to direct effects on the carbon cycle, logging alsoaffects non-carbon forest ecosystem attributes at a range of timescales, including immediate changes in the forest micro-climate,and longer-term changes in erosion, soil and nutrient cycling, firesusceptibility, as well as potential changes in future tree speciesstructure and composition (Ananda and Herath, 2003; Brouwer,1996; Cochrane et al., 1999; Holdsworth and Uhl, 1997; Nepstadet al., 1999; Palace et al, 2008a; Pereira et al., 2002; Pinard et al.,1996; Steege et al., 1995). Remote tools, in an integrated frame-work with field surveys and models, would allow independentoversight of jurisdictional forest management regulations andinternational greenhouse gas commitments (Alo and Pontius,2008; Bustamante et al., 2016).

Advances continue in our collective ability to monitor defor-estation and forest degradation using remote sensing tools(Frolking et al., 2009; Mitchell et al., 2017). However, uncertaintyin satellite-based estimates of land surface features is a functionof the length scales of the target features and the spatial and tem-poral resolution of the satellite sensor (Clark et al., 2004; Palaceet al., 2008b). Therefore, small-scale forest impacts on the canopy,such as a single tree fall or an under-canopy skid trail, can be dif-ficult or impossible to detect with moderate spatial resolution ima-gery where only a small proportion of an imaged pixel is affectedby the target feature (Hunter et al., 2015; Read et al., 2003). Addi-tionally, some logging impacts such as the loss of green cover asso-ciated with a felled tree are ephemeral, and infrequent repeatobservation opportunities over cloudy tropical regions make itchallenging to detect selective logging (Pereira et al., 2002; Pereset al., 2006). The earliest techniques employed to map selectivelogging relied on visual interpretation of Landsat Thematic Mapper(TM) imagery (Watrin and da Rocha, 1992) and annual forestchange monitoring products based on moderate resolution opticalimagery (Broich et al., 2011). Over the last two decades, large-region evaluation of selective logging has been demonstrated withmoderate resolution optical imagery (Asner et al., 2005; Souzaet al., 2005), but with uncertainty resulting from sensorlimitations.

Recent advances in research efforts to map selective logging intropical forests involve the use of airborne Light Detection AndRanging (LiDAR). LiDAR provides a measurement of the three-dimensional arrangement of trees, branches, and understory vege-

tation in forests, from the ground to the top of the canopy (Hosoiand Omasa, 2006; Palace et al., 2016). The two-dimensional aggre-gate of a LiDAR point cloud provides vertical distributions of vege-tation, referred to from here on as canopy profiles, which have alsobeen examined to estimate forest biometric properties, such asbiomass, stem density, and basal area (Asner et al., 2004; Dansonet al., 2007; Hopkinson et al., 2009; Hunter et al., 2015; Kentet al., 2015; Palace et al., 2015; Phua et al., 2016; Sullivan et al.,2014), as well as to delineate individual trees and tree gaps(Ferraz et al., 2016; Lefsky et al., 2007). The characteristics of acanopy profile can provide additional insights into forest condi-tions and, particularly when considering the proportion of mid-canopy and understory vegetation returns, may be indicative oflogging roads, skid trails, and felled trees (Ellis et al., 2016). Severalstudies have demonstrated that skid trails and logging gaps fromfelled trees can be identified through the use of LiDAR RelativeDensity Models (RDM; e.g. Andersen et al., 2014; d’Oliveira et al.,2012; Ellis et al., 2016). While these studies have demonstratedLiDAR can be an effective tool for quantifying logging, recent stud-ies suggest that LiDAR is currently not cost effective for integrationinto monitoring systems due to the high cost of both acquiring andprocessing the data (Ellis et al., 2016; Meyer et al., 2013). Thedevelopment of accurate automated methods for processing LiDARobservations could be of critical value because they would reducethe data processing costs, allowing for more efficient coverage oflarge areas while providing repeatable and consistent estimatesof important land surface attributes. An ability to quantify selectivelogging and estimate forest degradation proves key in understand-ing forest resource extraction, quantifying degradation, and aidingthe developing sustainable management practices.

In this paper, we focus on new automated methods to map andquantify the extent of selective logging damage at two concessionsin Kalimantan, Indonesia. These new methods rely on high spatialdensity airborne LiDAR data, developed in coordination with field-based measurements, and aerial photos of the extent and impact ofselectively logged areas. We present our results in terms of thearea of logging features, including skid trails, roads/decks, and gapsassociated with felled trees and collateral damage. Assessment ofthe automated approach is accomplished with a comparison toGPS-based observations and high-resolution aerial photography.

2. Sites and methods

We conducted our study in two regions in Kalimatan, Indonesiaon the island of Borneo (Fig. 1). The sites were chosen to representdifferent levels of terrain ruggedness and logging managementapproaches. Logging in these two concessions is planned to occurin a 35-year rotation cycle.

Timberdana is a 76,340-hectare concession located in thedistricts of Kutai Barat and Barito Utara in East Kalimantan(0� 50047.8900 S, 115� 23030.0300 E). The 2014 logging block of801 ha was covered by LiDAR imaging and aerial photos, andsampling of the logging impact was distributed throughout theannual logging area. The 2014 logging block is generally gentlyto moderately undulating terrain getting progressively steeperand more broken towards the western limits of the block.More than half of the area had been previously logged with theremaining area representing intact, unlogged forest. The com-pany’s logging machinery consists of Komatsu D85E-SS-2 crawlertractors typical of the selective logging machinery used elsewherein Kalimantan. To sell logs on the open market, the company willtarget select species and individuals with a regular, straight boleand abide by the 50-cm diameter limit as its harvesting criteria.However, the company has not practiced any form of ReducedImpact Logging (RIL) in Timberdana (Fig. 2).

Fig. 1. Two logging concessions (stars) logged in 2014 and imaged by LiDAR in 2014 were located in Kalimantan, the Indonesian section of the island of Borneo. Theconcessions were visited by field teams within four months of the LiDAR data collection.

Fig. 2. The canopy height model derived from LiDAR at the logged Timberdana concession shows evidence of logging features within the 2014 cut blocks (red polygons).Roads, logging decks, and canopy are apparent as dark regions, indicating low canopy heights. (For interpretation of the references to colour in this figure legend, the reader isreferred to the web version of this article.)

230 L. Melendy et al. / ISPRS Journal of Photogrammetry and Remote Sensing 139 (2018) 228–240

The Roda Mas concession is located on the upper reaches of theMahakam River in East Kalimantan (0� 58047.6500 N, 115� 03037.0600

E) (Fig. 3). The 2014 annual logging block of 1246 ha located inBlock C of the concession (on the West side of the Mahakam River)

was sampled. The logging company still harvests virgin forest inRoda Mas while attempting to follow RIL guidelines. The conces-sion has been certified by the Forest Stewardship Council for sus-tainable forest management. Terrain conditions are significantly

Fig. 3. The canopy height model derived from LiDAR at the logged Roda Mas concession shows evidence of logging features within the 2014 cut blocks (red polygons). Roads,logging decks, and canopy are apparent as dark regions, indicating low canopy heights. (For interpretation of the references to colour in this figure legend, the reader isreferred to the web version of this article.)


more rugged than in Timberdana. The selective logging performedin Roda Mas is constrained to a 50-cm diameter minimum limit.Once the logs are harvested, they are floated downriver to theindustrial complex of Tirta Mahakam located near Samarinda.Due to the inaccessibility of the company’s log dump on the riverby barges, sinker species are not harvested.

Logging in the 2014 cutting permit areas in the timber conces-sions of Timberdana and Roda Mas began in February and was con-cluded by early November. Sampling of logging impact was carriedout 10–24 December 2014 at the Timberdana site, and 17 Januaryto 15 February 2015 at the Roda Mas site, and was distributedthroughout the 2014 logging area under LiDAR coverage.

2.1. Field methods

The location of a sample of four types of logging features –roads, skid trails, log landings (or decks), and logging gaps thatresult from felled trees – were recorded using commercial hand-held Garmin GPS units (models 64S and 64SC) and converted tovector data (Table 1). The spatial uncertainty of individual trackpoints using these devices varies depending on conditions, butare reported by the company to be typically within 10 m (standarddeviation).

The shape of the gap associated with each felled tree was esti-mated by walking the perimeter of the gap with a GPS device. Roadcenterline locations were recorded with GPS, and sample road

Table 1Summary of field measurements in the two concessions.

Timberdana Roda Mas

Logging plots and gaps 58 62Roads length (and # of width

measurements)10 km (45) 21 km (42)

Skid trails length (and # of widthmeasurements)

9 km (2 4 8) 25 km (70)

Log landings (decks) 49 26

widths were collected an average of 350 m apart, measured intwo ways: road width, defined as the width of the proportion ofthe road on which a vehicle can travel as evidenced by tire tracks,and total width as measured from tree trunk to tree trunk perpen-dicular to the road direction (Fig. 4). Skid trail locations wererecorded with the GPS units, and sample skid trail widths weremeasured approximately 50 m apart in Timberdana and 300 mapart in Roda Mas. Road width and skid trail width were measuredwith a tape, while road total width was measured with a laserrange finder. The perimeters of the log landings were walked,and the measurements recorded with GPS.

Fig. 4. Logging road illustrating road width (yellow) and total width (red)measurements. (For interpretation of the references to colour in this figure legend,the reader is referred to the web version of this article.)


2.2. Airborne LiDAR and photo acquisition

Airborne LiDAR data, collected with an Optech ORION M300,and high-resolution aerial photos, collected using a Trimble 80-megapixel medium format digital aerial camera, were flown byPT Surtech Prima, a Jakarta, Indonesia-based company, betweenOctober and December 2014. LiDAR data and high resolution digi-tal color imagery covered the 2014 logged blocks in the two log-ging concessions as part of a larger data collection covering morethan 100,000 ha.

The LiDAR data were provided as LAS files in 500-m tiles createdby combining observations from overlapping flight paths. The dataachieved vertical accuracy of 0.08 m (standard error at 1 sigma) onclear ground. All data were supplied in Indonesian National Datum(DGN95) which is effectively the same as the WGS84 satellitedatum (ITRF 2000 Reference Frame). Vertical datum of the datawas geoidal height generated from the EGM08 spheroid/geoid sep-aration model. In Roda Mas, the LiDAR pulse density exceeded 10pulses per square meter while the density of point returns aver-aged 59 returns per square meter with a median of 43 returns. InTimberdana, the LiDAR pulse density exceeded 10 pulses persquare meter while return density averaged 54 returns per squaremeter with a median of 43 returns.

The digital photos were collected coincident with the LiDARobservations at an approximate resolution of 5 cm. The photoswere stitched into a tiled mosaic using Agisoft Photoscan (1.2.5).Photos were aligned using the high accuracy setting with genericpair selection. A dense point cloud was built using a medium set-ting and the output was as an orthomosaic image. Georeferencingwas done outside of Photoscan using LiDAR data.

2.3. LiDAR point data classification and digital elevation modelcreation

2.3.1. Raw point cloud dataRaw unclassified point clouds were delivered in 25-hectare

square tiles on a site-to-site basis. Data were processed usingopen source software developed as part of this study. Thesoftware, lidar2dems (https://github.com/Applied-GeoSolutions/lidar2dems), is a collection of open-source command line utilitiesbuilt upon the Point Data Abstraction Library (PDAL, http://www.pdal.io/) and its associated dependencies for point classification,and points2grid (https://github.com/CRREL/points2grid) for grid-ding returns into digital elevation models (DEMs).

2.3.2. Ground return classificationClassification of ground returns was performed using the clas-

sify script within lidar2dems. Points were classified using PDALground, which is built upon the Progressive Morphological Filter(PMF, Zhang et al. 2003). The PMF parameters, slope and cell size,were tuned based on a priori tests conducted in different terrainand vegetation coverage types. For ground classification, we usedfour broad classification groups that describe terrain and vegeta-tion within sub-areas of each site (1) non-forest with flat terrain:slope 1, cell size 3; (2) forest with flat terrain: slope 1, cell size2; (3) non-forest with complex terrain: slope 5, cell size 2; (4) for-est with complex terrain: slope 10, cell size 2. All tiles fallingwithin each sub-area were merged, then the merged LAS files wereclassified using the parameterization from the classification groupdefined by the sub-area.

2.3.3. Digital elevation model derivationMerged and classified LAS files were passed to the dem script

within lidar2dems, which was used to develop DEMs (digital ter-rain model, DTM; digital surface model, DSM; canopy heightmodel, CHM). LiDAR returns for each LAS file were filtered to use

ground returns to generate one-meter DTMs. Each DTMwas gener-ated using all ground-classified points in an XML pipeline withinPDAL, and was dependent on points2grid. For gridding returns inpoints2grid, we used inverse distance weighting of points withineach grid cell search radius. Grid cells with no ground returns wereprogressively filled using a gap-filling technique built into lidar2-dems, which iteratively increases the search radius used inpoints2grid and fills no data pixels with the smallest availablesearch radius for each pixel. Gaps were filled using search radiiof 0.56 m (the circumscribing circle radius of a 1-m grid cell),1.41 m (the radius of the circumscribing circle of a 2-m grid cell),2.5 m, and 3 m, and remaining missing data were filled using near-est neighbor interpolation. DSMs were generated using all non-ground returns in points2grid with a search radius of 0.56 m, usingthe maximum return value within each grid cell search radius. TheCHM was calculated as the difference between the DSM and theDTM. Each DEM was clipped to sub-area extents using Gippy(https://github.com/gipit/gippy) and merged with alike DEM typeswithin the site extent.

2.3.4. Return count voxelsFor each of the merged and classified LAS files, we created Geo-

Tiffs of cubic-meter voxels (volumetric pixels, Lefsky et al. 1999) ofreturn counts. Voxels were produced for summed return countsrelative to ground elevation on the same grid as each sub-areaDTM (1 m horizontal), both of which were determined from theDTM itself. All returns lower than the maximum canopy heightof each site and each coinciding CHM pixel were used to populatethe voxel dataset. Return count voxels were derived for 1-m verti-cal increments above each DTM pixel. Voxels were used becausethey offer a familiar format for remote sensing researchers thatcan be simply and rapidly used to generate analysis products.

2.3.5. Voxel productsRelative density models (RDM) of return counts were calculated

from voxel data in Python 2.7 (Python Software Foundation,Python version 2.7, www.python.org). In general, the RDM is calcu-lated for each pixel as:

RDM ¼PhN

hiR

PhNh0R

ð1Þ

where R is the number of returns in height class h, height classh0 is representative of classified and user-defined ground returns,and height classes hi, . . .hN are representative of above groundreturns in height classes from i to N, which define regions of inter-est in the vegetation profile. The RDM for each pixel ranges from 0to 1, and represents the proportion of returns occurring from thevertical region of interest, hi to hN, relative to all returns from theregion h0 to hN (Fig. 5). For pixels in which no returns occur fromh0 to hN, we defined the RDM pixel as having no data. An RDMwas calculated for our canopy region of interest, which can beindicative of skid trail logging impact (Andersen et al., 2014).

2.4. Automated mapping of logging features

We developed an automated methodology to extract loggingroads/decks, skid trails, and gaps from the LiDAR data. The auto-mated approach requires a canopy height model, digital terrainmodel, and voxels at 1 � 1 � 1-m resolution as input, and usesthese input sources to generate three 1-m binary raster layers.Parameters used in this method, including minimum clump sizeand height thresholds, were set based on observations derivedfrom canopy profiles (Fig. 6). The three raster layers generatedwere:

https://github.com/Applied-GeoSolutions/lidar2demshttps://github.com/Applied-GeoSolutions/lidar2demshttp://www.pdal.io/http://www.pdal.io/https://github.com/CRREL/points2gridhttps://github.com/gipit/gippyhttp://www.python.org

Fig. 5. Schematic diagram showing the relative density model used for identifying skid trails.

Fig. 6. Canopy height profiles for five different logging features from a region in the Roda Mas concession.


(1) Contiguous areas of extremely low canopy height (ELCH).All 1 � 1-m pixels in the CHM with a height of 0.5 m or lesswere isolated, and a binary raster layer was generated, with1 indicating the extremely low canopy height threshold wasmet and 0 indicating this low threshold was not met. Thisraster product was post-processed using a median filter

and a clump/sieve process that eliminated clumped adjacent(using four neighbors) low canopy height areas covering lessthan 450 m2.

(2) Contiguous areas with no understory (NU). A relative den-sity model map was created from the voxels to indicate theratio of LiDAR returns in 2- to 8-m vertical space to LiDAR


returns from the ground (0 m) up to 8 m. A binary map usinga maximum 1:100 threshold was generated from the RDM,meaning that 1 � 1-m pixels that had fewer than 1% ofreturns in the 2–8 m vertical space out of all returns fromthe ground to 8 m were mapped with a value of 1, whileall pixels with 1% or more of returns in this vertical spacewere assigned a value of 0. 1 � 1-m areas with no returnsin the 0–8 m space were also assigned a value of 0. This ras-ter product was post-processed using a median filter and aclump/sieve process that eliminated clumped adjacent(using four neighbors) areas meeting the defined thresholdcovering less than 90 m2.

(3) Contiguous areas of reduced canopy height in loggingrange (RCH). In a process analogous to ELCH above, all pixelsin the CHM with a height of 3 m or less were isolated, and abinary raster layer was generated with 1 indicating thereduced canopy threshold was met and 0 indicating thethreshold was not met (i.e. canopy was greater than 3 m).This raster product was post-processed using a median filter

Fig. 7. The three products extracted from the LiDAR voxels (ELCH (road raster); NU (skilogging (roads/decks in blue, skid trails in green, logging gaps in orange).

and a clump/sieve process that eliminated clumped adjacent(using four neighbors) low canopy height areas covering lessthan 20 m2. Pixels meeting these criteria, in addition tobeing within 50 m of a skid trail (Ellis et al., 2016), were clas-sified as contiguous areas of reduced canopy height in log-ging range.

Contiguous areas with no understory include both skid trailsand streams. To separate these features, a threshold value in theDTM model was identified for each concession and applied to theclassified rasters. All mapped contiguous areas with no understorybelow this threshold were identified as streams and removed fromthe logging features set.

Using these three raster products, we generated a map of:

(A) Roads and logging decks, which are defined as all pixels inthe ELCH product;

(B) Skid trails and logging decks, which are defined as all pixelsin the NU product that do not fall in the first product, ELCH;

d trail raster); RCH (gap raster)) are combined to map the features associated with


(C) Canopy gaps, which are all pixels in the RCH product that arenot also in either of the other two products.

These classified features were subsequently converted from ras-ters to polygons. The skid trails and logging decks class (B) was fur-ther refined into two separate classes via an additional rule set.Skid trails and logging deck features that fell immediately adjacentto the road (i.e., more than 50% of the area fell within 7 m of theroad) were reclassified into the roads/decks class. Those that didnot meet this criterion were classified as skid trails (Fig. 7). This cri-terion was added to the rule set because stacked trees on a deckoften have more than 99% of the returns below 2 m. The roadsand logging decks class (A) remained mixed and inseparable here.The effects of this conjoined class are minimal, particularly fromthe perspective of carbon accounting, as roads and log landingshave comparable carbon emission factors (Winrock, 2014; VCS,2015).

After the skid trails and road features identified in the auto-mated algorithm were converted to polygons, a module of ouralgorithm identified the centerline of these features, convertingthe skid trail and road polygons to lines. These lines were used in(a) the calculation of total linear extent of these features acrossthe logging block and (b) a comparison of the width of road andskid trail features produced from the automated algorithm to thewidths measured in the field.

2.5. Comparison of automated approach against GPS and photoevidence

The mapping of roads/decks and skid trails from the automatedalgorithm (AA) was compared to field measured GPS measure-ments and aerial photos (FM), which represented the only avail-able independent data for the spatial location and extent oflogging events at the Roda Mas and Timberdana sites over the per-iod of interest. A full validation of the results from automated rou-tine was not conducted here because of limitations in the referencedata. Therefore, the assessment was the agreement between thetwo datasets, rather than the usual accuracy of one dataset com-pared with another. The GPS measurements obtained during thesite visits and aerial photos acquired simultaneously with theLiDAR observations provided a basis for comparison withthe results from the automated routine, but they were limited.The GPS data had two limiting factors. First, the GPS data had sig-nificant spatial uncertainty, particularly at the Roda Mas conces-sion where dense canopy and topography reduce the accuracy ofthe GPS measurements. Second, the GPS measurements were notobtained comprehensively in any particular area. Instead, the fieldteam collected observations to reach targeted goals for number of

Fig. 8. GPS reference data of logging features were not highly spatially accurate and wedetected roads are brown polygons and the GPS measured road is a blue line).

features. This collection approach resulted in a complication aris-ing from the circumstance where the automated algorithm identi-fies a feature and there was no corresponding feature in thereference GPS. In these instances, it was unclear if the disagree-ment was due to (a) an algorithm error of commission or (b) thefact that the field team did not visit and record an existing feature.This complication was mitigated through a review of the aerialphotos in areas of potential error of commission. Roads and deckswere clear in the photos, as are most logging gaps. To evaluate skidtrails, we examined the area around the terminal end of the pre-dicted skid trail for presence of a gap, given that skid trails onlyexist to connect roads to extracted trees.

To prepare the data for this comparison, all road/decks and skidtrail features were assigned an ID, such that a given feature, whenpresent in both data sets, would have the same ID both in the auto-mated algorithm (AA) and field GPS data sets. This processinvolved an attempt by an analyst to locate a counterpart for eachfeature. For example, a section of road observed in the field maynot have overlapped exactly with one of the automated sections,but knowing that these sections were the same road, by visualinspection of shape and context, allowed for quantitativecomparison.

We used a python script to systematically determine Producer’sand Consumer’s agreement under various assumptions about thelocation accuracy of the features. The measure that drove the cal-culations used here is the intersected area of FM and AA features.Producer’s agreement is defined here as the FM and AA intersectedarea as a proportion of the area of the FM features, while Con-sumer’s agreement is the FM and AA intersected area as a propor-tion of the area of the AA features. Once again, the term‘‘agreement” is used here in place of ‘‘accuracy” because theground observations, while informative, were not sufficiently accu-rate or comprehensive to be considered truth for accuracyassessment.

To evaluate the dependence of spatial positional uncertainty onthe estimated level of agreement, a succession of buffers (0, 10, 20,and 30 m) was applied to the features of interest (roads/decks,skids trails), in addition to a base buffer width. Line features, i.e.road and skid trails, were assigned a base buffer width based onan estimate of their field measured mean width, while the deckswere not initially buffered because they were already defined aspolygon features (Fig. 8).

To summarize, in order to calculate Producer’s agreement foreach feature, the FM feature was held fixed and compared withbuffered AA features for each buffer width as mentioned above,reporting area of the FM feature, the area of intersection, andthe fraction represented by intersected area divided by the AA fea-ture area. For calculation of Consumer’s agreement, the converse

re often offset from the features detected by the automated algorithms (algorithm

Table 2Summary of FM logging features in the two concessions.

Timberdana Roda Mas

Total road mean width (SE) 16.6 m (1.9) 25.3 m (1.0)Road mean width (SE) 6.5 m (0.5) 7.6 m (0.2)Skid trail mean width (SE) 4.4 m (0.1) 4.2 m (0.2)Log landing mean area (SE) 0.12 ha (0.05) 0.08 ha (0.02)


condition was applied, i.e., the AA feature was held fixed and com-pared with the buffered FM feature. The total area of intersectionand total feature area were summed, and the fraction of overlapwas calculated and reported by feature type.

In some cases, there were automated algorithm detected log-ging features that were not associated with any field GPS features.To evaluate if these unmatched automated features were errors ofcommission, high resolution aerial imagery was consulted (Fig. 9).If there was definitive visual evidence of logging, these featureswere noted and included in the assessment. Road features areeasily determined from visual evidence. However, visually detect-ing skid trails, over which the forest canopy is often preserved, wasmade possible by identifying the existence of a notable gap at theend of the skid trail. The detection agreement in the Consumer col-umn was decremented by 100% of the area of features that did nothave associated GPS features or photo evidence of such features.

An exception was made for the converse case of existing butincorrect GPS positions for road and skid trail features. Five ofthe 106 road segments were removed, and three of the 207 exam-ined skid trail segments were removed, based on a quality controlevaluation of the GPS data. In three of these cases, duplicated andcontradictory GPS tracks labeled as both skid trail and road wereremoved from the analysis. In the other two cases, segments ofGPS-labeled road were not consistent with the photos, and weredeemed erroneous and removed from the analysis. These incorrect(duplicated or unmatched) GPS features were removed in order tomake the fairest comparison possible.

Photo evidence was reviewed across different domains, depend-ing on feature type. The domain for roads and decks was the entire2014 cut block in each concession, while the domain for skid trailswas a sub-region of the 2014 cut block in each concession. Theexamination of sub-regions for skid trails was necessitated dueto the extensive coverage of the skid trails and the time associatedwith visual inspection of each high-resolution image. The sub-regions were selected for examination at random, and representedapproximately 20% percent of the cut blocks in Timberdana andRoda Mas.

The area and width of the AA features were compared to the FMfeature area and width. All FM roads/decks were compared to theirAA counterparts, while the skid trail comparison was restricted tothe 20% examined.

Gaps identified by the automated algorithm were compared tothe gaps identified and delineated in the field using GPS. Addition-ally, each field measured gap was evaluated against the CHM by an

Fig. 9. Examples examining photo evidence in areas where no GPS data were collectphotographic evidence and (b) no photographic evidence of a skid trail.

analyst and, when the field digitized gap could be linked to a CHMvisible gap, the analyst hand-delineated the gap using the CHM as aguide. On a gap-by-gap basis, the area of each field measured gapwas compared with the area of its matching hand-digitized gap,and the Pearson’s correlation coefficient (R) between these areaswas reported. The fractional overlap between the automated algo-rithm gap and hand-digitized gaps was also calculated.

3. Results

Field measurements of logging features showed significant dif-ferences in road widths and log landing size between the two con-cessions, while the field measured skid trail widths between theconcessions were comparable (Table 2).

Agreement between the AA features and the FM and photo fea-tures increased with increasing feature buffer size. These increasesin agreement plateaued more quickly in Timberdana than in RodaMas, suggesting that GPS accuracy in Timberdana was higher thanin Roda Mas. At a buffer distance of 30 m, which is three times thestandard deviation of the typical GPS accuracy, the agreementbetween AA and FM skid trails exceeded 90% for both sites (Tables3 and 4). While Producer agreement for roads/decks at both sitesexceeded 80%, the Consumer agreement for these features were69% and 78% at Roda Mas and Timberdana, respectively.

The AA provided estimates of length, width, and area for theroad/deck and skid trail features. The AA did not explicitly separateroads from log landings, so the road area reported here is inclusiveof log landing area, and the reported average width of roads isinflated as a result.

In Timberdana, the average AA road width was greater (17.2 m)than FM road width (16.6 m). In Roda Mas, the AA road width(15.6 m) was substantially less than the FM width (26.6 m). AAskid trail widths exceeded the FM skid trail widths. In Timberdanaand Roda Mas, the AA skid trail widths averaged more than one

ed, but the automated algorithm identified a skid trail (pink line): (a) definitive


and a half times the FM skid trail widths (7.5 m vs. 4.4 m in Tim-berdana, and 7.6 m vs. 4.2 m in Roda Mas). FM canopy gap areawas moderately correlated with the hand-digitized gap area ingaps clearly identifiable in the CHM (R = 0.39). The overlap of thehand digitized gaps and the AA output showed mixed results. Halfof the hand-digitized gap area was not identified in the AA (50.3%),and a higher percentage of the hand digitized gap was identified asskid trail (27.3%) than was identified as gap (22.4%) in the AA.

Area of each logging feature was mapped by sub-block for thetwo concessions (Fig. 10). The AA-detected logging damage (i.e.sum of road/deck, skid trail, and gap area) totaled 174 ha at RodaMas and 203 ha at Timberdana, equivalent to 14.0% and 25.3% ofthe total area set aside for logging in each block. While the fractionof area affected by roads was similar at the two concessions (3.4%vs. 4.0%), skid trail and gap damage at Timberdana was twice asdense as at Roda Mas (21.3% vs. 10.5%) (Table 5).

4. Discussion

This study distinguished differences in logging impacts, as theTimberdana extent of logging impacts measured using the LiDARAA was nearly double that in Roda Mas, where Reduced ImpactLogging (RIL) practices were being implemented. In addition,intense topographic conditions at Roda Mas (Terrain RuggednessIndex, or TRI [Riley et al., 1999], of 0.53) limited accessibility insome areas, while topographic conditions did not limit access atTimberdana (TRI of 0.24). Furthermore, total gap area calculatedin Timberdana was considerably inflated by three clusters of clear-ing cut by local residents not employed by the logging concession.These clear-cuts are common in Indonesia, particularly in areaswith mining interests or areas where in-migration by illegal culti-vators creates conditions of establishing defacto ownership forpurposes of future compensation or livelihood, and are done witha mix of intents, including the attempt to increase access to areasfor mining (Intarini et al., 2014). If these three clusters, which con-tribute approximately 16 ha of additional gap area, were removedfrom the analysis, gap percentage in Timberdana fell to 5.1% (from7.1%) and total area affected by logging dropped to 23.3% (from25.3%).

Our results suggest that an automated approach to mappingLiDAR-derived features associated with logging compares

Table 3Level of agreement between FM features and AA features in Roda Mas.

Buffer size (m) Roda Mas

Roads/decks

Producer Consu

0 0.206 0.25710 0.532 0.50820 0.705 0.62830 0.802 0.686

Table 4Agreement between FM features and AA features in Timberdana.

Buffer size (m) Timberdana

Roads/decks

Producer Consu

0 0.422 0.47410 0.823 0.72820 0.900 0.76930 0.929 0.777

favorably to in-field measurement of these features. An automatedapproach can perform an analysis in a small fraction of the timethat is required by an in-field measurement team, and is a candi-date to replace a hand-digitizing RDM approach. Automated meth-ods can potentially offer more consistent and less biased resultsthan other approaches, considering it is not affected by fatigue orother effects that can often accompany in-field or hand-digitizingefforts. However, our results show that there is some potentialfor misclassification using automated methods.

Consumer agreement for roads/decks was lower than the agree-ment for skid trails, reflecting a tendency of the AA to classify heav-ily damaged areas as road that the field GPS team identified as skidtrails (Fig. 11). This occurred in both Timberdana (78% Consumer’sagreement) and Roda Mas (69% Consumer’s agreement). From theperspective of carbon monitoring, the frequent misclassification ofheavily damaged skid trails or logging gaps as roads is, perhaps, notrelevant. Indeed, the carbon impact of heavily damaged skid trailsin these areas is likely similar to the carbon impact of roads. Whilethere are places where secondary roads are abandoned, a key dif-ference is that roads are often maintained, while skid trails are leftto regrow. More research is needed to assign field-calibrated emis-sions factors to AA-derived impact zones and quantify loggingimpacts using carbon metrics (MgC emitted per ha accessed andm3 harvested).

The AA average road width matched closely with the total widthmeasured in the field in Timberdana, with a difference of less than1 m. In Roda Mas, the field measured total width was 11 m widerthan the algorithm average road width. Several factors influencedthe AA estimates and the FM estimates. The AA-based estimatesof road width were inflated by the inclusion of log-landing area.Alternatively, road width estimated from the algorithm did notinclude the area under canopy adjacent to or overhanging the road,resulting in a potential underestimate of road width in the algo-rithm. The FM-based estimates of total road width were recordedat set intervals using a laser range finder. Some of these set intervalwidth measurements were recorded at locations where roads anddecks meet, and resulted in widths exceeding 50 m. These roaddeck conflations were more prevalent in Roda Mas. Future algo-rithm improvements will include an added ruleset to identifycanopy overhanging the road for reclassification as road area.

The AA estimates of skid trail width exceeded FM widths bymore than 70% at both concessions. Skid trail widths estimated

Skid trails

mer Producer Consumer

0.252 0.6590.806 0.9390.901 0.9720.934 0.979

Skid trails

mer Producer Consumer

0.368 0.7930.917 0.9870.949 0.9920.960 0.992

Fig. 10. Roads and decks (brown), skid trails (green), and gaps (blue) at the 2014logged blocks of Roda Mas (a) and Timberdana (b). Black lines delineate sub-blockboundaries. Large contiguous gaps in Timberdana (b) are the result of illegalclearing conducted by residents not affiliated with the concession.

Table 5Area (ha) and % of total block area occupied by logging impacts measured by theLiDAR AA.

Roda Mas Timberdana

Roads & decks 42.0 (3.4%) 31.8 (4.0%)Skid trails 81.5 (6.5%) 113.5 (14.2%)Gaps 50.3 (4.0%) 57.2 (7.1%)Total logging 173.9 (14.0%) 202.5 (25.3%)

Table 6Comparison of cost across approaches for measuring logging impact.

Roads &trails

Roads, trails &gaps

Field sample (complete coverage) $0.25 to $2 –Field sample (10% of cut block) – $8 to $12LiDAR hand digitized (complete

coverage)$15 to $37 $25 to $47

LiDAR auto algorithm (completecoverage)

$11 to $36 $11 to $36


by the algorithm reach from tree stem to stem and had a spatialresolution of 1 m, while widths measured in the field are measuredbased on the width of the skidder evidence left in the soil. Areas ofopen understory under a closed canopy that are adjacent to theskid trail were included in the AA skid trail width.

The AA identification of gaps was the most challenging to eval-uate with the data collected for this analysis. By first linking field

Fig. 11. Example of (a) GPS identified skid trail and (b) classification

measured gaps to hand-digitized gaps, and then calculating theoverlap between the hand-digitized gaps and the automated algo-rithm gaps, we were able to partially overcome the limitationscaused by high location error in the hand-held GPS systemdeployed in the field. The boundaries and definitions between skidtrails and gaps are often blurred, and the comparison presentedhere reflects these uncertainties.

We estimate that the total cost of acquiring LiDAR data andapplying this automated method ranges from $11 to $36 per ha(Table 6), with the wide range primarily resulting from significantaircraft mobilization costs that represent a high fraction of the costwhen the area imaged is low. Hand-digitizing the roads and skidtrails using an RDM and CHM costs approximately $6 per ha inaddition to the base LiDAR costs, with higher costs associated withthe hand digitizing of canopy gaps. Collecting GPS data for all roadsand skid trails can cost a fraction of a dollar per hectare, but thecollection of detailed GPS and measurement information on gapscan significantly increase these costs even if the collection targetsa 10% sample of the cut block. The cost of LiDAR acquisition islikely to continue to decline and, if the methods outlined here pro-duce accurate results in other types of forest, it is conceivable thatthis automated approach can be applied across tropical forests as acost efficient and effective means for estimating the extent ofselective logging impacts. Linking the output from this automatedapproach with LiDAR- or field-derived biomass estimates andemission factors or more complex models (e.g. Piponiot et al.,2016) could provide a platform for REDD+ monitoring and RIL Car-bon certification. Building from these methods, region-wide esti-mates from a stratified-random sample of LiDAR measurementscould be integrated with information from new operational remotesensing platforms (e.g. GEDI, NiSAR, BIOMASS) to significantlyreduce uncertainty associated with the effects of logging.

5. Conclusions

The automatedmethods described here are effective at identify-ing the roads, decks, skid trails, and gaps associated with selectivelogging, and offer a reliable and potentially scalable approach forwide-area mapping of logging impact. This tool can be used formonitoring and measuring the impact of selective logging prac-tices, both legally permitted logging and illegal logging. When

by the AA as road in Roda Mas due to complete tree removal.


combined with estimates of carbon emission factors, the algorithmholds potential as a tool for use in quantifying and distinguishingcarbon impacts of selective logging methods.

Acknowledgements

This study was funded by a grant from NASA’s Carbon Monitor-ing System Program (grant # NNX13AP88G). We are grateful totwo anonymous reviewers who provided constructive recommen-dations that have substantially improved the presentation of thisresearch.

References

Alo, C.A., Pontius Jr, R.G., 2008. Identifying systematic land-cover transitions usingremote sensing and GIS: the fate of forests inside and outside protected areas ofSouthwestern Ghana. Environ. Plann. B: Plann. Des. 35 (2), 280–295. https://doi.org/10.1068/b32091.

Ananda, J., Herath, G., 2003. Soil erosion in developing countries: a socio-economicappraisal. J. Environ. Manage. 68 (4), 343–353. https://doi.org/10.1016/S0301-4797(03)00082-3.

Andersen, H.E., Reutebuch, S.E., McGaughey, R.J., d’Oliveira, M.V., Keller, M., 2014.Monitoring selective logging in western Amazonia with repeat lidar flights.Remote Sens. Environ. 151, 157–165. https://doi.org/10.1016/j.rse.2013.08.049.

Asner, G.P., Knapp, D.E., Broadbent, E.N., Oliveira, P.J., Keller, M., Silva, J.N., 2005.Selective logging in the Brazilian Amazon. Science 310 (5747), 480–482. https://doi.org/10.1126/science.1118051.

Asner, G.P., Keller, M., Pereira, J.R., Zweede, J.C., Silva, J.N.M., 2004. Canopy damageand recovery after selective logging in Amazonia: field and satellite studies.Ecol. Appl. 14, 280–298.

Blaser, J., Sarre, A., Poore, D., Johnson, S., 2011. Status of Tropical ForestManagement 2011. ITTO Technical Series, 38, International Tropical TimberOrganization, Yoko-hama, Japan.

Broich, M., Hansen, M., Stolle, F., Potapov, P., Margono, B.A., Adusei, B., 2011.Remotely sensed forest cover loss shows high spatial and temporal variationacross Sumatera and Kalimantan, Indonesia 2000–2008. Environ. Res. Lett. 6 (1),014010. https://doi.org/10.1088/1748-9326/6/1/014010.

Brouwer, L.C., 1996. Nutrient Cycling in Pristine and Logged Tropical Rain Forest: AStudy in Guyana. Tropenbos Guyana Programme.

Bush, M.B., McMichael, C.H., Piperno, D.R., Silman, M.R., Barlow, J., Peres, C.A.,Power, M., Palace, M.W., 2015. Anthropogenic influence on Amazonian forestsin pre-history: an ecological perspective. J. Biogeogr. 42 (12), 2277–2288.https://doi.org/10.1111/jbi.1263.

Bustamante, M., Roitman, I., Aide, T.M., Alencar, A., Anderson, L.O., Aragão, L., Asner,G.P., Barlow, J., Berenguer, E., Chambers, J., Costa, M.H., 2016. Toward anintegrated monitoring framework to assess the effects of tropical forestdegradation and recovery on carbon stocks and biodiversity. Global ChangeBiol. 22 (1), 92–109. https://doi.org/10.1111/gcb.13087.

Chambers, J.Q., Asner, G.P., Morton, D.C., Anderson, L.O., Saatchi, S.S., Espírito-Santo,F.D., Palace, M., Souza, C., 2007. Regional ecosystem structure and function:ecological insights from remote sensing of tropical forests. Trends Ecol. Evol. 22(8), 414–423. https://doi.org/10.1016/j.tree.2007.05.001.

Clark, D.B., Castro, C.S., Alvarado, L.D.A., Read, J.M., 2004. Quantifying mortality oftropical rain forest trees using high-spatial-resolution satellite data. Ecol. Lett. 7(1), 52–59. DOI: 0.1046/j.1461-0248.2003.00547.x.

Cochrane, M.A., Alencar, A., Schulze, M.D., Souza, C.M., Nepstad, D.C., Lefebvre, P.,Davidson, E.A., 1999. Positive feedbacks in the fire dynamic of closed canopytropical forests. Science 284 (5421), 1832–1835. https://doi.org/10.1126/science.284.5421.1832.

Danson, F.M., Hetherington, D., Morsdorf, F., Koetz, B., Allgöwer, B., 2007. Forestcanopy gap fraction from terrestrial laser scanning. Geosci. Remote Sens. Lett. 4(1), 157–160.

DeFries, R., Asner, G., Achard, F., Justice, C., Laporte, N., Price, K., Small, C.,Townshend, J., 2005. Monitoring tropical deforestation for emerging carbonmarkets. Tropical Deforestation Climate Change, 35–44.

d’Oliveira, M.V., Reutebuch, S.E., McGaughey, R.J., Andersen, H.E., 2012. Estimatingforest biomass and identifying low-intensity logging areas using airbornescanning lidar in Antimary State Forest, Acre State, Western Brazilian Amazon.Remote Sens. Environ. 124, 479–491. https://doi.org/10.1016/j.rse.2012.05.014.

Ellis, P., Griscom, B., Walker, W., Gonçalves, F., Cormier, T., 2016. Mapping selectivelogging impacts in Borneo with GPS and airborne lidar. For. Ecol. Manage. 365,184–196. https://doi.org/10.1016/j.foreco.2016.01.020.

Ferraz, A., Saatchi, S., Mallet, C., Meyer, V., 2016. Lidar detection of individual treesize in tropical forests. Remote Sens. Environ. 183, 318–333. https://doi.org/10.1016/j.rse.2016.05.028.

Frolking, S., Palace, M.W., Clark, D.B., Chambers, J.Q., Shugart, H.H., Hurtt, G.C., 2009.Forest disturbance and recovery: a general review in the context of spaceborneremote sensing of impacts on aboveground biomass and canopy structure. J.Geophys. Res. Biogeosci. 114 (G2). https://doi.org/10.1029/2008JG000911.

Hansen, M.C., Stehman, S.V., Potapov, P.V., Loveland, T.R., Townshend, J.R., DeFries,R.S., Pittman, K.W., Arunarwati, B., Stolle, F., Steininger, M.K., Carroll, M., 2008.Humid tropical forest clearing from 2000 to 2005 quantified by using

multitemporal and multiresolution remotely sensed data. Proc. Natl. Acad.Sci. 105 (27), 9439–9444. https://doi.org/10.1073/pnas.0804042105.

Holdsworth, A.R., Uhl, C., 1997. Fire in Amazonian selectively logged rain forest andthe potential for fire reduction. Ecol. Appl. 7 (2), 713–725. https://doi.org/10.1890/1051-0761(1997) 007[0713:FIASLR]2.0.CO;2.

Hopkinson, C., Chasmer, L., 2009. Testing LiDAR models of fractional cover acrossmultiple forest ecozones. Remote Sens. Environ. 113, 275–288.

Hosoi, F., Omasa, K., 2006. Voxel-based 3-D modeling of individual trees forestimating leaf area density using high-resolution portable scanning lidar. IEEETrans. Geosci. Remote Sens. 44 (12), 3610–3618. https://doi.org/10.1109/TGRS.2006.881743.

Hunter, M.O., Keller, M., Morton, D., Cook, B., Lefsky, M., Ducey, M., Saleska, S., deOliveira Jr, R.C., Schietti, J., 2015. Structural dynamics of tropical moist forestgaps. PloS One 10 (7), e0132144. https://doi.org/10.1371/journal.pone.0132144.

Intarini, D.Y., Resosudarmo, I.A.P., Komalasari, M., Ekaputri, A.D., Agustavia, M.,2014. Case Report: Ketapang Community Carbon Pools, West Kalimantan,Indonesia. In: Sills, E.O., Atmadja, S., Duchelle, Sassi, C. de A.E., Kweka, D.,Resosudarmo, I.A.P., Sunderlin, W.D., (Eds.) REDD+ on the Ground: A Case Bookof Subnational Initiatives Across the Globe. pp. 329–347.

Kent, R., Lindsell, J., Laurin, G., Valentini, R., Coomes, D., 2015. Airborne LiDARdetects selectively logged tropical forest even in an advanced stage of recovery.Remote Sens. 7, 8348–8367.

Lawson, S., MacFaul, L., 2010. Illegal Logging and Related Trade: Indicators of theGlobal Response. Chatham House, London.

Lefsky, M.A., Cohen, W.B., Acker, S.A., Parker, G.G., Spies, T.A., Harding, D., 1999.Lidar remote sensing of the canopy structure and biophysical properties ofDouglas-fir western hemlock forests. Remote Sens. Environ. 70 (3), 339–361.https://doi.org/10.1016/S0034-4257(99)00052-8.

Lefsky, M.A., Keller, M., Panga, Y., de Camargod, P.B., Hunter, M.O., 2007. Revisedmethod for forest canopy height estimation from Geoscience Laser AltimeterSystem waveforms. J. Appl. Remote Sens. https://doi.org/10.1117/1.2795724.

Meyer, V., Saatchi, S.S., Chave, J., Dalling, J.W., Bohlman, S., Fricker, G.A., Robinson, C.,Neumann, M., Hubbell, S., 2013. Detecting tropical forest biomass dynamicsfrom repeated airborne lidar measurements. Biogeosciences 10 (8), 5421–5438.https://doi.org/10.5194/bg-10-5421-2013.

Mitchell, A.L., Rosenqvist, A., Mora, B., 2017. Current remote sensing approaches tomonitoring forest degradation in support of countries measurement, reportingand verification (MRV) systems for REDD+. Carbon Balance Manage. 12 (1), 9.

Nepstad, D.C., Veríssimo, A., Alencar, A., Nobre, C., 1999. Large-scaleimpoverishment of Amazonian forests by logging and fire. Nature 398 (6727),505. https://doi.org/10.1038/19066.

Palace, M., Keller, M., Silva, H., 2008a. Necromass production: studies inundisturbed and logged Amazon forests. Ecol. Appl. 18 (4), 873–884.

Palace, M., Keller, M., Asner, G.P., Hagen, S., Braswell, B., 2008b. Amazon foreststructure from IKONOS satellite data and the automated characterization offorest canopy properties. Biotropica 40 (2), 141–150. https://doi.org/10.1111/j.1744-7429.2007.00353.x.

Palace, M.W., Sullivan, F.B., Ducey, M.J., Treuhaft, R.N., Herrick, C., Shimbo, J.Z.,Mota-E-Silva, J., 2015. Estimating forest structure in a tropical forest using fieldmeasurements, a synthetic model and discrete return lidar data. Remote Sens.Environ. 161, 1–11. https://doi.org/10.1016/j.rse.2015.01.020.

Palace, M., Sullivan, F.B., Ducey, M., Herrick, C., 2016. Estimating tropical foreststructure using a terrestrial lidar. PloS One 11 (4), e0154115. https://doi.org/10.1371/journal.pone.0154115.

Pearson, T.R., Brown, S., Murray, L., Sidman, G., 2017. Greenhouse gas emissionsfrom tropical forest degradation: an underestimated source. Carbon BalanceManage. 12 (1), 3. https://doi.org/10.1186/s13021-017-0072-2.

Pereira, R., Zweede, J., Asner, G.P., Keller, M., 2002. Forest canopy damage andrecovery in reduced-impact and conventional selective logging in eastern Para,Brazil. Forest Ecol. Manage. 168 (1), 77–89. https://doi.org/10.1016/S0378-1127(01)00732-0.

Peres, C.A., Barlow, J., Laurance, W.F., 2006. Detecting anthropogenic disturbancein tropical forests. Update 21 (5), 227. https://doi.org/10.1016/j.tree.2006.03.007.

Phua, M.H., Hue, S.W., Ioki, K., Hashim, M., Bidin, K., Musta, B., Suleiman, M., Yap, S.W., Maycock, C.R., 2016. Estimating logged-over lowland rainforestaboveground biomass in Sabah, Malaysia using airborne LiDAR data.Terrestrial, Atmospheric Oceanic Sci. 27 (4), 481–489. https://doi.org/10.3319/TAO.2016.01.06.02(ISRS).

Pinard, M.A., Putz, F.E., 1996. Retaining forest biomass by reducing logging damage.Biotropica, 278–295. https://doi.org/10.2307/2389193.

Piponiot, C., Sist, P., Mazzei, L., Peña-Claros, M., Putz, F.E., Rutishauser, E., Shenkin,A., Ascarrunz, N., De Azevedo, C.P., Baraloto, C., França, M., 2016. Carbonrecovery dynamics following disturbance by selective logging in Amazonianforests. ELife 5. https://doi.org/10.7554/eLife. 21394.

Putz, F.E., Zuidema, P.A., Synnott, T., Peña-Claros, M., Pinard, M.A., Sheil, D., Vanclay,J.K., Sist, P., Gourlet-Fleury, S., Griscom, B., Palmer, J., 2012. Sustainingconservation values in selectively logged tropical forests: the attained and theattainable. Conserv. Lett. 5 (4), 296–303. https://doi.org/10.1111/j.1755-263X.2012.00242.x.

Read, J.M., Clark, D.B., Venticinque, E.M., Moreira, M.P., 2003. Application of merged1-m and 4-m resolution satellite data to research and management in tropicalforests. J. Appl. Ecol. 40 (3), 592–600. https://doi.org/10.1046/j.1365-2664.2003.00814.x.

Riley, S.J., DeGloria, S.D., Elliot, R., 1999. A terrain ruggedness index that quantifiestopographic heterogeneity. Int. J. Sci. 5 (1–4), 23–27.

https://doi.org/10.1068/b32091https://doi.org/10.1068/b32091https://doi.org/10.1016/S0301-4797(03)00082-3https://doi.org/10.1016/S0301-4797(03)00082-3https://doi.org/10.1016/j.rse.2013.08.049https://doi.org/10.1126/science.1118051https://doi.org/10.1126/science.1118051http://refhub.elsevier.com/S0924-2716(18)30057-1/h0025http://refhub.elsevier.com/S0924-2716(18)30057-1/h0025http://refhub.elsevier.com/S0924-2716(18)30057-1/h0025https://doi.org/10.1088/1748-9326/6/1/014010http://refhub.elsevier.com/S0924-2716(18)30057-1/h0045http://refhub.elsevier.com/S0924-2716(18)30057-1/h0045https://doi.org/10.1111/jbi.1263https://doi.org/10.1111/gcb.13087https://doi.org/10.1016/j.tree.2007.05.001http://refhub.elsevier.com/S0924-2716(18)30057-1/h0065http://refhub.elsevier.com/S0924-2716(18)30057-1/h0065http://refhub.elsevier.com/S0924-2716(18)30057-1/h0065https://doi.org/10.1126/science.284.5421.1832https://doi.org/10.1126/science.284.5421.1832http://refhub.elsevier.com/S0924-2716(18)30057-1/h0075http://refhub.elsevier.com/S0924-2716(18)30057-1/h0075http://refhub.elsevier.com/S0924-2716(18)30057-1/h0075http://refhub.elsevier.com/S0924-2716(18)30057-1/h0080http://refhub.elsevier.com/S0924-2716(18)30057-1/h0080http://refhub.elsevier.com/S0924-2716(18)30057-1/h0080https://doi.org/10.1016/j.rse.2012.05.014https://doi.org/10.1016/j.foreco.2016.01.020https://doi.org/10.1016/j.rse.2016.05.028https://doi.org/10.1016/j.rse.2016.05.028https://doi.org/10.1029/2008JG000911https://doi.org/10.1073/pnas.0804042105https://doi.org/10.1890/1051-0761(1997)007[0713:FIASLR]2.0.CO;2https://doi.org/10.1890/1051-0761(1997)007[0713:FIASLR]2.0.CO;2http://refhub.elsevier.com/S0924-2716(18)30057-1/h0135http://refhub.elsevier.com/S0924-2716(18)30057-1/h0135https://doi.org/10.1109/TGRS.2006.881743https://doi.org/10.1109/TGRS.2006.881743https://doi.org/10.1371/journal.pone.0132144http://refhub.elsevier.com/S0924-2716(18)30057-1/h0160http://refhub.elsevier.com/S0924-2716(18)30057-1/h0160http://refhub.elsevier.com/S0924-2716(18)30057-1/h0160http://refhub.elsevier.com/S0924-2716(18)30057-1/h0165http://refhub.elsevier.com/S0924-2716(18)30057-1/h0165https://doi.org/10.1016/S0034-4257(99)00052-8https://doi.org/10.1117/1.2795724https://doi.org/10.5194/bg-10-5421-2013http://refhub.elsevier.com/S0924-2716(18)30057-1/h0185http://refhub.elsevier.com/S0924-2716(18)30057-1/h0185http://refhub.elsevier.com/S0924-2716(18)30057-1/h0185https://doi.org/10.1038/19066http://refhub.elsevier.com/S0924-2716(18)30057-1/h0195http://refhub.elsevier.com/S0924-2716(18)30057-1/h0195https://doi.org/10.1111/j.1744-7429.2007.00353.xhttps://doi.org/10.1111/j.1744-7429.2007.00353.xhttps://doi.org/10.1016/j.rse.2015.01.020https://doi.org/10.1371/journal.pone.0154115https://doi.org/10.1371/journal.pone.0154115https://doi.org/10.1186/s13021-017-0072-2https://doi.org/10.1016/S0378-1127(01)00732-0https://doi.org/10.1016/S0378-1127(01)00732-0https://doi.org/10.1016/j.tree.2006.03.007https://doi.org/10.1016/j.tree.2006.03.007https://doi.org/10.3319/TAO.2016.01.06.02(ISRS)https://doi.org/10.3319/TAO.2016.01.06.02(ISRS)https://doi.org/10.2307/2389193https://doi.org/10.7554/eLife.21394https://doi.org/10.1111/j.1755-263X.2012.00242.xhttps://doi.org/10.1111/j.1755-263X.2012.00242.xhttps://doi.org/10.1046/j.1365-2664.2003.00814.xhttps://doi.org/10.1046/j.1365-2664.2003.00814.xhttp://refhub.elsevier.com/S0924-2716(18)30057-1/h0260http://refhub.elsevier.com/S0924-2716(18)30057-1/h0260


Ryan, S.J., Palace, M.W., Hartter, J., Diem, J.E., Chapman, C.A., Southworth, J., 2017.Population pressure and global markets drive a decade of forest cover change inAfrica’s Albertine Rift. Appl. Geogr. 81, 52–59. https://doi.org/10.1016/j.apgeog.2017.02.009.

Souza, Carlos M., Roberts, Dar A., Cochrane, Mark A., 2005. Combining spectral andspatial information to map canopy damage from selective logging and forestfires. Remote Sens. Environ. 98 (2), 329–343. https://doi.org/10.1016/j.rse.2005.07.013.

Sullivan, F.B., Palace, M., Ducey, M., 2014. Multivariate statistical analysis ofasynchronous lidar data and vegetation models in a neotropical forest. RemoteSens. Environ. 154, 368–377. https://doi.org/10.1016/j.rse.2014.04.027. ISSN0034-4257.

Steege, H.T., Boot, R., Brouwer, L., Hammond, D., Van Der Hout, P., Jetten, V.G., Khan,Z., Polak, A.M., Raaimakers, D., Zagt, R., 1995. Basic and applied research forsound rain forest management in Guyana. Ecol. Appl. 5 (4), 904–910. https://doi.org/10.2307/2269341.

Verified Carbon Standard (VCS), 2015. Reduced Impact Logging Practices thatReduce Carbon Emissions (RIL-C) Methodology [WWW Document]. URL (accessed 6.23.15).

Watrin, O.D.S., da Rocha, A.M.A., 1992. Levantamento da vegetação natural e do usoda terra no município de Paragominas (PA) utilizando imagens TM/LANDSAT.Embrapa Amazônia Oriental-Séries anteriores (INFOTECA-E).

Winrock International, 2014. AFOLU Carbon Calculator. The Forest Managementtool: Underlying Data and Methods. Prepared by Winrock International underthe Cooperative Agreement No. EEM-A-00-06-00024-00.

Zhang, K., Chen, S.C., Whitman, D., Shyu, M.L., Yan, J., Zhang, C., 2003. A progressivemorphological filter for removing nonground measurements from airborneLIDAR data. IEEE Trans. Geosci. Remote Sens. 41 (4), 872–882. https://doi.org/10.1109/TGRS.2003.810682.

https://doi.org/10.1016/j.apgeog.2017.02.009https://doi.org/10.1016/j.apgeog.2017.02.009https://doi.org/10.1016/j.rse.2005.07.013https://doi.org/10.1016/j.rse.2005.07.013https://doi.org/10.1016/j.rse.2014.04.027.ISSN0034-4257https://doi.org/10.1016/j.rse.2014.04.027.ISSN0034-4257https://doi.org/10.2307/2269341https://doi.org/10.2307/2269341http://www.v-c-s.org/methodologies/reduced-impact-logging-practices-reduce-carbon-emissions-ril-c-methodologyhttp://www.v-c-s.org/methodologies/reduced-impact-logging-practices-reduce-carbon-emissions-ril-c-methodologyhttp://refhub.elsevier.com/S0924-2716(18)30057-1/h0295http://refhub.elsevier.com/S0924-2716(18)30057-1/h0295http://refhub.elsevier.com/S0924-2716(18)30057-1/h0295https://doi.org/10.1109/TGRS.2003.810682https://doi.org/10.1109/TGRS.2003.810682

Automated method for measuring the extent of selective logging damage with airborne LiDAR data1 Introduction2 Sites and methods2.1 Field methods2.2 Airborne LiDAR and photo acquisition2.3 LiDAR point data classification and digital elevation model creation2.3.1 Raw point cloud data2.3.2 Ground return classification2.3.3 Digital elevation model derivation2.3.4 Return count voxels2.3.5 Voxel products

2.4 Automated mapping of logging features2.5 Comparison of automated approach against GPS and photo evidence

3 Results4 Discussion5 ConclusionsAcknowledgementsReferences

isprs journal of photogrammetry and remote...

Documents