An empirical, integrated forest biomass monitoring system · turbance and recovery dynamics...

Environmental Research Letters

LETTER bull OPEN ACCESS

An empirical integrated forest biomass monitoring systemTo cite this article Robert E Kennedy et al 2018 Environ Res Lett 13 025004

View the article online for updates and enhancements

This content was downloaded from IP address 1666135166 on 07022018 at 2051

Environ Res Lett 13 (2018) 025004 httpsdoiorg1010881748-9326aa9d9e

LETTER

An empirical integrated forest biomass monitoringsystem

Robert E Kennedy119 Janet Ohmann2 Matt Gregory3 Heather Roberts4 Zhiqiang Yang5 David M Bell6Van Kane7 M Joseph Hughes8 Warren B Cohen9 Scott Powell10 Neeti Neeti11 Tara Larrue12 SamHooper13 Jonathan Kane14 David L Miller15 James Perkins16 Justin Braaten17 and Rupert Seidl18

1 College of Earth Ocean and Atmospheric Sciences Oregon State University United States of America2 Emeritus Scientist Pacific Northwest Forestry Sciences Lab USDA Forest Service United States of America3 College of Forestry Oregon State University United States of America4 College of Forestry Oregon State University United States of America5 College of Forestry Oregon State University United States of America6 USDA Forest Service Pacific Northwest Research Station United States of America7 School of Environmental and Forest Sciences University of Washington United States of America8 College of Earth Ocean and Atmospheric Sciences Oregon State University United States of America9 USDA Forest Service Pacific Northwest Forestry Sciences Lab United States of America10 Department of Land Resources and Environmental Sciences Montana State University United States of America11 Department of Geography Boston University Now at Department of Natural Resources TERI University India12 College of Earth Ocean and Atmospheric Sciences Oregon State University United States of America13 College of Earth Ocean and Atmospheric Sciences Oregon State University United States of America14 School of Environmental and Forest Sciences University of Washington United States of America15 Department of Geography Boston University Now at Department of Geography University of California Santa Barbara United States

of America16 Department of Geography Boston University Now at University of New Hampshire United States of America17 College of Earth Ocean and Atmospheric Sciences Oregon State University United States of America18 Institute of Silviculture Universty of Natural Resources and Life Sciences Vienna Austria19 Author to whom any correspondence should be addressed

OPEN ACCESS

RECEIVED

18 July 2017

REVISED

17 November 2017

ACCEPTED FOR PUBLICATION

28 November 2017

PUBLISHED

1 February 2018

Original content fromthis work may be usedunder the terms of theCreative CommonsAttribution 30 licence

Any further distributionof this work mustmaintain attribution tothe author(s) and thetitle of the work journalcitation and DOI

E-mail rkennedycoasoregonstateedu

Keywords forest biomass Landsat forest inventory lidar monitoring disturbance

Supplementary material for this article is available online

AbstractThe fate of live forest biomass is largely controlled by growth and disturbance processes both naturaland anthropogenic Thus biomass monitoring strategies must characterize both the biomass of theforests at a given point in time and the dynamic processes that change it Here we describe and test anempirical monitoring system designed to meet those needs Our system uses a mix of field datastatistical modeling remotely-sensed time-series imagery and small-footprint lidar data to build andevaluate maps of forest biomass It ascribes biomass change to specific change agents and attempts tocapture the impact of uncertainty in methodology We find that

bull A common image framework for biomass estimation and for change detection allows for consistentcomparison of both state and change processes controlling biomass dynamics

bull Regional estimates of total biomass agree well with those from plot data alone

bull The system tracks biomass densities up to 450ndash500 Mg haminus1 with little bias but begins underesti-mating true biomass as densities increase further

bull Scale considerations are important Estimates at the 30 m grain size are noisy but agreement atbroad scales is good Further investigation to determine the appropriate scales is underway

bull Uncertainty from methodological choices is evident but much smaller than uncertainty based onchoice of allometric equation used to estimate biomass from tree data

bull In this forest-dominated study area growth and loss processes largely balance in most years withloss processes dominated by human removal through harvest In years with substantial fire activityhowever overall biomass loss greatly outpaces growth

Taken together our methods represent a unique combination of elements foundational to anoperational landscape-scale forest biomass monitoring program

copy 2018 The Author(s) Published by IOP Publishing Ltd

Environ Res Lett 13 (2018) 025004

1 Introduction

Predicting the fate of carbon in forests under futureclimates is a fundamental scientific and managementchallenge as these systems contain large reservoirs ofcarbon provide many essential ecosystem services andrepresent a potentially critical feedback in global cli-mate change (Birdsey et al 2007) Yet carbon storage inforested ecosystems can be highly dynamic affectedby diverse anthropogenic and natural disturbanceprocesses that canradically change the carbon trajectoryof a landscape (Chambers et al 2007 Environment-Canada 2006 Williams et al 2016) Models anddecision support tools that cannot realistically incorpo-rate the full range of disturbance and growth processescan only provide a snapshot of the carbon dynamicsand may not fulfill the needs of the public or of policymakers (Joyce et al 2009) Moreover neither mod-els nor management plans can exist in an historicalvacuum both must consider how the potential futureconditions relate to past and current trends and status(Birdsey 2006)

Ourunderstandingofpossible forest carbonfuturesis hampered by our sparse observations of the recentcarbon past In the United States the core tool usedfor national and international reporting of forest car-bon fluxes (EPA 2016) has traditionally been the fieldmeasurements collected by the Forest Inventory andAnalysis (FIA) program (Goodale et al 2002 Wood-bury et al 2007) Though FIA measurements are nowrecorded consistently cover the contiguous US and arestatistically defensible the goal of any such inventoryprogram is an estimate of growing stock or volume forlarge areas (McRoberts et al 2014) This result may notbe sufficient for managers and modelers who need spa-tially explicit maps of carbon and carbon change at thescale of management (Sannier et al 2016) Additionallythe decadal repeat period of FIA plot measurements (inthe western US) is longer than that needed to captureanthropogenic and natural processes that significantlyalter carbon pools every year (Houghton 2005)

Other tools have the potential to complement FIAplot information in capturing actual forest carbon con-ditions but none can fully address carbon dynamicsalone Maps based on small-footprint lidar (light detec-tion and ranging) data have been shown to provideexcellent estimates of live aboveground forest biomassand hence carbon (Hudak et al 2002 Kim et al 2009Naesset 2002 Pflugmacher et al 2014) but these dataare generally available for relatively small areas andeven in the future are unlikely to be repeated fre-quently enough over large areas to capture disturbanceprocesses in a monitoring framework Maps of forestdisturbance from Landsat satellite imagery have fineenough spatial resolution to capture individual patchesof disturbance and growth (ie 30 times 30 m) and alsoto meet the criterion of covering large areas (Hansenand Loveland 2012 Huang et al 2010) Landsat-based maps of forest dynamics have traditionally had

greatest success focusing on high-severity abrupt dis-turbances such as clearcuts and fires (Cohen et al 2002Masek et al 2013)

While characterizing abrupt disturbance processesis critical for understanding biomass dynamics mapsshowing the timing of such disturbance processes aloneare insufficient for biomass monitoring They omitkey lower-severity anthropogenic disturbances suchas mechanical thinning subtle or long-lasting natu-ral disturbances caused by insects and drought (Cohenet al 2016 McDowell et al 2015) and slow aggradationprocesses such as forest growth Additionally maps ofdisturbance timing alone provide no indication of thebiomass status before the disturbance nor of the agentthat caused itmdashboth of which can substantially alter thebiomass consequences of a given change (Williams et al2016) Other efforts to map biomass or disturbance atbroad spatial extents from satellite-based data (opticalradar and lidar) have generally focused on single-dateestimates of biomass stock (Goetz et al 2009 Kellndor-fer et al 2000 Swatantran et al 2011 Zheng et al 2004)or have spatial grains too large to adequately capturechange caused by management (Blackard et al 2008Wilson et al 2013) Uncertainties introduced through-out the modeling process are rarely imparted to usersdespite the recognized need for such estimates (Saatchiet al 2011 Sexton et al 2015 Williams et al 2016)

Herewedescribe anempirical forest biomassmon-itoring system that meets many of these challenges Welink FIA plot data with time-series analysis of satel-lite imagery to create yearly maps of forest biomass atthe regional scale and link changes in those biomassmaps to disturbance and growth dynamics In the pro-cess we examine methodological uncertainties andutilize high-quality lidar-based local-scale maps ofbiomass to understand robustness of our regional esti-mates Taken together the yearly estimates of biomassbiomass uncertainty and cause of change delivered byour monitoring system can provide a nuanced under-standing of the spatial and temporal patterns of forestbiomass dynamics at a regional scale

2 Methods

21 OverviewThe biomass monitoring system is built from threemethodological components (figure 1) (1) regional-scale time-series image analysis using LandTrendr(Kennedy et al 2010 2015) (2) statistical linkage ofthat imagery with field plot data using the gradi-ent nearest neighbor (GNN) methodology (Ohmannet al 2012) and (3) local-scale assessment using air-borne lidar data (Kane et al 2010)

Methodological uncertainty is explicitly incorpo-rated into the system Rather than providing only asingle solution for biomass estimation at any givenlocation we create a suite of scenarios of the biomasslandscape for each pixel and year yielding both the

2


Figure 1 An observation-based empirical carbon monitoring system From left satellite-data are analyzed through a time-seriesalgorithm (LandTrendr Kennedy et al 2010) to create disturbance maps change agent maps and temporally-stabilized yearly imagetime-series Those images are linked via statistical algorithms (GNN gradient nearest neighbor imputation (Ohmann and Gregory2002)) with plot data (FIA forest inventory analysis) and other geospatial data to create yearly maps of plot-like information Thatinformation includes tree-level species and height information which can be used in allometric equations to create yearly biomassmaps at at 30 m grain size These maps can be used to characterize biomass status in a consistent manner across large or smallownerships or to examine how change in biomass is affected by anthropogenic or natural agents of change Where available theseregional-scale biomass maps can be compared to small-area maps of biomass derived from small-footprint lidar acquisitions

central tendency and an estimate of methodologicaluncertainty We then characterize trends in biomassaccording to the timing and type of agents causingchange including both human and natural processes

22 Study area and datasetsWe prototyped our biomass monitoring system in theWestern Cascades province of Oregon and a small partof California (see section S1 for details) This studyarea contains some of the most carbon-rich conifer-dominated forests in the world (Smithwick et al 2002Waring and Franklin 1979) with a wide range of dis-turbance and recovery dynamics (Kennedy et al 2012Moeur et al 2011)

The monitoring system is built from four broadclasses of data Forest inventory data measured in thefield by US federal agencies provide the foundationalmeasurements of individual trees at a sample of loca-tions across the study area Imagery from the Landsatsatellites provide the spatial and temporal descriptionsof the landscape as it changes over time Ancillarygeospatial data describe the broad environmental con-ditions including soil elevation and climate Finallyairborne lidar (laser range-finding) data are usedopportunistically to assess performance of the systemat focal areas

23 Analysis231 Image analysis with LandTrendrApplying simple statistical tools to spectral data theLandTrendr algorithms distill the dominant distur-bance and recovery processes for each pixel in a Landsat

Thematic Mapper time-series (Kennedy et al 2010)LandTrendr image analysis fills two roles First itprovides stabilized imagery on which the later GNNmapping is based By removing factors that introduceyear-to-year spectral noise this lsquotemporally stabilizedrsquoimagery allows development of a single consistentGNN model across time These stabilized images notonly capture the state of the landscape in a given yearbut also whether that state is changing from one yearto the next or is stable Second data derived fromLandTrendr can be used to label the agents that causeabrupt and subtle changes including clearcuts partialharvests fires and insect attack For both stabilizedimagery and change agents the ultimate analyticalproducts were 30 m resolution raster images for everyyear from 1990ndash2012 While the LandTrendr algo-rithms have been well-described elsewhere (Kennedyet al 2010 Kennedy et al 2012) we highlight in sectionS21 the changes and key steps in processing that applyto the biomass monitoring system as well as the toolsused to evaluate effectiveness

232 Gradient nearest neighbor imputationThe GNN method links forest inventory plot datawith a suite of geospatial predictor variables to pro-duce regional estimates of forest characteristics fromwhich biomass can be calculated Also described else-where (Ohmann and Gregory 2002 Ohmann et al2012 Ohmann et al 2014) the GNN methoduses canonical correspondence analysis (ter Braak1986) to develop a multivariate gradient space inwhich spatial predictor datasets (climate topography

3


00 200 400 600

Observed (Mg ha)800 1000

200

400

Pre

dict

ed (M

g h

a)

600

800 00 200 400

200

4001000Inset r2 = 082 RMSE=84

r2 = 078 RMSE=103 Slope =083 Intercept = 368

Slope =10 Intercept = 149

Figure 2 Estimates of the relationship between aboveground live-tree biomass (CRM allometrics) predicted from the LandTrendr +GNN imputation approach versus that observed at gt 6000 plots across the study area Shown are mean values from GNN Scenarios1 biomass shows little bias up through 500 Mg haminus1 (see inset) but slowly becomes underpredicted as biomass increases beyond thatvalue (main graph)

spectral data) and forest inventory plot observationsare linked In section S22 we highlight innovationsof the method specific to the biomass monitoringsystem presented in this contribution At the end ofthe GNN modeling each pixel on the landscape isassigned tree measurements from an inventory plotfrom which any derivative metrics (including biomass)can be determined and mapped

233 Uncertainty scenariosLandTrendr and GNN are imperfect models affectedby the parameters used to constrain them To capturethe uncertainty in these parameter choices we ran bothLandTrendr andGNNunder threeparameter scenarioseach (labeled LT1 2 and 3 for LandTrendr scenariosand GNN 1 2 and 3 for GNN) resulting in a com-bination of nine lsquouncertainty scenariorsquo predictions foreach pixel and year Details of these scenarios are givenin sections S214 and S224

Additionally the translation from tree measure-ments to biomass through allometric equations is itselfuncertain (Duncanson et al 2015) We used two differ-ent families of allometric scaling to estimate biomassfrom tree measurements in each pixel those based onJenkins et al (2003) and those using the componentratio method (CRM (Heath et al 2008))

234 Lidar-based mappingAirborne lidar data provide insight into the three-dimensional structure of forest stands (Lefsky et al2002) and when combined with reliable concurrentfield measurements can yield highly reliable estimatesof forest biomass at apoint in time (DubayahandDrake2000) For one of the two lidar acquisitions in our studydomain with appropriate field data we calculated esti-mates of aboveground forest biomass For the secondsite we utilized an existing forest biomass map Detailsof analysis are provided in section S23

3 Results

31 Detection and attribution of changeResults of detection and attribution of change aredetailed in section S3 and were accomplished at ratesconsistent with our prior published work (Kennedyet al 2015 Kennedy et al 2012) Confidence in detec-tionand labelingarehighest for abrupt high magnitudeevents such as fires and clearcuts and lower for sub-tle and long-duration events such as insect attack orgeneral decline in vigor (sensu (Cohen et al 2016))

32 Forest biomass estimation with GNNFor the 6108 forest inventory plot measurements in theentire study area agreement between measured andmean imputed biomass (using GNN scenarios 1 and 2(section 224) for all LandTrendr scenarios) was rel-atively unbiased and 11 for biomass densities up to500 Mg haminus1 (figure 2 inset) and then showed signsof slight underprediction at biomass densities greaterthan 500 Mg haminus1 (figure 2 main) To understandhow these relationships would affect estimates of totalbiomass for the whole region we used scaling factorson observed plot values to estimate the proportion ofthe landscape in different biomass categories calcu-lated total regional biomass within those categoriesand compared the total regional biomass to that derivedfrom the nine variants of our biomass maps grouped inthe same biomass categories (figure 3) Mapped totalbiomass agreed well with estimated biomass from sam-pled plots both in terms of totals and in terms of thedistribution of biomass across biomass categories butunder-represented the very highest biomass categories(greater than 830 Mg haminus1) Underrepresentation ofhigh biomass conditions was accentuated for GNNScenario 3 runs Although it is impossible to knowbiomass of unsampled plots the conservative approachof assigning biomass in proportion to all sampled plotssuggests that our mapped values underpredict biomass

4


0-21

0 M

g h

a21

0-41

041

0-62

062

0-83

0

830-1500

Unsampled

00

Plot

Esti

mat

e

LT1

GNN

1LT

1 G

NN2

LT1

GNN

3LT

2 G

NN1

LT2

GNN

2LT

2 G

NN3

LT3

GNN

1LT

3 G

NN2

LT3

GNN

3

02

04

06

Bio

mas

s (P

g)

08

10

Figure 3 Total aboveground live biomass (using CRM allometry) in the study region estimated both from plot data expansionfactors (lsquoPlot Estimatersquo) and from nine biomass maps derived using the biomass monitoring system described herein Colors representthe proportion of the total estimated biomass contributed by forests in different biomass densities shown next to the Plot Estimatebar Note that all biomass maps underestimate the proportion of biomass in forests above biomass density of 830 Mg haminus1 withGNN 3 scenarios showing the greatest underestimation Unsampled plots are those not acceesible they have unknown biomass As aconservative estimate we assume they have the same distribution of biomass values as the rest of the measured dataset

across all scenarios Regardless of LandTrendr orGNN-scenario Jenkins-based biomass was sim16higher than CRM-based biomass (figure 3 vsfigure S5) consistent with findings of others (Zhouand Hemstrom 2009) The standard deviation amongLandTrendr amp GNN scenarios however was ltlt 1of the mean for either CRM or Jenkins based mapsWhen reporting the remainder of our results wereport only on the CRM-based estimates as this isthe approach currently employed by FIA to estimatenation-wide biomass stocks

Imputed maps of aboveground live forest biomassshowed spatial patterns consistent with expecta-tions for the managed landscapes of this studyarea (figure 4) For example for the CRM-basedbiomass estimation using LandTrendr scenario 1(LT1) and GNN scenario 1 (GNN 1) figures4(a) and (b) old-growth conifer forest shows highlive tree biomass (gt600 Mg haminus1) while recently-disturbed areas show low biomass (0minus50 Mg haminus1)Taking the mean value of biomass from all ninescenarios retains the same landscape-scale pattern(figure 4(c)) but reduces the pixel-scale spatial vari-ability resulting from the nearest-neighbor approachA map of the variability of biomass across scenar-ios shows considerable spatial noise (figure 4(d)) butretains some coherent spatial patterns Generally vari-ability peaks in the middle of the range of values (figure4(e)) but for the majority of the study area vari-ability in estimates is below 25 of the mean values

(figure 4(e) inset) Because all LandTrendr and GNNscenarios are methodologically defensible this suggeststhat no single scenario could be used to represent theothers Rather a user should be aware of both a centraltendency (such as the mean value figure 4(c)) and thestandard deviation image (figure 4(d)) consistent withBell et al (2015)

33 Comparison with lidar-derived estimatesWe examined the relationship between biomass esti-mated from lidar and from LandTrendr +GNN at twoforest sites At the 30 m pixel scale the relationship wasnoisy but it improved with spatial aggregation (sectionS5figureS6) For theHJA the relationship trackedwellup until approximately 450 Mg haminus1 and then abruptlyplateaued (figure S6) while biomass values remainedless than 450 Mg haminus1 at the Deschutes site Maps ofdisagreement confirm that discrepancies were greatestin high-biomass stands at the HJ Andrews

34 Change in biomassBy creating yearly biomass maps from 1990ndash2012 themonitoring system allows comparison at a range oftemporal aggregation scales from the whole periodto yearly At any given point or for any given patchor area the biomass trajectory over the entire periodcan be tracked showing the impacts on biomass ofboth disturbance and regrowth (figure 5) and how dif-ferent modeling scenarios affect estimates Notably itis unlikely that any single modeling scenario is best

5


Figure 4 Example biomass products from the monitoring system (a) Biomass density (Mg biomass haminus1) for a single scenario (LT 1GNN 1 GRM) of the system for the year 2005 (b) Detail of (a) as indiciated by black box and arrow (c) Mean biomass density acrossall nine scenarios for the CRM allometric family for the same year (d) Standard deviation of the nine scenario estimates (e) Densityplot of uncertainty as represented by standard deviation of biomass vs mean biomass for the entire area shown in (a) Althoughuncertainty exceeded mean biomass in some cases more than 50 and 75 of the landscape had uncertainty ratios less than 25and 50 respectively (inset)

a)

b)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 20100

100

200

300

400

500

600

700

Bio

mas

s (M

g h

a)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010240

260

280

300

320

340

360

380

Bio

mas

s (M

g h

a)

1991 1993 2010

300m

600++

0

Mg

biom

ass

ha

G1 G2 G3L1L2L3

c)

Year Year

Figure 5 Tracking change in biomass (a) Maps of mean per-pixel biomass across all nine LandTrendr and GNN scenario combinations(see table S1 for LandTrendr scenarios section S224 for GNN scenarios) for selected years (b) Trajectory of biomass for the clearcutindicated by the lower polygon in (a) Charted values are mean across pixels within the clearcut patch for each combination ofLandTrendr and GNN scenario (L1 = LandTrendr scenario 1 etc G1 = GNN scenario 1 etc) Note the variability in pre-biomassamong GNN and LT scenarios and the variation in post-disturbance regrowth patterns (c) As with (b) except for the upper polygonwhich illustrates regrowth from a clearcut that occured before the observation period Note that the y-axis scale is changed to highlightregrowth variability among LT and GNN scenarios

6


8

6

4

2

0

ndash2

ndash4

ndash6

ndash8

1990 1995 2000 2005 2010Year

Bio

mas

s ch

ange

(Tg)

AgentStableRegrowthClearcutPartial HarvestFireInsectDiseaseOtherUnknown AbruptUnknown Slow

Net Change

Figure 6 Change in biomass caused by specific agents For each of the nine scenarios static biomass (based on CRM) in a year wasdifferenced from the prior year and the mean of all changes summarized at the pixel scale by change agent and summarized across theentire study area When abrupt disturbances occurred in patches less thansim1 ha we labeled them as lsquounknown abruptrsquo Unknown slowchanges corresponded with long slow decline in vegetative signal that could not be ascribed to observable insect or disease outbreaksbut are typically resultant from a mixture of lsquodeclinersquo factors (Cohen et al 2016)

under all conditions as illustrated by the variabilityin representation of both pre-disturbance conditionand post-disturbance recovery rate among scenarios(figures 5(b) and (c))

For a monitoring system it is important to doc-ument not just how much biomass is changing overtime but why By linking maps of change agent withmaps of mean biomass change we can ascribe biomassloss and gain to specific agents over time (figure 6) Forthis study area we found that anthropogenic changes(clearcuts and partial harvests) were a consistent losssignal over the entire time period with occasionallarge loss from fire Net change recorded by our sys-tem switched from general uptake in the 1990s toloss in the 2000s

Because our system produces multiple estimates ofbiomass change for each time period it allows assess-ment of the uncertainty of the impact of each agent(figure 7) Uncertainty in biomass loss due to clearcutsand partial harvests is low relative to the mean esti-mated loss andgenerally consistent across time (figures7(a) and (b)) Uncertainty in loss due to fire scaleswith overall loss rate which itself is quite variablefrom year to year (figure 7(c)) Both of these find-ings emphasize that estimates of biomass immediatelyafter abrupt disturbances can be relatively uncertainmaking the estimates of loss span a wider rangethan the pre-disturbance biomass Uncertainties asso-ciated with long slow disturbance and with regrowth(figures 7(e) and (f)) were greater than for the otheragents but were relatively consistent over time Evenconsidering the uncertainty in estimates loss fromlong slow disturbance which is often associated withpests and other stressors (Cohen et al 2016) exceedsloss from the two insect agents that were explicitly

modeled (bark beetle and western spruce budwormMeigs et al 2015)

4 Discussion

Forests occupy an important component of any carbonmonitoring program They have the capacity to storesubstantial amounts of carbon in live abovegroundbiomass but they are also subject to disturbances thatremove it For carbon accounting and risk analysis acarbon monitoring system must track forest biomassconsistently across space and time and must trackhow natural and anthropogenic processes change itBelow we highlight strategies strengths and weak-nesses and findings of our system Substantially moredetail can be found in supplemental section S6 availableat stacksioporgERL13025004mmedia

Our monitoring strategy used advanced algorith-mic approaches to leverage strengthsof complementarydatasets Spectral information from satellite data areubiquitous and can track change over time but sat-urate in high biomass systems Plot data are the goldstandard in biomass estimation but are sparse in spacein time In our system temporal segmentation pro-vided the stability of optical data needed to build largeplot databases and the nearest neighbor imputationapproach allowed high biomass data to be included inmapped outputs Although our estimates of biomassshow signs of saturation they did so at relativelyhigh densities (450 or 500 Mg haminus1 figures 2 and3) suggesting utility in many forested systems Spa-tial scale matters however At pixel scales noise vastlyswamps signal (figure S6) meaning our systemrsquos esti-mates of absolute biomass at one point in time are not

7


a) c)b)

d) e) f)

Δ Tg

Bio

mas

sΔ

Tg B

iom

ass

Loss

GainLoss

Gain

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

0-2

2468

10

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

Figure 7 Mean biomass loss or gain attributable to individual agents of change (a) clearcuts (b) partial harvest (c) fire (d) knowninsect (e) unknown slow disturbance and (f) recovery Black line indicates mean change in biomass across all nine LT and GNNscenarios for all cases of a given agent per year and grey shading shows variability in biomass estimation across the nine modelingscenarios

appropriate at the pixel scale ongoing work exploresthis scale effect (Bell et al in preparation) Howevertracking relative change in biomass over time appearspossible even at the scale of individual forest stands(figure 5)

Because our system relied on temporal segmenta-tion of Landsat imagery it was possible to explicitly tiechange in biomass to disturbance and growth changesWhile thedetectionof disturbance fromremote sensingcontained error (tables S2 and 3) such tight couplingallowed comparison of the carbon impacts of differentchange agents (figure 6) Importantly labeling agentsof change depended heavily on scaling local knowledgeof disturbance through machine learning algorithms

Ourensembleapproach toestimatinguncertainty isimperfect and incomplete but provides a first approx-imation of the spatial and temporal pattern of theunknown Not strictly error our uncertainty mapsessentially point to places and times where the systembehaved poorly This can provide insight into whereand how the information content or the assumptionsof the empirical system break down

Althoughour results in ahighbiomass forest systemare encouraging some caution is needed when consid-ering application of our monitoring system in otherforest types The availability and quality of our coredatasets may represent a best-case scenario comparedtoother areasHistorical Landsat data in the continentalUS are temporally dense and the climate of our studyregion allowed high probability of obtaining cloud-freeobservations Similarly the FIA plot network is unusu-ally robust inqualitydesign and temporaldepthTheseconditions would not be easily met in many tropicalsystems for example where cloud cover and image

availability limit spectral information and access andcost constraints make direct measurement challengingAdditionally the strength of the relationships betweenspectral data and forest type data may differ in foresttypes unlike those tested here which were dominatedby evergreen conifer types In sparse forests shrub-lands or broadleaf deciduous systems robustness ofprediction would need to be tested explicitly beforeimplementing our system

Nevertheless by providing a consistent monitor-ing framework across which different agents of changecan be compared our system yields several compellinginsights in our forest-dominated study area Humanactivities control the biomass removal in most yearsFire is sporadic but when it occurs shifts the systemstrongly into carbon loss Note that while our systemsimply measures removal from the landscape and doesnot track the fate of that biomass (or carbon) the tightcoupling of agent and biomass loss could be used toestimate differential fates of carbon in fires harvestsand other processes Interestingly loss of carbon fromknown insects was not as great as loss due to otherslow degradation processes The challenge in inter-preting this finding is the non-specific nature of thoseprocesses broadly described as lsquodeclinersquo (Cohen et al2016) they appear to be real phenomena but are amix of other insects disease drought effects and otherstressors Our own estimates show substantial uncer-tainty in their biomass impacts leading us to cautionagainst over-interpretationof their importance withoutfurther study

Overall our system appears viable as another toolfor managers and modelers of forest carbon It pro-vides consistent spatial and temporal estimates of forest

8


biomass dynamics and attempts transparent inclusionof errors and uncertainties By linking agents of changewith biomass change it provides tools to managers andmodelers to understand impacts and evaluate possiblealternative future scenarios

5 Conclusion

We have presented an empirical approach to monitor-ing live forest carbon at broad scales using a mix of fielddata statistical modeling remotely-sensed time-seriesimagery and small-footprint lidardataThe frameworkbuilds maps of forest biomass state for each year thatare consistent and comparable across years and is con-structed to allow direct coupling of change in biomassto specific change agents Several important findingsemerge from this work

1 A common image framework for biomass estima-tion and for change detection allows for consistentcomparison of both state and change processes con-trolling carbon dynamics

2 Although biomass estimation with optical sensorstends to saturate using traditional statistical meth-ods the nearest-neighbor imputation approachallows us to better capture the full range of val-ues (including very high biomass) at the ecoregionscale

3 Comparison with estimates from lidar data showsgeneral agreement but also shows local-scale dis-agreement at the scale of pixels and forest standsThis suggests that there is a lower limit on the spa-tial detail at which this regional scale method canbe applied Further investigation to determine theappropriate scales is underway

4 Variability in controlling parameters for both thetime-series fitting and the imputation approachdo result in variability in biomass estimates at thepixel scale but using the mean of these ensemblesimproves the overall fit between observed and pre-dictedvalues Importantly the variability inbiomassestimates caused by methodological differences ismarkedly less than the different in biomass esti-mates caused by use of different allometric scalingequations

5 In this forest-dominated study area growth and lossprocesses largely balance in most years with lossprocesses dominated by human removal throughharvest In years with substantial fire activityhowever overall biomass loss greatly outpacesgrowth

Acknowledgments

This work was funded through USDA NIFA Grant2011-67003-20458 and NASA Carbon MonitoringSystem Grant NNX12AP76G

ORCID iDs

Robert E Kennedy httpsorcidorg0000-0002-5507-474X

References

Bell D M Gregory M J Kane V R Kane J T Kennedy R E RobertsH Simpson M and Yang Z Multiscale divergence betweenLandsat- and lidar-based biomass mapping explained byregional variation in canopy cover and composition RemoteSensEnviron in preparation

Bell D M Gregory M J and Ohmann J L 2015 Imputed foreststructure uncertainty varies across elevational and longitudinalgradients in the Western Cascade mountains Oregon USAForest Ecol Manage 358 154ndash64

Birdsey R A 2006 Carbon accounting rules and guidelinesfor the United States forest sector J Environ Qual 351518ndash24

Birdsey R A et al 2007 North American forests The First State of theCarbon Cycle Report (SOCCR) The North American CarbonBudget and Implications for the Global Carbon Cycle ed A WKing L Dilling G P Zimmerman D M Fairman R AHoughton G Marland A Z Rose and T J Wilbanks (AshevilleNC National Oceanic and Atmospheric AdministrationNational Climatic Data Center) pp 117ndash26

Blackard J A et al 2008 Mapping US forest biomass usingnationwide forest inventory data and moderateresolution information Remote Sens Environ 1121658ndash77

Chambers J Q Fisher J I Zeng H Chapman E L Baker D B andHurtt G C 2007 Hurricane Katrinarsquos carbon footprint on USGulf Coast forests Science 318 1107

Cohen W B Spies T Alig R J Oetter D R Maiersperger T K andFiorella M 2002 Characterizing 23 years 1972ndash95 of standreplacement disturbance in western Oregon forests withLandsat imagery Ecosystems 5 122ndash37

Cohen W B Yang Z Q Stehman S V Schroeder T A Bell D MMasek J G Huang C Q and Meigs G W 2016 Forestdisturbance across the conterminous United States from1985ndash2012 the emerging dominance of forest decline ForestEcol Manage 360 242ndash52

Dubayah R O and Drake J B 2000 Lidar remote sensing for forestryJ Forestry 98 44ndash6

Duncanson L I Dubayah R O and Enquist B J 2015 Assessing thegeneral patterns of forest structure quantifying tree and forestallometric scaling relationships in the United States Glob EcolBiogeogr 24 1465ndash75

Environment-Canada 2006 National Inventory Report 1990ndash2004Greenhouse Gas Sources and Sinks in Canada (OttowaOntario Greenhouse Gas Division Environment Canada)

EPA 2016 Inventory of US Greenhouse Gas Emissions and Sinks1990minus2014 (Washington DC EP Agency)

Goetz S J Baccini A Laporte N T Johns T Walker W KellndorferJ Houghton R A and Sun M 2009 Mapping and monitoringcarbon stocks with satellite observations a comparison ofmethods Carbon Balance Manage 4

Goodale C L et al 2002 Forest carbon sinks in the NorthernHemisphere Ecol Appl 12 891ndash9

Hansen M C and Loveland T R 2012 A review of large areamonitoring of land cover change using Landsat data RemoteSens Environ 122 66ndash74

Heath L S Hansen M H Smith J E Smith W B and Miles P D 2008Investigation into calculating tree biomass and carbon in theFIADB using a biomass expansion factor approach ForestInventory and Analysis (FIA) Symposium ed W McWilliamsG Moisen and R Czaplewski (Park City UT US Departmentof Agriculture Forest Service Rocky Mountain ResearchStation)

Houghton R A 2005 Aboveground forest biomass and the globalcarbon balance Glob Change Biol 11 945ndash58

9


Huang C Goward S N Masek J G Thomas N Zhu Z andVogelmann J E 2010 An automated approach forreconstructing recent forest disturbance history using denseLandsat time series stacks Remote Sens Environ 114 183ndash98

Hudak A T Lefsky M A Cohen W B and Berterretche M 2002Integration of lidar and Landsat ETM+ data for estimatingand mapping forest canopy height Remote Sens Environ 82397ndash416

Jenkins J C Chojnacky D C Heath L S and Birdsey R A 2003National-scale biomass estimators for United States treespecies Forest Sci 49 12ndash35

Joyce L A Blate G M McNulty S G Millar C I Moser S Neilson R Pand Peterson D L 2009 Managing for multiple resources underclimate change national forests Environ Manage 44 1022ndash32

Kane V R McGaughey R J Bakker J D Gersonde R F Lutz J A andFranklin J F 2010 Comparisons between field- andLiDAR-based measures of stand structural complexity Can JForest Res 40 761ndash73

Kellndorfer J Walker W LaPoint L Bishop J Cormier T Fiske G JKirsch K and Westfall J 2000 NACP Aboveground Biomassand Carbon Baseline Data (NBCD 2000) Dataset at Oak RidgeNational Laboratory Distributed Active Archive Center(ORNL-DAAC) (httpsdoiorg103334ORNLDAAC1161)

Kennedy R E Yang Z Braaten J Thompson C Antonova N JordanC and Nelson P 2015 Attribution of disturbance change agentfrom Landsat time-series in support of habitat monitoring inthe Puget Sound region USA Remote Sens Environ 166271ndash85

Kennedy R E Yang Z and Cohen W B 2010 Detecting trends inforest disturbance and recovery using yearly Landsat timeseries 1 LandTrendrmdashTemporal segmentation algorithmsRemote Sens Environ 114 2897ndash910

Kennedy R E Zhiqiang Y Cohen W Pfaff E Braaten J and NelsonP 2012 Spatial and temporal patterns of forest disturbance andregrowth within the area of the Northwest Forest Plan RemoteSens Environ 122 117ndash33

Kim Y Yang Z Cohen W B Pflugmacher D Lauver C L andVankat J L 2009 Distinguishing between live and deadstanding tree biomass on the North Rim of Grand CanyonNational Park USA using small-footprint lidar data RemoteSens Environ 113 2499ndash510

Lefsky M A Cohen W B Parker G and Harding D 2002 Lidarremote sensing for ecosystem studies BioScience 52 19ndash30

Masek J G Goward S N Kennedy R E Cohen W B Moisen G GSchleeweis K and Huang C 2013 United States forestdisturbance trends observed using Landsat time seriesEcosystems (httpsdoiorg101007s10021-013-9669-9)

McDowell N G et al 2015 Global satellite monitoring of climate-induced vegetation disturbances Trends Plant Sci 20 114ndash23

McRoberts R E Liknes G C and Domke G M 2014 Using a remotesensing-based percent tree cover map to enhance forestinventory estimation Forest Ecol Manage 331 12ndash8

Meigs G Kennedy R E Gray A N and Gregory M J 2015Spatiotemporal dynamics of recent mountain pine beetle andwestern spruce budworm outbreaks across the PacificNorthwest Region USA Forest Ecol Manage 339 71ndash86

Moeur M Ohmann J L Kennedy R E Cohen W B Gregory M JYang Z Roberts H Spies T A and Fiorella M 2011 NorthwestForest Planndashthe first 15 years 1994minus2008 status and trends oflate-successional and old-growth forests General TechnicalReport (Portland OR)

Naesset E 2002 Predicting forest stand characteristics with airbornescanning laser using a practical two-stage procedure and fielddata Remote Sens Environ 80 88ndash99

Ohmann J L and Gregory M 2002 Predictive mapping of forestcomposition and structure with direct gradient analysis andnearest-neighbor imputation in coastal Oregon USA Can JForest Res 32 724ndash41

Ohmann J L Gregory M J Roberts H M Cohen W B Kennedy R Eand Yang Z 2012 Mapping change of older forest withnearest-neighbor imputation and Landsat time-series ForestEcol Manage 272 13ndash25

Ohmann J L Gregory M J and Roberts H M 2014 Scaleconsiderations for integrating forestry inventory plot data andsatellite data for regional forest mapping Remote SensEnviron 151 3ndash15

Pflugmacher D Cohen W B Kennedy R E and Yang Z Q 2014Using Landsat-derived disturbance and recovery history andlidar to map forest biomass dynamics Remote Sens Environ151 124ndash37

Saatchi S S et al 2011 Benchmark map of forest carbon stocks intropical regions across three continents Proc Natl Acad SciUSA 108 9899ndash904

Sannier C McRoberts R E and Fichet L V 2016 Suitability of globalforest change data to report forest cover estimates at nationallevel in gabon Remote Sens Environ 173 326ndash38

Sexton J O Noojipady P Anand A Song X P McMahon S HuangC Q Feng M Channan S and Townshend J R 2015 A modelfor the propagation of uncertainty from continuous estimatesof tree cover to categorical forest cover and change RemoteSens Environ 156 418ndash25

Smithwick E A H Harmon M E Remillard S M Acker S A andFranklin J F 2002 Potential upper bounds of carbonstores in forests of the Pacific Northwest Ecol Appl 121303ndash17

Swatantran A Dubayah R Roberts D Hofton M and Blair J B 2011Mapping biomass and stress in the Sierra Nevada using lidarand hyperspectral data fusion Remote Sens Environ 1152917ndash30

ter Braak C J F 1986 Canonical correspondence analysis a neweigenvecctor technique for multivariate direct gradient analysisEcology 67 1167ndash79

Waring R H and Franklin J F 1979 Evergreen coniferous forests ofthe pacific northwest Science 204 1380ndash6

Williams C A Gu H MacLean R Masek J G and Collatz G J 2016Disturbance and the carbon balance of US forests aquantitative review of impacts from harvests fires insects anddroughts Glob Planet Change 143 66ndash80

Wilson B T Woodall C W and Griffith D M 2013 Imputing forestcarbon stock estimates from inventory plots to a nationallycontinuous coverage Carbon Balance Manage 8 1

Woodbury P B Smith J E and Heath L S 2007 Carbonsequestration in the US forest sector from 1990ndash2010 ForestEcol Manage 241 14ndash27

Zheng D L Rademacher J Chen J Q Crow T Bresee M le Moine Jand Ryu S R 2004 Estimating aboveground biomass usingLandsat 7 ETM+ data across a managed landscape in northernWisconsin USA Remote Sens Environ 93 402ndash11

Zhou X and Hemstrom M A 2009 Estimating Aboveground TreeBiomass on Forest Land in the Pacific Northwest AComparison of Approaches (Portland OR US Department ofAgriculture Forest Service) p 18

10

Environ Res Lett 13 (2018) 025004 httpsdoiorg1010881748-9326aa9d9e

LETTER

An empirical integrated forest biomass monitoringsystem

Robert E Kennedy119 Janet Ohmann2 Matt Gregory3 Heather Roberts4 Zhiqiang Yang5 David M Bell6Van Kane7 M Joseph Hughes8 Warren B Cohen9 Scott Powell10 Neeti Neeti11 Tara Larrue12 SamHooper13 Jonathan Kane14 David L Miller15 James Perkins16 Justin Braaten17 and Rupert Seidl18

1 College of Earth Ocean and Atmospheric Sciences Oregon State University United States of America2 Emeritus Scientist Pacific Northwest Forestry Sciences Lab USDA Forest Service United States of America3 College of Forestry Oregon State University United States of America4 College of Forestry Oregon State University United States of America5 College of Forestry Oregon State University United States of America6 USDA Forest Service Pacific Northwest Research Station United States of America7 School of Environmental and Forest Sciences University of Washington United States of America8 College of Earth Ocean and Atmospheric Sciences Oregon State University United States of America9 USDA Forest Service Pacific Northwest Forestry Sciences Lab United States of America10 Department of Land Resources and Environmental Sciences Montana State University United States of America11 Department of Geography Boston University Now at Department of Natural Resources TERI University India12 College of Earth Ocean and Atmospheric Sciences Oregon State University United States of America13 College of Earth Ocean and Atmospheric Sciences Oregon State University United States of America14 School of Environmental and Forest Sciences University of Washington United States of America15 Department of Geography Boston University Now at Department of Geography University of California Santa Barbara United States

of America16 Department of Geography Boston University Now at University of New Hampshire United States of America17 College of Earth Ocean and Atmospheric Sciences Oregon State University United States of America18 Institute of Silviculture Universty of Natural Resources and Life Sciences Vienna Austria19 Author to whom any correspondence should be addressed

OPEN ACCESS

RECEIVED

18 July 2017

REVISED

17 November 2017

ACCEPTED FOR PUBLICATION

28 November 2017

PUBLISHED

1 February 2018

Original content fromthis work may be usedunder the terms of theCreative CommonsAttribution 30 licence

Any further distributionof this work mustmaintain attribution tothe author(s) and thetitle of the work journalcitation and DOI

E-mail rkennedycoasoregonstateedu

Keywords forest biomass Landsat forest inventory lidar monitoring disturbance

Supplementary material for this article is available online

AbstractThe fate of live forest biomass is largely controlled by growth and disturbance processes both naturaland anthropogenic Thus biomass monitoring strategies must characterize both the biomass of theforests at a given point in time and the dynamic processes that change it Here we describe and test anempirical monitoring system designed to meet those needs Our system uses a mix of field datastatistical modeling remotely-sensed time-series imagery and small-footprint lidar data to build andevaluate maps of forest biomass It ascribes biomass change to specific change agents and attempts tocapture the impact of uncertainty in methodology We find that

bull A common image framework for biomass estimation and for change detection allows for consistentcomparison of both state and change processes controlling biomass dynamics

bull Regional estimates of total biomass agree well with those from plot data alone

bull The system tracks biomass densities up to 450ndash500 Mg haminus1 with little bias but begins underesti-mating true biomass as densities increase further

bull Scale considerations are important Estimates at the 30 m grain size are noisy but agreement atbroad scales is good Further investigation to determine the appropriate scales is underway

bull Uncertainty from methodological choices is evident but much smaller than uncertainty based onchoice of allometric equation used to estimate biomass from tree data

bull In this forest-dominated study area growth and loss processes largely balance in most years withloss processes dominated by human removal through harvest In years with substantial fire activityhowever overall biomass loss greatly outpaces growth

Taken together our methods represent a unique combination of elements foundational to anoperational landscape-scale forest biomass monitoring program

copy 2018 The Author(s) Published by IOP Publishing Ltd


1 Introduction







2 Methods



2









3


00 200 400 600


200

400

Pre

dict

ed (M

g h

a)

600

800 00 200 400

200

4001000Inset r2 = 082 RMSE=84








3 Results



4


0-21

0 M

g h

a21

0-41

041

0-62

062

0-83

0

830-1500

Unsampled

00

Plot

Esti

mat

e

LT1

GNN

1LT

1 G

NN2

LT1

GNN

3LT

2 G

NN1

LT2

GNN

2LT

2 G

NN3

LT3

GNN

1LT

3 G

NN2

LT3

GNN

3

02

04

06

Bio

mas

s (P

g)

08

10







5



a)

b)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 20100

100

200

300

400

500

600

700

Bio

mas

s (M

g h

a)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010240

260

280

300

320

340

360

380

Bio

mas

s (M

g h

a)

1991 1993 2010

300m

600++

0

Mg

biom

ass

ha

G1 G2 G3L1L2L3

c)

Year Year


6


8

6

4

2

0

ndash2

ndash4

ndash6

ndash8

1990 1995 2000 2005 2010Year

Bio

mas

s ch

ange

(Tg)


Net Change






4 Discussion



7


a) c)b)

d) e) f)

Δ Tg

Bio

mas

sΔ

Tg B

iom

ass

Loss

GainLoss

Gain

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

0-2

2468

10

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010









8



5 Conclusion







Acknowledgments


ORCID iDs


References


















9



































10


1 Introduction







2 Methods



2









3


00 200 400 600


200

400

Pre

dict

ed (M

g h

a)

600

800 00 200 400

200

4001000Inset r2 = 082 RMSE=84








3 Results



4


0-21

0 M

g h

a21

0-41

041

0-62

062

0-83

0

830-1500

Unsampled

00

Plot

Esti

mat

e

LT1

GNN

1LT

1 G

NN2

LT1

GNN

3LT

2 G

NN1

LT2

GNN

2LT

2 G

NN3

LT3

GNN

1LT

3 G

NN2

LT3

GNN

3

02

04

06

Bio

mas

s (P

g)

08

10







5



a)

b)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 20100

100

200

300

400

500

600

700

Bio

mas

s (M

g h

a)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010240

260

280

300

320

340

360

380

Bio

mas

s (M

g h

a)

1991 1993 2010

300m

600++

0

Mg

biom

ass

ha

G1 G2 G3L1L2L3

c)

Year Year


6


8

6

4

2

0

ndash2

ndash4

ndash6

ndash8

1990 1995 2000 2005 2010Year

Bio

mas

s ch

ange

(Tg)


Net Change






4 Discussion



7


a) c)b)

d) e) f)

Δ Tg

Bio

mas

sΔ

Tg B

iom

ass

Loss

GainLoss

Gain

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

0-2

2468

10

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010









8



5 Conclusion







Acknowledgments


ORCID iDs


References


















9



































10









3


00 200 400 600


200

400

Pre

dict

ed (M

g h

a)

600

800 00 200 400

200

4001000Inset r2 = 082 RMSE=84








3 Results



4


0-21

0 M

g h

a21

0-41

041

0-62

062

0-83

0

830-1500

Unsampled

00

Plot

Esti

mat

e

LT1

GNN

1LT

1 G

NN2

LT1

GNN

3LT

2 G

NN1

LT2

GNN

2LT

2 G

NN3

LT3

GNN

1LT

3 G

NN2

LT3

GNN

3

02

04

06

Bio

mas

s (P

g)

08

10







5



a)

b)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 20100

100

200

300

400

500

600

700

Bio

mas

s (M

g h

a)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010240

260

280

300

320

340

360

380

Bio

mas

s (M

g h

a)

1991 1993 2010

300m

600++

0

Mg

biom

ass

ha

G1 G2 G3L1L2L3

c)

Year Year


6


8

6

4

2

0

ndash2

ndash4

ndash6

ndash8

1990 1995 2000 2005 2010Year

Bio

mas

s ch

ange

(Tg)


Net Change






4 Discussion



7


a) c)b)

d) e) f)

Δ Tg

Bio

mas

sΔ

Tg B

iom

ass

Loss

GainLoss

Gain

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

0-2

2468

10

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010









8



5 Conclusion







Acknowledgments


ORCID iDs


References


















9



































10


00 200 400 600


200

400

Pre

dict

ed (M

g h

a)

600

800 00 200 400

200

4001000Inset r2 = 082 RMSE=84








3 Results



4


0-21

0 M

g h

a21

0-41

041

0-62

062

0-83

0

830-1500

Unsampled

00

Plot

Esti

mat

e

LT1

GNN

1LT

1 G

NN2

LT1

GNN

3LT

2 G

NN1

LT2

GNN

2LT

2 G

NN3

LT3

GNN

1LT

3 G

NN2

LT3

GNN

3

02

04

06

Bio

mas

s (P

g)

08

10







5



a)

b)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 20100

100

200

300

400

500

600

700

Bio

mas

s (M

g h

a)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010240

260

280

300

320

340

360

380

Bio

mas

s (M

g h

a)

1991 1993 2010

300m

600++

0

Mg

biom

ass

ha

G1 G2 G3L1L2L3

c)

Year Year


6


8

6

4

2

0

ndash2

ndash4

ndash6

ndash8

1990 1995 2000 2005 2010Year

Bio

mas

s ch

ange

(Tg)


Net Change






4 Discussion



7


a) c)b)

d) e) f)

Δ Tg

Bio

mas

sΔ

Tg B

iom

ass

Loss

GainLoss

Gain

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

0-2

2468

10

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010









8



5 Conclusion







Acknowledgments


ORCID iDs


References


















9



































10


0-21

0 M

g h

a21

0-41

041

0-62

062

0-83

0

830-1500

Unsampled

00

Plot

Esti

mat

e

LT1

GNN

1LT

1 G

NN2

LT1

GNN

3LT

2 G

NN1

LT2

GNN

2LT

2 G

NN3

LT3

GNN

1LT

3 G

NN2

LT3

GNN

3

02

04

06

Bio

mas

s (P

g)

08

10







5



a)

b)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 20100

100

200

300

400

500

600

700

Bio

mas

s (M

g h

a)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010240

260

280

300

320

340

360

380

Bio

mas

s (M

g h

a)

1991 1993 2010

300m

600++

0

Mg

biom

ass

ha

G1 G2 G3L1L2L3

c)

Year Year


6


8

6

4

2

0

ndash2

ndash4

ndash6

ndash8

1990 1995 2000 2005 2010Year

Bio

mas

s ch

ange

(Tg)


Net Change






4 Discussion



7


a) c)b)

d) e) f)

Δ Tg

Bio

mas

sΔ

Tg B

iom

ass

Loss

GainLoss

Gain

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

0-2

2468

10

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010









8



5 Conclusion







Acknowledgments


ORCID iDs


References


















9



































10



a)

b)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 20100

100

200

300

400

500

600

700

Bio

mas

s (M

g h

a)

1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010240

260

280

300

320

340

360

380

Bio

mas

s (M

g h

a)

1991 1993 2010

300m

600++

0

Mg

biom

ass

ha

G1 G2 G3L1L2L3

c)

Year Year


6


8

6

4

2

0

ndash2

ndash4

ndash6

ndash8

1990 1995 2000 2005 2010Year

Bio

mas

s ch

ange

(Tg)


Net Change






4 Discussion



7


a) c)b)

d) e) f)

Δ Tg

Bio

mas

sΔ

Tg B

iom

ass

Loss

GainLoss

Gain

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

0-2

2468

10

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010









8



5 Conclusion







Acknowledgments


ORCID iDs


References


















9



































10


8

6

4

2

0

ndash2

ndash4

ndash6

ndash8

1990 1995 2000 2005 2010Year

Bio

mas

s ch

ange

(Tg)


Net Change






4 Discussion



7


a) c)b)

d) e) f)

Δ Tg

Bio

mas

sΔ

Tg B

iom

ass

Loss

GainLoss

Gain

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

0-2

2468

10

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010









8



5 Conclusion







Acknowledgments


ORCID iDs


References


















9



































10


a) c)b)

d) e) f)

Δ Tg

Bio

mas

sΔ

Tg B

iom

ass

Loss

GainLoss

Gain

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010

0-2

2468

10

0-1-2-3-4

1990 1995 2000 2005 2010 1990 1995 2000 2005 2010 1990 1995 2000 2005 2010









8



5 Conclusion







Acknowledgments


ORCID iDs


References


















9



































10



5 Conclusion







Acknowledgments


ORCID iDs


References


















9



































10



































10

An empirical, integrated forest biomass monitoring system · turbance and recovery dynamics...

Documents

Transcript of An empirical, integrated forest biomass monitoring system · turbance and recovery dynamics...