the challenges of using satellite data sets to assess...

Full Terms & Conditions of access and use can be found athttps://www.tandfonline.com/action/journalInformation?journalCode=tcmt20

Carbon Management

ISSN: 1758-3004 (Print) 1758-3012 (Online) Journal homepage: https://www.tandfonline.com/loi/tcmt20

The challenges of using satellite data sets toassess historical land use change and associatedgreenhouse gas emissions: a case study of threeIndonesian provinces

Sybrand van Beijma, Julia Chatterton, Susan Page, Chris Rawlings, RichardTiffin & Henry King

To cite this article: Sybrand van Beijma, Julia Chatterton, Susan Page, Chris Rawlings, RichardTiffin & Henry King (2018) The challenges of using satellite data sets to assess historical land usechange and associated greenhouse gas emissions: a case study of three Indonesian provinces,Carbon Management, 9:4, 399-413, DOI: 10.1080/17583004.2018.1511383

To link to this article: https://doi.org/10.1080/17583004.2018.1511383

View supplementary material

Published online: 20 Dec 2018.

Submit your article to this journal

Article views: 77

View Crossmark data

https://www.tandfonline.com/action/journalInformation?journalCode=tcmt20https://www.tandfonline.com/loi/tcmt20https://www.tandfonline.com/action/showCitFormats?doi=10.1080/17583004.2018.1511383https://doi.org/10.1080/17583004.2018.1511383https://www.tandfonline.com/doi/suppl/10.1080/17583004.2018.1511383https://www.tandfonline.com/doi/suppl/10.1080/17583004.2018.1511383https://www.tandfonline.com/action/authorSubmission?journalCode=tcmt20&show=instructionshttps://www.tandfonline.com/action/authorSubmission?journalCode=tcmt20&show=instructionshttp://crossmark.crossref.org/dialog/?doi=10.1080/17583004.2018.1511383&domain=pdf&date_stamp=2018-12-20http://crossmark.crossref.org/dialog/?doi=10.1080/17583004.2018.1511383&domain=pdf&date_stamp=2018-12-20

The challenges of using satellite data sets to assess historical land usechange and associated greenhouse gas emissions: a case study of threeIndonesian provinces

Sybrand van Beijmaa, Julia Chattertonb, Susan Pagec, Chris Rawlingsa, Richard Tiffind and Henry Kingb

aAgrimetics Ltd., Rothamsted Research, Harpenden, United Kingdom; bSafety and Environmental Assurance Centre, Unilever R&D,Colworth Science Park, Sharnbrook, Bedfordshire, United Kingdom; cDepartment of Geography, University of Leicester, Leicester,United Kingdom; dAgrimetics Ltd., University of Reading, Reading, United Kingdom

ABSTRACTAdvances in satellite remote sensing and the wealth of earth observation (EO) data nowavailable have improved efforts toward determining and quantifying historical land use andland cover (LULC) change. Satellite imagery can overcome the absence of accurate recordsof historical land use; however, the variability observed in the case study regions demon-strates a number of current challenges.Differences in spatial coverage, resolution and land cover classification can lead tochallenges in analyzing historical data sets to estimate LULC change and associated GHGemissions. This paper demonstrates the calculation of LULC change from three existing,open-source data sets to show how this can lead to significant variation in estimates of GHGemissions related to differences in land classification methodologies, EO input data andperiod of investigation. This article focuses on selected regions of Indonesia, where quantify-ing land use change is important for GHG assessments of agricultural commodities and forevidencing progress against corporate and government deforestation commitments.Given the significance of GHG emissions arising from LULC change and the increasing needfor emissions monitoring, this research highlights a need for consensus building to developconsistency in historical and future LULC change estimates. This paper concludes with a setof recommendations for improvements to ensure consistent LULC mapping.

KEYWORDSLand use/land coverchange; GHG emissions;remote sensing; palm oil;sustainability

Introduction

Advances in satellite remote sensing (RS) and thewealth of earth observation (EO) data now avail-able have improved efforts toward accuratelymapping land use and land cover (LULC) andquantifying change [1]. This reduces reliance on,for example, ground-level monitoring, andimproves the resolution of assessments that arecurrently based on country-level statistics.However, challenges remain, and factors such asthe type of data (e.g. optical or radar) and spatialand temporal resolution of satellite data may sig-nificantly influence the classification of LULC [1,2].Several organizations have produced and madeopenly available LULC data sets based upon theinterpretation of optical EO satellite data. Theseare derived from different satellites, based on dif-ferent sensors, with variations in return time andLULC classification methodology. This paper ana-lyzes uncertainty in greenhouse gas (GHG)

emission estimates by calculating LULC changewith three historical LULC data sets, with a focuson selected regions of Indonesia where the devel-opment of the palm oil (PO) industry has been asignificant driver of LULC change in recentdecades [2].

LULC mapping

Mapping of LULC is one of the key applications ofRS technologies and has been carried out for atleast 40 years [3]. However, there is little agree-ment on best practices for LULC mapping.A recent overview of different LULC mappingmethodologies is provided by Joshi et al. [1]. Theprocess of RS image classification is complex andinvolves many steps, including the determinationof a land-cover classification system, collection ofdata sources and selection of a classification algo-rithm [4]. One of the most important considera-tions in LULC mapping is the definition of LULC

CONTACT Sybrand van Beijma [email protected] Agrimetrics Ltd., c/o Rothamsted Research, West Common, HarpendenAL5 2JQ, United Kingdom

Supplemental data for this article can be accessed here.

� 2018 Informa UK Limited, trading as Taylor & Francis Group

CARBON MANAGEMENT2018, VOL. 9, NO. 4, 399–413https://doi.org/10.1080/17583004.2018.1511383

http://crossmark.crossref.org/dialog/?doi=10.1080/17583004.2018.1511383&domain=pdfhttps://doi.org/10.1080/17583004.2018.1511383http://www.tandfonline.com

classes. This can be done with a focus on land use(the purpose for which humans use land) or afocus on land cover (physical properties of a landsurface) [1]. LULC class definitions can be eitherbroad (e.g. forest, agriculture, grassland, etc.) orspecific (e.g. subdividing agricultural land into oilpalm, corn, banana, etc.). Optimal class definitiondepends on the specific needs of the user, but, ingeneral, broad classes are better suited for large-scale (continental) LULC mapping. While higherspecificity in land classes is preferable for regional-or national-scale land mapping studies [1], it hasbeen shown that using a large number of highlyspecific classes can lead to misclassification, as dif-ferences between classes become small [4,5].

Another major consideration when developing anLULC classification scheme is the selection of optimalRS input data. Low-resolution (LR) optical sensors(e.g. MOderate Resolution ImagingSpectroradiometer (MODIS), MEdium ResolutionImaging Spectrometer (MERIS)) have been useful forvegetation mapping at the global or continentalscale, while medium-resolution (MR) satellites (e.g.Landsat Thematic Mapper (TM)) are most frequentlyused for regional LULC mapping [6]. High-resolution(HR) satellite data (e.g. DigitalGlobe, Satellite Pourl’Observation de la Terre (SPOT)) require greaterresources in terms of processing capacity and can becostly when large area coverage needs to beacquired. Therefore, HR data is more likely to beused for validation of smaller areas [6]. The quality ofRS imagery can be hampered by persistent cloudcover in tropical regions [2]. Integrated use of syn-thetic aperture radar (SAR) satellite data, which hashigh resolution capability and is unaffected by cloudcover, has shown to improve LULC mapping signifi-cantly [1] and is becoming more commonly used intropical LULC mapping [7].

The classification methodology used for LULCmapping is a third major consideration. There is aplethora of image classification algorithms andmethodologies available [1]. Common methodolo-gies or algorithms range from statistical methods(e.g. maximum likelihood classification (MLC), prin-cipal component analysis (PCA)) [8,9] to machinelearning algorithms (support vector machine(SVM), random forest (RF)) [10–12], knowledge-based/decision tree methods [5, 13] and visual/manual interpretation of satellite data [12].Changes in LULC class definitions, RS data inputand classification methods over time can lead toissues of inconsistency and variability in estimatesof historical LULC change [2].

GHG emissions attributable to LULC change

Carbon dioxide emissions from fossil fuel use arerelatively well quantified, but GHG emissions fromLULC change remain highly uncertain yet are oneof the largest anthropogenic sources of GHG emis-sions [14]. Land-use changes can cause emissionsdue to carbon losses in both biomass and soils[15]. Rapid expansion of agriculture for large-scalecommodity crops can lead to large changes in car-bon stocks [16].

Understanding emissions from LULC change iskey to quantifying life-cycle emissions of large-scale agricultural commodities, such as PO. Growthin PO production in Southeast Asia, led primarilyby Indonesia and Malaysia, has been a key compo-nent of meeting growing global demand for bio-based oil in recent decades. Indonesia andMalaysia currently meet more than 85% of globalPO demand, 51% and 34% respectively [17]. Inthese countries, plantations cover an estimatedarea of 140,000 km2 on both mineral and organic(peat) soils, which has led to large-scale LULCchange in the region [2].

A historical record of 20–25 years is necessaryfor LUC emissions to be included in life-cycleassessments (LCAs). Openly available satellite datawith global coverage, and of sufficient quality,does not widely exist from prior to 2000 and,therefore, this period is rarely covered by LULCdata sets.

Significance of peat soils

Soils in wetland ecosystems (e.g. peat swamp for-ests) contain large amounts of organic material,and therefore have high belowground carbonstocks with carbon densities that may exceedthose of the aboveground vegetation [18]. Whenorganic soils are disturbed – and particularly whendrained, removing water from the soil pores – oxy-gen can enter the soil surface and oxidize the soilorganic material through biological and chemicalprocesses. Oxidation of soil organic matter leads toa carbon flux to the atmosphere, mostly asCO2 [19].

GHG emissions after drainage are not constant;they will vary as water tables and peat characteris-tics change [20]. In typical PO plantation develop-ments on peat soils in Southeast Asia, the initialpeatland drainage usually involves a rapid lower-ing of the water table to depths of around orbelow 1 m to over 3 m. In the first few months oryears after drainage, the peat surface will change

400 S. VAN BEIJMA ET AL.

rapidly through a combination of peat oxidationand soil compression. In this transition phase,carbon emissions are higher than during the sub-sequent, more stable phase (i.e. following palmplanting), when water levels will generally bemaintained at depths of around 0.80 m. From thatpoint onward, oxidation will proceed at a more orless stable rate until the peat surface is at or closeto the local drainage level; depending on the peatdepth, this may take several decades [20].

Any holistic assessment of the carbon emissionsarising from LULC change must include changes inboth above- and belowground carbon stocks. Therelative proportion of PO plantations on organicsoils in Southeast Asia has increased over the last20 years; these now occupy some 31,000 km2, orapproximately 23% of the total area under POplantations [21]. It has been shown that this pro-cess has been responsible for generating substan-tial carbon losses and associated GHG emissionsfrom peat decomposition [19].

Aim of this paper

The aim of this paper is to evaluate and compareexisting LULC data sets, derived from EO data, toassess historical LULC change and associated GHGemissions. To achieve this, this article focuses onthree Indonesian provinces where large-scale LULC

change has been observed in recent decades,much of which is attributable to the developmentof plantations.

Materials and methods

Study area

This study focuses on three areas of interest (AOIs),namely the Indonesian provinces of NorthernSumatra, Riau on the island of Sumatra, andCentral Kalimantan on the island of Borneo (Figure1). These three AOIs, covering approximately onesixth of the total area of Indonesia, lie within anarea that is the focus of much attention surround-ing land-use change emissions [22–24]. All of theAOIs include areas with peat soils, according tothe peat soil map distributed by the Centre forRemote Imaging, Sensing and Processing (CRISP)in Singapore [19]. Additionally, in all three AOIs POproduction occurs on both mineral and peat soils,according to PO concession data obtained fromGlobal Forest Watch [25] (see Table 1).

LULC data sources

Three open-source, satellite-derived LULC data setswere identified as thematically and spatially rele-vant for the AOIs, as detailed in Table 2.

Figure 1. The areas of interest in Indonesia, with palm oil (PO) plantation concessions and peat soil areas indicated.

CARBON MANAGEMENT 401

The Climate Change Initiative (CCI) LULC dataset was developed by the European Space Agency(ESA) CCI Land Cover Initiative, currently availablewith updates for the period 1992–2015. CCI is aglobal LULC data set, with a class definition basedon the Land Cover Classification System (LCCS)developed by the United Nations (UN) Food andAgriculture Organization (FAO) [26]. Class defini-tions are broad, with no specific LULC classes fortree plantations. Quality assessment of the CCIdata set (included in [26]) was based on referenc-ing using higher resolution satellite data or derivedproducts (Landsat, Google Earth, SPOT-Vegetation(SPOT-VGT)) for specific reference areas, whichwere chosen to cover all global climatic zones,with subsamples chosen randomly from theseareas. The overall accuracy of the CCI 2010 dataset compared with a reference data set for 2009was 74.4%.

The CRISP LULC data set was developed byCRISP in Singapore and covers Southeast Asia,with updates for 2000, 2010 and 2015. The map-ping methodology is well documented [21, 27].The 2015 LULC data update has been developedusing a methodology that differs significantly fromthat used for the 2000 and 2010 updates; CRISPhas therefore advised users to avoid comparisonsof the 2015 data with older updates for LULCchange analysis [21]. The class definition is specific,with two classes for plantations (‘Large scale palmplantations’ and ‘Plantation/regrowth’). Qualityassessment of the CRISP data set was carried outby comparing the LULC maps with a total of 1000random sample plots from very high-resolutionsatellite data [21]. The total accuracy for the 2015CRISP data set was 81.6%.

The MoF LULC data set was developed at theIndonesian Ministry of Forestry, and currently pro-vides irregular updates between 1990 and 2015. Intotal, 10 updates are available, of which eight arebetween 2000 and 2015. There is no accompany-ing documentation detailing the image classifica-tion methodology used for the LULC mapping.However, according to [23], it is primarily based onvisual interpretation of Landsat 30� 30 m satellitedata. There is no indication of whether any qualityassurance checks have been carried out. Whenconsidering forest cover in Indonesia, comparisonbetween MoF and Global Forest Watch forestcover data [28] indicated agreement in 90.2% ofthe area considered [29]. The MoF LULC classes arespecific, identifying two plantation types (generalplantation and timber plantation) as well as undis-turbed and disturbed forests.

Data pre-processing

Figure 2 presents an overview of the processingand analysis workflow. After collection of theLULC, AOI boundary and peat soil extent data, thedata is pre-processed using the following steps:

� Conversion of LULC data to raster: the MoFdata is delivered in vector format, in order tomake the data set comparable in terms of reso-lution; it was converted to raster with a100� 100 m spatial resolution;

� Subsetting of LULC and peat soil data per AOI;� Splitting of LULC data between peat soil and

non-peat soil areas;� Reprojection of data to the same Universal

Transverse Mercator (UTM) zone projection:

Table 1. Geographical extent and area of peat soil cover [19] and palm oil (PO) concessions [25] of the study areas.

Province Total area (ha) Peat soil area (ha)Peat (% oftotal area)

PO concessionarea (ha)

PO concession (%of total area)

PO on peatsoil (ha)

PO on peat (% oftotal

concession area)

North Sumatra 7,243,839 347,925 4.8 132,538 1.8 61,203 46.2Riau 8,995,724 4,004,336 44.5 2,117,307 23.5 819,769 38.7Central

Kalimantan15,354,930 3,005,097 19.6 3,199,420 20.8 464,079 14.5

Table 2. Overview of land use and land cover (LULC) data sets used in this research.Organization Abbreviation Spatial resolution (m) Spatial extent Updates URL of data repository

European Space Agency(ESA) Climate ChangeInitiative Land Cover

CCI 300� 300 Global Annual between 1992and 2015

http://maps.elie.ucl.ac.be/CCI/viewer/

Centre for RemoteImaging, Sensing andProcessing, Singapore

CRISP 250� 250 Southeast Asia2000, 2010, 2015

https://ormt-crisp.nus.edu.sg/ormt/Home/Disclaimer

Indonesia Ministryof Forestry

MoF 30� 30 (100� 100 usedfor this research)

Indonesia1990, 1996, 2000, 2003,2006, 2009, 2011, 2012,2013, 2015

http://www.greenpeace.org/seasia/id/Global/sea-sia/Indonesia/Code/Forest-Map/en/index.html


http://maps.elie.ucl.ac.be/CCI/viewer/http://maps.elie.ucl.ac.be/CCI/viewer/https://ormt-crisp.nus.edu.sg/ormt/Home/Disclaimerhttps://ormt-crisp.nus.edu.sg/ormt/Home/Disclaimerhttps://ormt-crisp.nus.edu.sg/ormt/Home/Disclaimerhttp://www.greenpeace.org/seasia/id/Global/seasia/Indonesia/Code/Forest-Map/en/index.htmlhttp://www.greenpeace.org/seasia/id/Global/seasia/Indonesia/Code/Forest-Map/en/index.htmlhttp://www.greenpeace.org/seasia/id/Global/seasia/Indonesia/Code/Forest-Map/en/index.htmlhttp://www.greenpeace.org/seasia/id/Global/seasia/Indonesia/Code/Forest-Map/en/index.htmlhttp://www.greenpeace.org/seasia/id/Global/seasia/Indonesia/Code/Forest-Map/en/index.html

UTM 47N for North Sumatra and Riau AOI data,UTM 49N for Central Kalimantan;

� Class aggregation of specific LULC classes intobroad classes for the cross-comparison of LULCdata, as detailed below.

Cross-comparison of LULC data

Pairwise comparison of the three LULC data setswas carried out using Mapcurve analysis [30].Mapcurve analysis provides a method to comparetwo categorized maps by cross-referencing, toquantify the similarity between the classifications.This analysis provides insight by calculating theproportion of overlap between each LULC class

from one data set (Map A) and the best-overlap-ping LULC class from another data set (Map B).The best overlaps for all classes from Map A withclasses from Map B are calculated, and the overlapfractions are summed to derive the total agree-ment between Map A and Map B. This total isnamed the goodness of fit (GoF); a GoF of 1.0means a perfect fit, while a GoF of 0.0 indicates nofit at all. This analysis can be run both ways – thatis, using map A as the original and using Map B asthe reference, or vice versa. The GoF is expressedas a percentage and can therefore be comparedacross categories and maps.

It should be noted that the GoF does not giveinformation about the total area of agreement, aseach LULC class has equal influence on the GoF,regardless of its area of presence in the originalmap. Nor does this analysis provide insight intothe relative quality of data sets; rather, it gives anindication of the proportion of overlap.

To make the three LULC data sets as compar-able as possible, LULC classes were aggregatedinto nine broad classes, based on a general classaggregation utilized for the CCI data [26]:Agriculture, Forest, Grassland, Shrubland, SparseVegetation, Wetland, Settlement, Bare and Water.

The cross-comparison analysis was run for datespertaining to two specific years for which all threeLULC data sets have an update, 2000 and 2015(Figure 3a–f).

LULC change analysis

To calculate LULC change for each LULC data set,changes from each initial update (t0) to the nextupdate (t1) were calculated from the pre-proc-essed data. This was done by comparing each pixellocation from the t0 raster data with each corre-sponding pixel from the t1 data. If a change inLULC class was observed, the pixel was reclassedas a pixel with a unique value combining the t0and t1 class code. If no change was observed, thepixel was reclassed as having no value (see FigureS1). From this analysis, LULC change maps andtables were produced. Table S1 provides the timeperiods used to assess LULC change. For CCI andMoF, these time periods coincide with the updatesof the MoF data set; for CRISP only one period hasbeen used, 2000 to 2010, as the update of CRISPfor 2015 cannot be compared for LULC changeanalysis [21]. The LULC change is expressed in hec-tares per year, to correct for varying time intervalsbetween updates.

Figure 2. Data analysis workflow diagram.


Carbon emission modelling

To convert LULC change into carbon emission esti-mates, values for aboveground biomass (AGB) and

organic soil degradation (OSD) emissions factorswere obtained for all the LULC classes of the threedata sets. This was done by conducting a review of

Figure 3. Best-fitting mapcurve plots for North Sumatra (a,d), Riau (b,e) and Central Kalimantan (c,f) for 2000 and 2015,respectively.


published literature related to LULC change inSoutheast Asia (primarily based on [15, 20, 31–33]).From this review, average values for AGB and OSDfor each LULC class were calculated (Tables S2 andS3). AGB emission factors are expressed in Mg C/ha, and the OSD emissions are given in Mg C/ha/yras these continue for an indefinite period after anLULC change from a natural to man-made state[19]. The LULC change data from the selectedareas and the AGB and OSD emission values werecombined to estimate GHG emissions. The model,Equation (1), is a simplified version of the model in[34], not taking into account the GHG emissionsrelated to peat fire due to additional uncertainty:

E ¼ Ea�Sa þ Ebo (1)where E is the emission estimate, Ea is emission

from AGB due to LUC, Sa is the sequestration ofCO2 from the atmosphere into crop biomassbetween succeeding land uses, and Ebo is emissionfrom OSD. A graphical example of the model isprovided in Figure S2. For example, if 1 ha changesfrom Primary Forest (average AGB 233Mg C/ha) toShrubland (average AGB 31Mg C/ha), then, forAGB, a total of 233� 31¼ 202Mg C will be emit-ted. If subsequently this 1 ha of Shrublandbecomes Plantation (average AGB 37Mg C/ha),then the net carbon emissions will be 31� 37=� 6Mg C, which indicates carbon sequestration.

The latest insights with respect to emissionsfrom drained peatlands are reported by IPCC [20,35]. The OSD emission factor values used in thispaper relate to ongoing oxidation of peat.Additional emissions occurring during the first 5years after drainage for plantation establishment[20, 36], relating to fires [37], and the potentialemissions from organic carbon flushed into aquaticecosystems (e.g. as dissolved organic carbon(DOC), and associated emissions of CO2 and CH4[38]), are all excluded. These emissions are highlyuncertain and would therefore obscure the uncer-tainty in GHG estimates from different LULC datasets. Thus, in the authors’ calculations, if (forexample) Peatswamp Forest on organic soilchanges to PO Plantation, the OSD emissionrelated to land conversion is 11Mg C/ha/yr.

On the basis of Khasanah et al. [31], whoreported that, for mineral soils, the net temporaltrend in the soil carbon stock (in the top 30 cm ofsoil) was not significantly different from zero inboth forest- and non-forest-derived plantations,soil carbon stock neutrality on mineral soils usedfor oil palm cultivation is assumed.

Results and discussion

Cross-comparison LULC data (Mapcurves)

The Mapcurve plots with the highest consistenciesfor each area and date are visualized in Figure 3.The highest GoF values are observed for either acombination of CCI as the original and CRISP asthe reference map or the combination of MoF asoriginal and CRISP as reference map. The highestGoF observed is 0.575 for North Sumatra in 2015,by combining MoF and CRISP, which means thereis 57.5% class agreement between these maps. Allother combinations lead to lower GoF values (seeTables S4 and S5). The two data types most dis-similar are the MoF and CCI data sets (generallyless than 40% class agreement).

The Mapcurve analysis shows large inconsisten-cies between LULC data sets, even after aggrega-tion of specific LULC classes to make the data setsmore comparable. Other comparative studies ofLULC data sets have also observed this [39–41],either by means of the Mapcurve analysis or byanalyzing spatial overlap of similar classes on apixel-by-pixel basis. Among these studies, the max-imum observed Mapcurve value was 0.53 [41],while in the pixel-by-pixel analysis the highestagreement was found to be 62% [40]. This showsthat, even after aggregation of specific LULCclasses into broader classes, high levels of agree-ment between LULC maps of similar age cannotbe assumed.

Differences between LULC maps can be causedby a number of factors [42], including data quality,spatial and temporal resolution, LULC classificationapproaches, algorithms and aggregation. Dataquality can be limited in tropical regions due topersistent cloud cover, and therefore present a lim-ited number of useful satellite acquisitions. If suffi-cient temporal resolution is available, there is abetter chance that high-quality imagery can beobtained in a given period.

Spatial resolution dictates the smallest mappingunit. In general, if a pixel is sufficiently small, morespecific LULC can be distinguished. Lower spatialresolution pixels often cover more than one spe-cific LULC class, and therefore the LULC class defin-ition must be more generic, as for CCI. Spectralresolution influences how well LULC classes can betechnically distinguished. MODIS data, whichunderlies the CRISP data set, operates in 34 spec-tral bands [43], whereas Landsat-8 operates in 11bands [44]. This means that even though MODIShas a spatially lower resolution than Landsat,


through its superior spectral resolution, MODISmight be able to detect more subtle variations inLULC than Landsat can.

As noted above, several classification algorithmswere used to develop the LULC data sets, whichhas an effect on differences in mapping results. Ingeneral, pixel-based classifiers tend to lead to highheterogeneity in the resulting LULC map, as eachpixel is individually classified. Therefore, it is cur-rently more common to include a clustering stepin the classification process, as this has been foundto positively influence the map accuracy [45,46].The MoF data set is based on visual interpretationof satellite data, which depends on the interpret-ation skills of each person working on the LULCmaps, which can be subjective [47]. LULC class def-inition can impact how readily LULC data sets canbe compared. Aggregation of specific LULC classesinto broad classes can overcome this problem to alarge extent, although it is not always clear towhich broad class a specific class might belong.

LULC change

For each LULC data set, LULC change has been cal-culated for each period between updates (shownin Table S3). The LULC change observed in eacharea is presented for North Sumatra (Figure 4),Riau (Figure 5) and Central Kalimantan (Figure 6).The LULC change is averaged to give LULC changein hectares per year, to make the differing periodsbetween updates directly comparable.

According to the MoF data set, for each AOI, one“peak change period” with an extreme LULCchange is recorded. In North Sumatra this is2006–2009; it is 2012–2013 in Riau, and 2013–2015

in Central Kalimantan. From the data (Table S6), theNorth Sumatra peak change period is 4.7 timeslarger than the average for 2000–2015, 2.5 timeslarger than average for the Riau peak changeperiod and 3.8 times larger than the average inCentral Kalimantan. It is questionable whetherthese changes, visible in the RS data, are related to‘actual observed’ LULC change in the AOI, which isdefined here as change that can be seen at groundlevel (corroborated by field observations), or haveother causes. Table 3 shows the largest contribu-tors to these peak change periods, according tothe MoF data sets. Analysis shows that for eachoccurrence the MoF class “Dry Rice Land Mixedwith Scrub” was involved, either by transition fromthis class to “Dry Rice Land” (in North Sumatra) orby transition into it from “Scrubland” (both Riauand Central Kalimantan). The class “Dry Rice LandMixed with Scrub” can be interpreted as a transi-tion class between Rice Land and Scrubland, or anecotone. Defining class boundaries for ecotones isoften difficult when making observations in thefield; it is even more challenging when interpretingRS data [48]. Due to the magnitude of these peakchange periods, it is unlikely that they are relatedto actual changes in LULC, and more likely thatthey are related to different interpretations of RSdata or a methodological shift between MoFupdates. However, the influence of this mappingeffect on the LULC change observed in CentralKalimantan in 2013–2015 is relatively small, andthe majority of LULC change estimated for thisperiod can be attributed to ‘actual observed’ LULCchange at ground level.

The CRISP maps consistently give higher esti-mates for land use change than either the CCI or

Figure 4. Land use and land cover (LULC) change in North Sumatra between 2000 and 2015.


MoF map. The annual LULC change estimated byCRISP is between 6.1 and 12.8 times larger thanthe LULC change from CCI for the AOIs. For CRISPcompared to MoF, the difference ratios for LULCchange lie between 2.2 and 3.8.

Temporal correlations between MoF andCCI data are plotted in Figure 7. As CRISP only pro-vides one update (between 2000 and 2010), thisdata set has not been included. The MoF LULCchange values have been corrected for the peak

change periods described above, to get a bettercomparison of actual observed LULC changebetween CCI and MoF data sets. The strongesttemporal correlation is shown in North Sumatra,with an R2 value of 0.7978, while those for Riauand Central Kalimantan are much lower.

The LULC change analysis shows little agreementbetween LULC data sets in the AOIs. While someinconsistency can be attributed to methodologicalfactors, not all of it can be explained directly.

Figure 5. Land use and land cover (LULC) change in Riau between 2000 and 2015.

Figure 6. Land use and land cover (LULC) change in Central Kalimantan between 2000 and 2015.

Table 3. Main contributors to land use and land cover (LULC) change for largest observed Indonesia Ministry ofForestry LULC changes for all areas of interest (AOIs).

AOI PeriodTotal LULC

change/yr (ha)Largest LULCchange/yr (ha) From LULC class t0 –> to LULC class t1 % of total

North Sumatra 2006–2009 638,860 407,018 Dry Rice Land Mixed w/Scrub –> Dry Rice Land 63.7Riau 2012–2013 1,118,233 726,066 Scrubland –> Dry Rice Land Mixed w/Scrub 64.9Central Kalimantan 2013–2015 1,237,019 274,672 Scrubland –> Dry Rice Land Mixed w/Scrub 22.2


GHG emissionsLarge variability in GHG emissions can be observedfor estimates made using the different LULC datasets (see Tables 4 and 5; see also Figures S3–S5).GHG emissions estimates from CRISP data(2000–2010 only) are considerably higher thanthose from either CCI or MoF data, while those

from MoF for the period 2011–2015 are generallymuch higher than the estimates from CCI (Table5). For Riau and Central Kalimantan, this is partlydue to the inconsistency in MoF data related tothe classification of “Scrubland” and “Dry Rice LandMixed with Scrub.” The peak change periods arealso visible, with a peak in emissions in Riau in2012–2013 (Figure S4) and in Central Kalimantanin 2013–2015 (Figure S5).

These results, which illustrate considerable vari-ability in GHG emission estimates from the differ-ent LULC data sets, are supported by other studies,such as Agus et al. [34], who estimated that carbonemissions from LULC change studies related to thePO industry in Kalimantan differed by a factorof 4.7.

GHG emission maps

The GHG emission estimates per data set, area andtime period can also be displayed geographicallyin maps of GHG emissions (Figure 8). The highestmodelled GHG emissions occur in areas with peatsoils, primarily in Riau and Central Kalimantan.However, there are also regions where net carbonsequestration occurs, likely related to conversionof low-biomass LULC (bare areas, shrub) to higherbiomass LULC (plantations). Bare and shrublandareas may be the result of previous deforestation,which highlights the need for sufficient historicaldata to understand and account for emissionsfrom LULC change over a longer period, especiallyin peat soil areas. Several methods exist to attri-bute these emissions to a product, depending onthe data available [49]. For LCA, the impact of landuse change should include all direct land usechange occurring 20 years (or one full harvest,

Figure 7. Scatter plot land use and land cover (LULC) change estimates in all three areas of interest in the period2000–2015 from CCI and MoF.

Table 4. Greenhouse gas (GHG) emissions for three areasof interest for 2000–2010/2011, with percentage ofemissions from mineral/peat.Emissions per year (Mg C/yr) and percentage of total

North Sumatra (4.8% peat)CCI (2000–2010) CRISP (2000–2010) MoF (2000–2011)

Mineral 1,332,803 34.5 4,943,071 58.7 1,127,552 38.4Peat 2,526,168 65.5 3,473,117 41.3 1,807,528 61.6Total 3,858,971 8,416,188 2,935,080Riau (44.5% peat)

CCI (2000–2010) CRISP (2000–2010) MoF (2000–2011)Mineral 9,533,167 33.2 11,784,303 28.7 4,812,749 17.2Peat 19,174,983 66.8 29,246,758 71.3 23,105,593 82.8Total 28,708,150 41,031,060 27,918,343Central Kalimantan (19.6% peat)

CCI (2000–2010) CRISP (2000–2010) MoF (2000–2011)Mineral 6,699,626 62.3 14,791,021 55.9 8,275,277 61.7Peat 4,055,275 37.7 11,693,997 44.2 5,142,602 38.3Total 10,754,901 26,485,018 13,417,880

Table 5. Greenhouse gas (GHG) emissions for three areasof interest for 2010/2011–2015, with percentage ofemissions from mineral/peat.Emissions per year (Mg C/yr) and percentage of total

North Sumatra (4.8% peat)CCI (2010–2015) MoF (2011–2015)

Mineral 491,500 47.8 3,450,122 80.3Peat 536,465 52.2 845,314 19.7Total 1,027,966 4,295,435Riau (44.5% peat)

CCI (2010–2015) MoF (2011–2015)Mineral 4,055,092 32.12 6,200,630 27.1Peat 8,571,629 67.88 16,672,805 72.9Total 12,626,721 22,873,435Central Kalimantan (19.6% peat)

CCI (2010–2015) MoF (2011–2015)Mineral 2,814,481 58.5 10,357,934 48.6Peat 1,994,610 41.5 10,941,092 51.4Total 4,809,091 21,299,027


whichever is longer) prior to the assessment. Thetotal GHG emissions (or removals) arising fromLULC change over this period would be allocatedequally to each year of the period [50].

Carbon emission factorsThe values for emissions from AGB and OSD,derived from the literature, are key to the GHGemissions calculations in this study. For planta-tions, the AGB value used in this study is 37Mg C/ha, based on a time-averaged value [15]. However,AGB values of 57.5Mg C/ha have been reportedfor plantations at full maturity [51]. To understandthis sensitivity, results were calculated using thisvalue (Table S7). In all but one instance, annualemissions are reduced by 1–33% when using thehigher carbon stock value for plantation classes. Inone case emissions increase (MoF, 2011–2015 forNorth Sumatra) because a large area of plantationwas converted to a land use with lower car-bon stock.

Temporal intervalThe results are also sensitive to the interval periodused between LULC map updates. This has animpact on GHG emissions related to OSD. It hasbeen shown that emissions from OSD can con-tinue for an indefinite time period after conversion

from a natural to man-made state [20]. This pro-cess is sketched in Figure S2, where emissionsrelated to soil degradation from the first stage ofLUC continue into the second stage of LUC. It hasbeen observed that when OSD emissions from aprevious period are not included, GHG emissionsestimates can vary significantly. This has been ana-lyzed with CCI data for Central Kalimantan, whereGHG estimates derived for 5-year intervals(2000–2005 and 2005–2010) were found to beapproximately 1 million Mg C/yr higher than emis-sions summed at intervals of 1 year, over the same10-year period. This shows that the GHG emissionmodel should incorporate LULC changes onorganic soils predating the period of interestwherever possible.

Limitations in estimating GHG emissionsThe GHG emission estimates were calculated basedon published carbon stock and emission factors fordifferent LULC classes. Variability and uncertaintyin carbon stocks can be observed in the range ofvalues in the literature (Tables S2 and S3) arisingfrom influences including soil type and climate, orwhere different studies include different elementsof carbon pools [15]. Furthermore, peat soils mayvary in depth and volume and therefore influencecarbon stocks [52]. Since the purpose of this study

Figure 8. GHG emission map of areas of interest, based on MoF data for the period 2000–2015.


was to establish uncertainty in GHG estimatesresulting from the use of different LULC products,variability and uncertainty in carbon stock valuesfor different LULC classes were not consid-ered further.

Further work could be done to integrate varia-bilities in carbon stock accounting with the varia-bilities in estimated LULC changes estimated inthis study. Key considerations would include spa-tial heterogeneity (edge effects) in above- andbelowground carbon stocks within land coverclasses [53,54]; variability in water table depth andcarbon loss rates for OSD [20]; and uncertainties inemissions from land clearance fires on peat soils[19]. Emissions from degraded peat soils areknown to continue a long period, often longerthan a typical LCA analysis period of 20–25 years[52]. Therefore, wherever possible, it is advised toincorporate any known historical LULC changes onpeat soils for a period as long as possible.

The finding that LULC maps based on RS datainterpretation differ is not new [40,41], and someattempts have been made to improve comparabilitybetween LULC maps [39]. An LULC data set is alwaysa trade-off between the input data quality andaccessibility, requirements of end-users and thetechnological and financial means available fordevelopment. Each of these data sets represents avaluable source of spatially explicit information forcalculating GHG emissions related to LULC change.The current variability between LULC maps suggeststhat estimates should be used to provide a rangerather than a single value for GHG emissions.

Conclusions and recommendations

The need to quantify GHG emissions associatedwith LULC change is important for LCAs of agricul-tural commodities and for providing evidence ofGHG reductions associated with zero net deforest-ation commitments. Without RS data, such calcula-tions would require detailed historical landrecords, and therefore these data sets are valuablefor estimating regional trends in LULC change andassociated GHG emissions. However, this study hasshown the potential variability in estimates thatcan be obtained through the use of three open-source RS data sets. These variabilities arise fromdifferences in EO input data, land classificationmethodologies, data resolution and period ofinvestigation. It is therefore advisable to comparedifferent LULC data sets in parallel and use thevariability between GHG emission estimates as a

confidence interval, rather than using a singlevalue. Users should be aware of the potential forvariability in LULC estimates.

GHG emission maps, such as that shown inFigure 8, are useful visualization tools that are notcommonly available and can provide spatialinsight into LULC change and related carbon emis-sions or sequestration. Current web-based plat-forms may show forest loss or LULC for a givenperiod, but, to the authors’ knowledge, do not yetprovide maps showing associated GHG emissions.Given the inconsistencies highlighted in this paper,there is a need for further work to ensure themaps provide robust estimates of LULC changeand associated emissions.

The method described in this paper can beused to provide spatially improved estimates ofLULC change and GHG emissions, particularlywhere the change occurs between LULC typeswith significantly different carbon stock values,such as between primary forest and plantation. Forthis to be most effective, there is a need for con-sensus building and harmonization on how todevelop a consistent and robust approach toassessing historical LULC change, to provide evi-dence for zero net deforestation commitments,and refine GHG assessments.

The following recommendations are made toimprove LULC mapping for GHG emission esti-mates for agricultural commodities.

First, the LULC class definition should focus onLULC classes closely associated with the main driv-ers of LULC change in the AOI. This should includeat least the following classes: primary and second-ary forest, several types of plantation (whereapplicable), bare land, and cropland. Global LULCdata sets often use class definitions that are toobroad or lack specific class distinctions importantfor GHG modelling. Additionally, class definitionsof different LULC data sets should be more com-parable. The FAO LCCS definitions were developedto be globally relevant and flexible enough to suitmost environments [55]. The CCI classes are basedon LCCS; it could be useful for other organizationsinvolved in land cover mapping to adopt this sys-tem as well.

Second and ideally, maps should be updated atleast every 2–3 years – and annual updates wouldbe preferable – to capture rapid changes, such asdeforestation (fire or logging), bare land, and plan-tation development.

Third, to enable LULC change analysis overtime, mapping methodology should remain


unchanged (a period of 20–25 years is required forLCA). If, for example, better mapping algorithmsare developed, such that the methodology can beimproved significantly, it would be preferable toreprocess the historical data using the new meth-odology to maintain consistency.

Fourth, optimal spatial resolution is dependenton the requirements of the user; research on aprovincial level can be done at lower spatial reso-lution than needed at a smaller scale, for exampleat plantation level. For studies related to a specificagri-food industry it is often sufficient to focus ondata sets with a spatial coverage of the main pro-ducing areas.

Finally, metadata including quality and meth-odological information should be published withthe data sets.

When the three LULC data sets are comparedagainst these recommendations, it is clear there iscurrently room for improvement. Signs of improve-ments are visible, as the recent reprocessing of theESA’s CCI Land Cover initiative has shown. As RScapabilities are advancing quickly and the import-ance of LULC change analysis is becoming betterrecognized, this is an excellent time to addressthese recommendations to make LULC data aneven more valuable resource for environmen-tal monitoring.

Acknowledgements

The authors would like to thank Agrimetrics and Safetyand Environmental Assurance Centre (SEAC) and UnileverR&D for funding this research, as well as the anonymousreviewers for their constructive comments.

Disclosure statement

No potential conflict of interest was reported bythe authors.

References

1. Joshi N, Baumann M, Ehammer A, et al. A review ofthe application of optical and radar remote sensingdata fusion to land use mapping and monitoring.Remote Sens. 8(1), 1–23 (2016).

2. Carlson KM, Curran LM, Ratnasari D, et al. Committedcarbon emissions, deforestation, and community landconversion from oil palm plantation expansion inWest Kalimantan, Indonesia. Proc. Natl. Acad. Sci.109(19), 7559–7564 (2012). [Internet]. Available from:http://www.pnas.org/cgi/doi/10.1073/pnas.1200452109.

3. Anderson JR. A land use and land cover classificationsystem for use with remote sensor data. USGovernment Printing Office, Washington, DC.

4. Srivastava PK, Han D, Rico-Ramirez MA, Bray M, IslamT. Selection of classification techniques for land use/land cover change investigation. Adv. Sp. Res. 50(9),1250–1265 (2012). [Internet]. Available from: http://www.sciencedirect.com/science/article/pii/S0273117712004218.

5. Kuenzer C, van Beijma S, Gessner U, Dech S. Landsurface dynamics and environmental challenges ofthe Niger Delta, Africa: Remote sensing-based analy-ses spanning three decades (1986–2013). Appl. Geogr.53, 354–368 (2014).

6. Kim C. Land use classification and land use changeanalysis using satellite images in Lombok Island,Indonesia. Forest Sci. Technol. 12(4), 183–191 (2016).[Internet]. Available from: http://dx.doi.org/10.1080/21580103.2016.1147498.

7. Sameen MI, Nahhas FH, Buraihi FH, Pradhan B, ShariffARBM. A refined classification approach by integrat-ing Landsat Operational Land Imager (OLI) andRADARSAT-2 imagery for land-use and land-covermapping in a tropical area. Int. J. Remote Sens. 37(10),2358–2375 (2016). [Internet]. Available from: http://dx.doi.org/10.1080/01431161.2016.1176273.

8. Okamoto K, Kawashima H. Estimation of rice-plantedarea in the tropical zone using a combination ofoptical and microwave satellite sensor data. Int. J.Remote Sens. 20(5), 1045–1048 (1999).

9. Erasmi S, Twele A. Regional land cover mapping inthe humid tropics using combined optical and SARsatellite data case study from Central Sulawesi,Indonesia. Int. J. Remote Sens. 30(10), 2465–2478(2009). [Internet]. Available from: http://www.tand-fonline.com/doi/abs/10.1080/01431160802552728.

10. van Beijma S, Comber A, Lamb A. Random forest clas-sification of salt marsh vegetation habitats usingquad-polarimetric airborne SAR, elevation and opticalRS data. Remote Sens. Environ. 149, 118–129 (2014).[Internet]. Available from: http://www.sciencedirect.com/science/article/pii/S0034425714001485.

11. Dong J, Xiao X, Sheldon S, Biradar C, Xie G. Mappingtropical forests and rubber plantations in complexlandscapes by integrating PALSAR and MODISimagery. ISPRS J. Photogramm. Remote Sens. 74,20–33 (2012).

12. Balzter H, Cole B, Thiel C, Schmullius C. MappingCORINE land cover from Sentinel-1A SAR and SRTMdigital elevation model data using random forests.Remote Sens. 7(11), 14876–14898 (2015).

13. Haack BN, Herold ND, Bechdol MA. Radar and OpticalData Integration for Land-Use/Land-Cover Mapping.Photogramm. Eng. Remote Sensing. 66, 709–716(2000).

14. Maria Roman-Cuesta R, Rufino M, Herold M, et al.Hotspots of gross emissions from the land use sector:Patterns, uncertainties, and leading emission sourcesfor the period 2000–2005 in the tropics.Biogeosciences. 13(14), 4253–4269 (2016).

15. Agus F, Henson IE, Sahardjo BH, et al. Review of emis-sion factors for assessment of CO2 emission fromland use change to oil palm in Southeast Asia.Reports from Tech. Panels 2nd Greenh. Gas Work. Gr.Roundtable Sustain. Palm Oil (RSPO) 7–28 (2013).


http://www.pnas.org/cgi/doi/10.1073/pnas.1200452109http://www.pnas.org/cgi/doi/10.1073/pnas.1200452109http://www.sciencedirect.com/science/article/pii/S0273117712004218http://www.sciencedirect.com/science/article/pii/S0273117712004218http://www.sciencedirect.com/science/article/pii/S0273117712004218http://dx.doi.org/10.1080/21580103.2016.1147498http://dx.doi.org/10.1080/21580103.2016.1147498http://dx.doi.org/10.1080/01431161.2016.1176273http://dx.doi.org/10.1080/01431161.2016.1176273http://www.tandfonline.com/doi/abs/10.1080/01431160802552728http://www.tandfonline.com/doi/abs/10.1080/01431160802552728http://www.sciencedirect.com/science/article/pii/S0034425714001485http://www.sciencedirect.com/science/article/pii/S0034425714001485

[Internet]. Available from: http://www.rspo.org/file/GHGWG2/3_review_of_emission_factors_Agus_et_al.pdf.

16. Germer J, Sauerborn J. Estimation of the impact of oilpalm plantation establishment on greenhouse gasbalance. Environ. Dev. Sustain. 10(6), 697–716 (2008).

17. FAO. Crops processed; Indonesia and World total; Oil,palm; Production quantity; 2014. FAOSTAT. (2016).[Internet]. Available from: http://www.fao.org/faostat/en/?#data/QD.

18. Page SE, Rieley JO, Banks CJ. Global and regionalimportance of the tropical peatland carbon pool.Glob. Chang. Biol. 17(2), 798–818 (2011).

19. Miettinen J, Hooijer A, Vernimmen R, Liew SC, PageSE. From carbon sink to carbon source: extensivepeat oxidation in insular Southeast Asia since 1990.Environ. Res. Lett. 12(2), 24014 (2017). [Internet].Available from: http://stacks.iop.org/1748-9326/12/i¼2/a¼024014?key¼crossref.d5803ab4c180cc18647fc14663f0d775.

20. Hooijer A, Page S, Canadell JG, et al. Current andfuture CO 2 emissions from drained peatlands inSoutheast Asia. Biogeosciences. 7(5), 1505–1514(2010).

21. Miettinen J, Shi C, Liew SC. 2015 Land cover map ofSoutheast Asia at 250 m spatial resolution. RemoteSens. Lett. 7(7), 701–710 (2016).

22. Carlson KM, Curran LM, Asner GP, Pittman AM, TriggSN, Marion Adeney J. Carbon emissions from forestconversion by Kalimantan oil palm plantations. Nat.Clim. Chang. 3(3), 283–287 (2012). [Internet].Available from: http://www.nature.com/doifinder/10.1038/nclimate1702.

23. Gunarso P, Hartoyo ME, Agus F, Killeen TJ. Oil Palmand Land Use Change in Indonesia, Malaysia andPapua New Guinea. Reports from Tech. Panels RSPOs2nd Greenh. Gas Work. Gr. 29–64 (2013).

24. Koh LP, Miettinen J, Liew SC, Ghazoul J. Remotelysensed evidence of tropical peatland conversion tooil palm. Proc. Natl. Acad. Sci. 108(12), 5127–5132(2011). [Internet]. Available from: http://www.pnas.org/cgi/doi/10.1073/pnas.1018776108.

25. Global Forest Watch. Oil palm concessions [Internet].2017. Available from: www.globalforestwatch.org.

26. ESA CCI. ESA CCI-Land Cover 2015 Product UsersGuide Version 2.0 [Internet]. Available from: http://maps.elie.ucl.ac.be/CCI/viewer/download/ESACCI-LC-Ph2-PUGv2_2.0.pdf.

27. Miettinen J, Shi C, Tan WJ, Liew SC. 2010 land covermap of insular Southeast Asia in 250-m spatial reso-lution. Remote Sens. Lett. 3(1), 11–20 (2012).

28. Hansen MC, Potapov PV, Moore R, et al. High-reso-lution global maps of 21st-century forest coverchange. Science (80-.) 342(6160), 850–853 (2013).[Internet]. Available from: http://www.sciencemag.org/cgi/doi/10.1126/science.1244693.

29. Margono BA, Potapov PV, Turubanova S, Stolle F,Hansen MC. Primary forest cover loss in Indonesiaover 2000–2012. Nat. Clim. Chang. 4, 730–735 (2014).

30. Hargrove WW, Hoffman FM, Hessburg PF. Mapcurves:A quantitative method for comparing categoricalmaps. J. Geogr. Syst. 8(2), 187–208 (2006).

31. Khasanah N, van Noordwijk M, Ningsih H, Rahayu S.Carbon neutral? No change in mineral soil carbonstock under oil palm plantations derived from forestor non-forest in Indonesia. Agric. Ecosyst. Environ. 211,195–206 (2015). [Internet]. Available from: http://dx.doi.org/10.1016/j.agee.2015.06.009.

32. IPCC. 2006. IPCC Guidelines for National GreenhouseGas Inventories (2006) [Internet]. Available from:https://www.ipcc-nggip.iges.or.jp/public/2006gl/.

33. Ziegler AD, Phelps J, Yuen JQ, et al. Carbon outcomesof major land-cover transitions in SE Asia: Greatuncertainties and REDDþpolicy implications. Glob.Chang. Biol. 18(10), 3087–3099 (2012).

34. Agus F, Gunarso P, Sahardjo BH, Harris N, VanNoordwijk M, Killeen TJ. Historical Co 2 EmissionsFrom Land Use and Land Use Change From the OilPalm Industry in Indonesia, Malaysia and Papua NewGuinea. 65–88 (2010).

35. IPCC. 2006 IPCC Guidelines for National GreenhouseGas Inventories [Internet]. Available from: https://www.ipcc-nggip.iges.or.jp/public/2006gl/.

36. Hooijer A, Page S, Jauhiainen J, et al. Subsidence andcarbon loss in drained tropical peatlands.Biogeosciences. 9(3), 1053–1071 (2012).

37. Page SE, Siegert F, Rieley JO, Boehm H-D V, Jaya A,Limin S. The amount of carbon released from peatand forest fires in Indonesia during 1997. Nature 420,61 (2002). [Internet]. Available from: http://dx.doi.org/10.1038/nature01131.

38. Moore S, Evans CD, Page SE, et al. Deep instability ofdeforested tropical peatlands revealed by fluvialorganic carbon fluxes. Nature 493(7434), 660–663(2013). [Internet]. Available from: http://dx.doi.org/10.1038/nature11818.

39. See LM, Fritz S. A method to compare and improveland cover Datasets: Application to the GLC-2000 andMODIS land cover products. IEEE Trans. Geosci.Remote Sens. 44(7), 1740–1746 (2006).

40. Giri C, Zhu Z, Reed B. A comparative analysis of theGlobal Land Cover 2000 and MODIS land cover datasets. Remote Sens. Environ. 94(1), 123–132 (2005).

41. DeVisser MH, Messina JP. Optimum land cover prod-ucts for use in a Glossina-morsitans habitat model ofKenya. Int. J. Health Geogr. 8(1), 39 (2009). [Internet].Available from: http://ij-healthgeographics.biomed-central.com/articles/10.1186/1476-072X-8-39.

42. Townshend J, Justice C, Li W, Gurney C, McManus J.Global land cover classification by remote sensing:present capabilities and future possibilities. RemoteSens. Environ. 35(2–3), 243–255 (1991).

43. Salomonson VV, Barnes W, Xiong J, Kempler S,Masuoka E. An overview of the Earth ObservingSystem MODIS instrument and associated data sys-tems performance. IEEE Int. Geosci. Remote Sens.Symp. 2(C), 1174–1176 (2002).

44. Roy DP, Wulder MA, Loveland TR, et al. Landsat-8:Science and product vision for terrestrial globalchange research. Remote Sens. Environ. 145, 154–172(2014).

45. Whiteside T. A Comparison of Object-Oriented andPixel-Based Classification Methods for Mapping LandCover. (September), 1225–1231 (2005).


http://www.rspo.org/file/GHGWG2/3_review_of_emission_factors_Agus_et_al.pdfhttp://www.rspo.org/file/GHGWG2/3_review_of_emission_factors_Agus_et_al.pdfhttp://www.rspo.org/file/GHGWG2/3_review_of_emission_factors_Agus_et_al.pdfhttp://www.fao.org/faostat/en/?#data/QDhttp://www.fao.org/faostat/en/?#data/QDhttp://stacks.iop.org/1748-9326/12/i=2/a=024014?key=crossref.d5803ab4c180cc18647fc14663f0d775http://stacks.iop.org/1748-9326/12/i=2/a=024014?key=crossref.d5803ab4c180cc18647fc14663f0d775http://stacks.iop.org/1748-9326/12/i=2/a=024014?key=crossref.d5803ab4c180cc18647fc14663f0d775http://www.nature.com/doifinder/10.1038/nclimate1702http://www.nature.com/doifinder/10.1038/nclimate1702http://www.pnas.org/cgi/doi/10.1073/pnas.1018776108http://www.pnas.org/cgi/doi/10.1073/pnas.1018776108http://www.globalforestwatch.orghttp://maps.elie.ucl.ac.be/CCI/viewer/download/ESACCI-LC-Ph2-PUGv2_2.0.pdfhttp://maps.elie.ucl.ac.be/CCI/viewer/download/ESACCI-LC-Ph2-PUGv2_2.0.pdfhttp://maps.elie.ucl.ac.be/CCI/viewer/download/ESACCI-LC-Ph2-PUGv2_2.0.pdfhttp://www.sciencemag.org/cgi/doi/10.1126/science.1244693http://www.sciencemag.org/cgi/doi/10.1126/science.1244693http://dx.doi.org/10.1016/j.agee.2015.06.009http://dx.doi.org/10.1016/j.agee.2015.06.009https://www.ipcc-nggip.iges.or.jp/public/2006gl/https://www.ipcc-nggip.iges.or.jp/public/2006gl/https://www.ipcc-nggip.iges.or.jp/public/2006gl/http://dx.doi.org/10.1038/nature01131http://dx.doi.org/10.1038/nature01131http://dx.doi.org/10.1038/nature11818http://dx.doi.org/10.1038/nature11818http://ij-healthgeographics.biomedcentral.com/articles/10.1186/1476-072X-8-39http://ij-healthgeographics.biomedcentral.com/articles/10.1186/1476-072X-8-39

46. Myint SW, Gober P, Brazel A, Grossman-Clarke S,

Weng Q. Per-pixel vs. object-based classification of

urban land cover extraction using high spatial reso-

lution imagery. Remote Sens. Environ. 115(5),

1145–1161 (2011).47. Foody GM. Status of land cover classification accur-

acy assessment. Remote Sens. Environ. 80(1), 185–201

(2002).48. Arnot C, Fisher P. Mapping the ecotone with fuzzy

sets. In: NATO Security through Science Series C:

Environmental Security. A Morris, S Kokhan (Eds.)

Springer, Dordrecht, 19–32 (2007).49. Davis SJ, Burney JA, Pongratz J, Caldeira K. Methods

for attributing land-use emissions to products.

Carbon Manag. 5(2), 233–245 (2014).50. The British Standards Institution; PAS 2050:2011;

Specification for the assessment of the life

cycle greenhouse gas emissions of goods andservices.

51. GanLian T, Cai H. Calculating GHG emission in oilpalm using PalmGHG. Planter. 93(1092), 167–176(2017).

52. Page SE, Banks CJ, Rieley JO. Tropical peatlands: distri-bution, extent and carbon storage – Uncertainties andknowledge gaps. Peatlands Int. 2(2), 26–27 (2007).

53. Chaplin-Kramer R, Ramler I, Sharp R, et al.Degradation in carbon stocks near tropical forestedges. Nat. Commun. 6, 10158 (2015).

54. Brinck K, Fischer R, Groeneveld J, et al. High reso-lution analysis of tropical forest fragmentation and itsimpact on the global carbon cycle. Nat. Commun. 8,14855 (2017).

55. Di Gregorio A, Henry M, Donegan E, et al. ClassificationConcepts Land Cover Classification System Softwareversion 3. Available from: www.fao.org/.


http://www.fao.org/

AbstractIntroductionLULC mappingGHG emissions attributable to LULC changeSignificance of peat soilsAim of this paper

Materials and methodsStudy areaLULC data sourcesData pre-processingCross-comparison of LULC dataLULC change analysisCarbon emission modelling

Results and discussionCross-comparison LULC data (Mapcurves)LULC changeGHG emissions

GHG emission mapsCarbon emission factorsTemporal intervalLimitations in estimating GHG emissions

Conclusions and recommendationsAcknowledgementsDisclosure statementReferences

the challenges of using satellite data sets to assess...

Documents