This article was downloaded by: [George Mason University] On: 28 March 2014, At: 12:17 Publisher: Taylor

& Francis Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Digital Earth

Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/tjde20

Integrating global land cover products for improved forest cover characterization: an application in North America
Xiao-Peng Song , Chengquan Huang , Min Feng , Joseph O. Sexton , Saurabh Channan & John R. Townshend
a a a a a a a

Global Land Cover Facility, Department of Geographical Sciences, University of Maryland, College Park, MD, USA Accepted author version posted online: 22 Oct 2013.Published online: 02 Dec 2013.

To cite this article: Xiao-Peng Song, Chengquan Huang, Min Feng, Joseph O. Sexton, Saurabh Channan & John R. Townshend (2013): Integrating global land cover products for improved forest cover characterization: an application in North America, International Journal of Digital Earth, DOI: 10.1080/17538947.2013.856959 To link to this article: http://dx.doi.org/10.1080/17538947.2013.856959

PLEASE SCROLL DOWN FOR ARTICLE Taylor & Francis makes every effort to ensure the accuracy of all the information (the Content) contained in the publications on our platform. Taylor & Francis, our agents, and our licensors make no representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Versions of published Taylor & Francis and Routledge Open articles and Taylor & Francis and Routledge Open Select articles posted to institutional or subject repositories or any other third-party website are without warranty from Taylor & Francis of any kind, either expressed or implied, including, but not limited to, warranties of merchantability, fitness for a particular purpose, or non-infringement. Any opinions and views expressed in this article are the opinions and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and should be independently verified with primary sources of information. Taylor & Francis shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://www.tandfonline.com/page/termsand-conditions Taylor & Francis and Routledge Open articles are normally published under a Creative Commons Attribution License http://creativecommons.org/licenses/by/3.0/. However, authors may opt to publish under a Creative Commons Attribution-Non-Commercial License http://creativecommons.org/licenses/by-nc/3.0/ Taylor & Francis and Routledge Open Select articles are currently published under a license to publish, which is based upon the Creative Commons Attribution-Non-Commercial No-Derivatives License, but allows for text and data mining of work. Authors also have the option of publishing an Open Select article under the Creative Commons Attribution License http:// creativecommons.org/licenses/by/3.0/. Downloaded by [George Mason University] at 12:17 28 March 2014 It is essential that you check the license status of any given Open and Open Select article to confirm conditions of access and use.

International Journal of Digital Earth, 2013 http://dx.doi.org/10.1080/17538947.2013.856959

Integrating global land cover products for improved forest cover characterization: an application in North America
Xiao-Peng Song*, Chengquan Huang, Min Feng, Joseph O. Sexton, Saurabh Channan and John R. Townshend
Global Land Cover Facility, Department of Geographical Sciences, University of Maryland, College Park, MD, USA

Downloaded by [George Mason University] at 12:17 28 March 2014

(Received 18 June 2013; accepted 8 October 2013) Six widely used coarse-resolution global land cover data-sets Global Land Cover Characterization (GLCC), Global Land Cover 2000 (GLC2000), GlobCover land cover product (GlobCover), MODIS land cover product (MODIS LC), the University of Maryland land cover product (UMD LC), and the MODIS Vegetation Continuous Fields tree cover layer (MODIS VCF) disagree substantially in their estimates of forest cover. Employing a regression tree model trained on higher-resolution, Landsat-based data, these multisource multiresolution maps were integrated for an improved characterization of forest cover over North America. Evaluated using a withheld test sample, the integrated percent forest cover (IPFC) data-set has a root mean square error of 11.75% substantially better than the 17.37% of GLCC, 17.61% of GLC2000, 17.96% of GlobCover, 15.23% of MODIS LC, 19.25% of MODIS VCF, and 15.15% of UMD LC, respectively. Although demonstrated for forest, this approach based on integration of multiple products has potential for improved characterization of other land cover types as well. Keywords: land cover; forest; data fusion; regression tree; North America

1. Introduction Satellite images have been used to characterize global patterns of land cover and land use since the mid-1990s. Developed using different data-sets and different methodologies, six global maps are now freely available at 300-m to 1-km resolutions: Global Land Cover Characterization (GLCC) (Loveland et al. 2000), Global Land Cover 2000 (GLC2000) (Bartholom and Belward 2005), GlobCover land cover product (GlobCover) (Bicheron et al. 2008), the Moderate Resolution Imaging Spectroradiometer land cover product (MODIS land cover product [MODIS LC]) (Friedl et al. 2002), the University of Maryland land cover product (UMD LC) (Hansen et al. 2000), and the MODIS Vegetation Continuous Fields product (MODIS VCF) (Hansen et al. 2003). A number of finer (e.g. Landsat-class) resolution land cover maps have also been generated from satellite data at regional to national scales (Hansen and Loveland, 2012), such as the North America Forest Disturbance product (Masek et al. 2008), the South

*Corresponding author. Email: xpsong@umd.edu

2013 The Author(s). Published by Taylor & Francis. This is an Open Access article. Non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly attributed, cited, and is not altered, transformed, or built upon in any way, is permitted. The moral rights of the named author(s) have been asserted.

X.-P. Song et al.

Dakota State University Congo/Indonesia/European Russia forest cover and loss maps (Broich et al. 2011; Potapov, Turubanova, and Hansen 2011; Potapov et al. 2012); the United States Department of Agriculture Cropland Data Layer (USDA 2013); the Brazilian PRODES (Amazon Deforestation Monitoring Project) deforestation maps (INPE 2013), as well as the United States National Land Cover Dataset (NLCD) (Homer et al. 2004; Xian, Homer, and Fry 2009). Often derived with greater local expertise than is possible with a global product, these regional Landsat-resolution data-sets provide a more accurate representation of the land surface type with greater spatial and thematic details. Some of these regional products have been well validated with finer-resolution imagery or field data, e.g. NLCD has an overall accuracy of 85.3% (Wickham et al. 2010). The accuracy of these regional data-sets has enabled their use as reference data for training or validating coarse-resolution products (Friedl et al. 2002; Hansen et al. 2003; Strahler et al. 2006). Previous studies comparing sets of two or three of the existing global products found general, yet variable agreement among them (e.g. Defries and Townshend 1994; Fritz and See 2008; Giri, Zhu, and Reed 2005; Hansen and Reed 2000; Herold et al. 2008; Jung et al. 2006; Pflugmacher et al. 2011; Song et al. 2011). High agreement tends to be located in relatively homogeneous and spectrally distinctive regions, whereas low agreement tends to be located in heterogeneous landscapes, land cover transition zones and in spectrally similar classes. Spatial patterns of agreement and disagreement revealed by cross-product comparison are useful for applications of these products; but ultimately, more accurate data-sets are always needed to meet users specific research demands. Yet, few attempts have been made to synthesize these data-sets into a consistent map. Jung et al. (2006) collected multiple versions of GLC2000, GLCC, and MODIS LC and merged them into a joint 1-km global land cover map by cross-walking different land cover legends. Fritz et al. (2011) created a hybrid cropland map for Africa using an approach modified from Jung et al.s (2006) synergy method by introducing an expert ranking step and a calibration step with national cropland statistics. Similarly, Schepaschenko et al. (2011) produced a hybrid land cover data-set over Russia by combing satellite-derived land cover maps, a Geographic Information System database and national statistics based on a set of knowledge rules. Here we propose a supervised, harmonization-based method for integrating multiresolution, multisource global products to improve land cover characterization. To demonstrate this approach, we use the six global maps above-mentioned as input and one Landsat-based map as reference to derive an integrated percent forest cover (IPFC) map over North America. Our long-term goal is to generate an improved forest cover map at the global scale to provide better parameterization for simulating land process in Earth System models (Lawrence and Chase 2007). Although we focus on the representation of forests in North America where we have sufficient reference data available for assessment of our results, the approach derived through this study can potentially be applied to any large area and any other land cover type. 2. Data We chose six widely used and available global data-sets to integrate. Each of these datasets was derived based on moderate-resolution satellite imagery (Table 1). The GLCC and UMD LC were developed using a global Advanced Very High Resolution Radiometer (AVHRR) data-set (Townshend et al. 1994). The MODIS LC and MODIS VCF were

Downloaded by [George Mason University] at 12:17 28 March 2014

International Journal of Digital Earth

Table 1. Summary of land cover products used in this study. Product GLCC GLC2000 GlobCover MODIS LC MODIS VCF UMD LC NLCD2001 Sensor AVHRR SPOT-4 MERIS MODIS MODIS AVHRR Landsat 5&7 Date April 1992 March 1993 November 1999 December 2000 December 2004 June 2006 October 2000 October 2001 October 2000 December 2001 April 1992 March 1993 Circa 2001 Resolution 1-km 1-km 300-m 1-km 500-m 1-km 30-m Classification approach Clustering Labeling Depends on individual region Supervised and unsupervised Decision tree Regression tree Decision tree Decision tree Publication

Downloaded by [George Mason University] at 12:17 28 March 2014

Loveland et al. (2000) Bartholom and Belward (2005) Bicheron et al. (2008) Friedl et al. (2002) Hansen et al. (2003) Hansen et al. (2000) Homer et al. (2004)

derived using data from MODIS. Data from the Satellite Pour lObservation de la Terre (SPOT-4) and Medium Resolution Imaging Spectrometer (MERIS) sensors were used to produce the GLC2000 and GlobCover, respectively. It should be noted that in some cases, e.g. MODIS-derived products, multiple versions have been generated following algorithmic improvements (Friedl et al. 2002, 2010; Hansen et al. 2003, 2005). The selection of the six global data-sets for this study takes into account several specific considerations: UMD LC and GLCC were the first 1-km maps generated; both MODIS LC and MODIS VCF are standard MODIS land products; MODIS VCF is a fractional land cover map; and GlobCover has the finest spatial resolution at the global scale. Collectively, these six maps represent land cover data-sets in different legends, derived from different sensor systems and with diverse classification approaches (Table 1). We use the Landsat-based National Land Cover Database 2001 (NLCD2001) over the conterminous United States as a reference for model training and product evaluation (Homer et al. 2004). Validated against expert-interpreted 1-m Digital Orthophoto Quarter Quadrangles, the NLCD2001 has 87% users accuracy and 88.5% producers accuracy for the forest class (Wickham et al. 2010). Further details of NLCD2001 are provided in Table 1 and associated references.

3. Method The data integration method consists of a series of steps (Figure 1) described in each of the sections below. As a preprocessing step, all products were reprojected to Lambert Azimuthal Equal Area projection with the WGS84 datum. They were also registered to a common spatial extent using nearest neighbor resampling. 3.1. Standardization to a common spatial and thematic scale These various data-sets need to be standardized to a common spatial and thematic scale prior to integration. We first defined a set of translation rules to convert each categorical

4
Global land-cover products

X.-P. Song et al.

Regional Landsat-based products

Standardization to a common spatial and thematic scale

Legend translation
Spatial aggregation

1-km categorical agreements

Legend translation
Spatial aggregation

Downloaded by [George Mason University] at 12:17 28 March 2014

5-km percentage forest/nonforest cover

5-km categorical agreements

5-km percentage forest/nonforest cover (reference)

Regression tree model

Integrated percent forest cover product

Figure 1. Flowchart for deriving the integrated 5-km percent forest cover product.

land cover class into proportional forest and nonforest cover and then spatially coarsened each product from its native scale to derive fractional coverage of forest and nonforest per 5-km grid. This resolution was chosen for consistency with the Climate Modeling Grid used for MODIS and AVHRR products (Pedelty et al. 2007). Aggregating categorical cover maps from 1- to 5-km also reduces spatial misregistration and yields a dynamic range sufficient for fractional cover. 3.1.1. Legend translation A set of translation rules was used to convert each maps categorical forest class into a forest canopy cover percentage based on its classification scheme (Table 2). All of the maps vegetation classes are defined based on woody and herbaceous canopy closure as well as vegetation height. For example, the International Geosphere-Biosphere Programme (IGBP) defined closed forest as an area with woody cover >60% and open forest as woody cover between 30% and 60%, both with tree height >2 m (Belward 1996). Binning land surface type into categorical groups at coarse spatial resolution unrealistically conceals the continuous nature of land surface properties. Moreover, aside from misclassification which is inevitable in any practical land cover map, a correctly classified pixel of closed forest in the IGBP scheme can still have forest canopy closure somewhere as low as 60%, with up to 40% of the remaining area being occupied by other land cover types. Therefore, to approximate the actual forest cover in a coarse grid and maintain statistical rigor, we assigned a proportional weight to each classified pixel

International Journal of Digital Earth

Table 2. Legend translation rules for harmonizing various land cover schemes into a common representation of forest cover. Land cover legend IGBP legend (GLCC, MODIS LC) Categorical class Forest (>60%) (evergreen needleleaf, deciduous needleleaf, everygreen broadleaf, everygreen needleaf, mixed) Woody savannas (3060%) Savannas (1030%) Cropland/natural vegetation mosaic Shrublands (closed, open), grasslands, permanent wetlands, urban and built-up, snow and ice, barren or sparsely vegetated, croplands, water bodies Forest (>60%) (evergreen needleleaf, deciduous needleleaf, everygreen broadleaf, everygreen needleaf, mixed) Woodland (4060%) Wooded grassland (1040%) Shrubland (closed, open), grassland, bare ground, urban and built, cropland, water Tree cover, closed (>40%) (evergreen broadleaved, deciduous broadleaved) Tree cover, open (1540%) deciduous broadleaved Tree cover (>15%) (everygreen needleleaf, deciduous needleleaf, mixed leaf type, regularly flooded fresh or saline) Mosaic: tree cover / other natural vegetation Mosaic: cropland/tree cover/other natural vegetation Burnt, shrub cover (evergreen, deciduous), herbaceous cover, sparse herbaceous or sparse shrub cover, regularly flooded shrub and or herbaceous cover, bare areas, artificial surfaces and associated areas, cultivated and managed areas, mosaic: cropland/shrub and/or herbaceous cover Closed forest (>40%) (broadleaved deciduous, needleleaved evergreen) Closed to open forest (>15%) (broadleaved evergreen or semideciduous forest, mixed broadleaved and needleleaved, broadleaved forest regularly flooded) Open (1540%) broadleaved deciduous forest/woodland, open (1540%) needleleaved deciduous or evergreen forest Mosaic forest or shrubland (5070%)/grassland (2050%) Mosaic grassland (5070%)/forest or shrubland (2050%) Mosaic vegetation (grassland/shrubland/forest) (5070%)/ cropland (2050%) Mosaic cropland (5070%)/vegetation (grassland/shrubland/ forest) (2050%) Closed to open shrubland, closed to open herbaceous vegetation, sparse vegetation, closed broadleaved forest or shrubland permanently flooded, closed to open grassland or woody vegetation on regularly flooded or waterlogged soil, artificial surfaces and associated areas, bare areas, permanent snow and ice, post-flooding or irrigated croplands, rainfed croplands, water bodies Forest canopy proportion (%) 80 45 20 25 0

Downloaded by [George Mason University] at 12:17 28 March 2014

UMD LC

80 50 25 0 70 27.5 57.5 50 25 0

GLC2000

GlobCover

70 57.5

27.5 30 17.5 20 11.7 0

Note: Nonforest proportion = 100% Forest canopy proportion. Land cover classes in each legend are grouped in the table according to forest canopy proportion in each respective class.

X.-P. Song et al.

corresponding to the mean value of its woody canopy closure as defined in its original legend, e.g. 80% to closed forest and 45% to open forest for the IGBP legend (Table 2). Classes such as closed and open shrublands, croplands, grasslands, permanent wetlands, urban and built-up, snow and ice, bare, and water bodies do not contain any forest cover. Thus, they were assigned a 0% forest cover and 100% nonforest cover. The mosaic classes in different land cover legends raise challenges in any legend harmonization work (Jung et al. 2006). The cropland/natural vegetation mosaic class in the IGBP legend contains a mixture of four classes including croplands, forests, shrublands, and grasslands (Belward 1996), so it was split into 25% forest cover and 75% nonforest cover. The complete legend translation rule set is given in Table 2. With these rules, each land cover product was converted to a percent forest cover map at its native resolution. MODIS VCF directly gives percent canopy cover for each pixel and hence no further translation is needed.

Downloaded by [George Mason University] at 12:17 28 March 2014

3.1.2. Spatial aggregation Each percent forest cover map was overlaid on the 5-km grid to calculate percent cover within each 5-km grid cell. For example, for the 1-km categorical GLCC, the aggregation was carried out by employing a 55 pixel window moving across the map. Within the local window, each classified pixel was first multiplied by its class-specific proportional weight defined by the legend translation rule, and then averaged to derive the proportional forest and nonforest cover within the 5-km grid. Other categorical maps were aggregated in the same way as GLCC. As MODIS VCF directly measures the percentage of forest canopy, we simply aggregated it from 500-m to 5-km with a 1010 local moving window by averaging the 100 pixel values within the window. 3.1.3. Deriving forest agreement metrics At the pixel level, it is reasonable to believe that a given land pixel is more likely to be forest if all six products independently classify the pixel as forest than if only one product identifies it as forest and the other five products label it as nonforest. Thus, different levels of agreement reflect varying degrees of certainty regarding the true forest cover in one pixel. In order to directly incorporate this agreement information into data integration, we calculated the pixel-based agreement metrics for the forest class. This analysis is based on categorical maps in parallel with the above legend translation and spatial aggregation process. As four of the six input products (i.e. GLCC, GLC2000, MODIS LC, and UMD LC) have an original resolution of 1-km, we first align all the six products at 1-km resolution to calculate a 1-km forest vote map and then derive the 5-km agreement metrics based upon the 1-km vote map. The 300-m GlobCover was resampled to 1-km resolution and the 500-m fractional MODIS VCF was spatially averaged to 1-km first and converted to binary forest and nonforest by applying a 30% threshold according to the IGBP definition. To assess the degree of correspondence between the six products at 1-km, we evaluated each pixel as the number of times it was labeled as forest by the six maps, resulting in a value between 0 and 6: the higher the value, the higher the agreement between the products for the forest class. The 5-km agreement metrics were derived by grouping the 1-km metrics using a 55 pixel moving window. Within the local window, the 1-km agreement pixels with values between 0 and 6 were noted and each 5-km grid was characterized by the frequencies of each of those values. Each 5-km grid then corresponds to the frequency histogram of 06 for forest agreement at 1-km resolution.

International Journal of Digital Earth

Downloaded by [George Mason University] at 12:17 28 March 2014

3.2. Supervized training and prediction A supervized regression tree algorithm was used to model the relationship between reference cover values from NLCD2001 and forest cover values as well as agreement metrics from the coarse data-sets. Tree-based classification and regression methods are well established in land cover characterization studies (e.g. Friedl et al. 2002; Hansen et al. 2003, 2000; Homer et al. 2004; Sexton et al. 2013a, 2013b; Xian and Crane 2005). Regression trees have the theoretical advantage of handling nonlinear relationships by recursively splitting the sample into binary partitions until criteria of accuracy or purity are met (Breiman et al. 1984). This algorithm produces a hierarchical set of decision rules, each of which terminates in a linear regression model. Predictor variables feeding into the regression tree model consist of the proportional forest and nonforest cover layers derived through legend translation and spatial aggregation. The seven agreement metrics layers are used in the conditional statements of the regression rules to parameterize the tree model. Reference data were derived by aggregating NLCD2001 from 30-m to 5-km resolution to calculate the percentage of forest pixels per 5-km grid. A total of 40,713 pixels (12% of land pixels) were systematically selected from the aggregated NLCD2001, from which half were randomly selected for model training and half for data validation. 3.3. Product evaluation Accuracies of the six input data-sets and the output IPFC data-set were evaluated against the aggregated NLCD2001 values using mean bias error (MBE), root mean square error (RMSE) and r2: Pn Pi Ri 1 MBE i1 n s Pn 2 i1 Pi Ri RMSE n Pn 2 1 Pi Ri r2 1 Pin 2 , i1 Ri R

where i is the sample index; Pi is the value of IPFC or forest cover of each input product; Ri is the reference forest cover per sample; R is the mean of reference; and n is the sample size (Willmott 1982). The test sample was further divided according to reference values into three subsets representing low, moderate, and high forest cover (i.e. 030%, 3160%, 61100%), respectively. Accuracy metrics were calculated using the entire test sample as well as these three subsets to report the disaggregate error by categories of percent forest cover. 4. Results 4.1. Model fitting and performance assessment Evaluation of the regression tree model using 20,357 training cases yielded an average error of 6.46% with a correlation coefficient between reference and predicted cover of 0.94. Internal 10-fold cross-validation on training data estimated a slightly higher average

X.-P. Song et al.

Downloaded by [George Mason University] at 12:17 28 March 2014

error of 6.78% and an identical correlation coefficient. Independent evaluation of the model using a withheld sample of 20,356 test cases estimated a further slightly higher average error of 6.91% and an identical correlation coefficient as using the training data, suggesting negligible model over-fitting with the training sample. The most important predictor variable in the model was forest cover derived from MODIS VCF, which was used in 65% of the conditional statements and in 69% of the terminal-node regression models (Table 3). All six products were heavily used in the terminal multivariate regressions, with forest cover from GLCC as the most frequently used variable and forest cover from MODIS LC as the least frequently used one. Relative to forest/nonforest cover variables, agreement metrics had moderate to low usage rates in the conditional statements. Agreement metric at level 0 was used in 32% of the conditional statements, while other agreement metrics had usage rates less than 10%. The low usage of agreement metrics at level 16 in the conditional statements is probably due to high usages of forest cover variables from MODIS VCF, MODIS LC, UMD LC, GlobCover as well as agreement metric at level 0, which, collectively, provide adequate information for splitting the regression tree. 4.2 The IPFC map The integrated product depicts well-known patterns of forest cover over North America, including contiguous expanses of forest in the eastern and Pacific north-western United States and boreal Canada (Figure 2). The IPFC also shows the climatological fragmentation of forests in the southern Rocky Mountains as well as the longitudinal gradients in ecotones between boreal forests and tundra. Tropical and subtropical dry forests in Southern Mexico are represented by moderate to high forest cover. Agricultural fields in the Great Plains are depicted with very low forest cover.
Table 3. Usage of predictor variables in regression tree model. Predictor variable MODIS VCF forest cover MODIS LC forest cover UMD LC forest cover GLC2000 forest cover Agreement level 0 GlobCover forest cover MODIS LC nonforest cover GLCC forest cover GlobCover nonforest cover Agreement level 6 Agreement level 1 UMD nonforest cover Agreement level 5 Agreement level 2 Agreement level 4 Agreement level 3 GLCC nonforest cover GLC2000 nonforest cover Conditional use rate 0.65 0.48 0.36 0.36 0.32 0.24 0.13 0.10 0.07 0.05 0.05 0.04 0.03 0.03 0.03 0.02 0.01 0.00 Terminal-node regression use rate 0.69 0.22 0.56 0.83 0.00 0.49 0.42 0.90 0.30 0.00 0.00 0.26 0.00 0.00 0.00 0.00 0.85 0.49

International Journal of Digital Earth

Downloaded by [George Mason University] at 12:17 28 March 2014

Figure 2. The IPFC map over North America.

A comparison of the six input data-sets and IPFC over the Chesapeake Bay region in Eastern United States highlights differences between the products in fragmented areas (Figure 3). This region is covered by a complex landscape including forested national parks, high-density urban lands of the Washington DC-Baltimore metropolitan area, suburban low-density residential lands, as well as agriculture fields in Maryland, Delaware, and Virginia. The fragmented and heterogeneous landscape raises great difficulties in land cover classifications with coarse-resolution satellite data and can lead to substantial disagreement between the existing products (Figure 3a). Each of the six products shows a varying degree of over or underestimation of forest cover as compared with NLCD2001, while the integrated product shows the closest visual similarity with the reference. 4.3. Comparison of integrated and original products Figure 4 compares forest cover derived from the six input data-sets as well as the newly integrated product against reference forest cover derived from NLCD2001. Combining the six products and calibrated by NLCD2001, the newly integrated product has substantially improved forest cover estimate as reflected by the scatter plots a highest value of 100% forest cover and a concentrated and nearly symmetric 1:1 line. General

10

Downloaded by [George Mason University] at 12:17 28 March 2014

(a) Forest agreement

(b) NLCD2001

(c) IPFC

X.-P. Song et al.

(d) GLCC

(e) GLC2000

(f) GlobCover

(g) MODIS LC

(h) MODIS VCF

(i) UMD LC

Figure 3. Subsets of forest cover maps in the Chesapeake Bay region in eastern USA. (a) Forest agreement of the six input products at 1-km spatial resolution. (b) NLCD2001 reference at 5-km spatial resolution. (c) The integrated product at 5-km spatial resolution. (di) Forest cover derived from the six existing products at 5-km resolution.

International Journal of Digital Earth

11

Downloaded by [George Mason University] at 12:17 28 March 2014

Figure 4. Scatter plots of forest cover (%) at 5-km resolution derived from global land cover products against reference percent forest cover (n = 20,356). The x axis in each plot represents reference forest cover derived from NLCD2001 and the y axis represents each global product as well as the integrated product.

patterns of agreement between coarse-resolution products and reference were revealed across the conterminous United States. The highest values of fractional forest cover derived from categorical maps (i.e. GLCC, GLC2000, GlobCover, MODIS LC and UMD LC) all saturated at around 80% cover due to semantic uncertainty in hard classifications. Since the actual forest cover within a coarse pixel grid was unknown, a mean 80% cover value was assigned to the closed forest class (more details in Section 3.1.1, Table 2). MODIS VCF also saturated at 80% a consistent conclusion with previous VCF validation studies (Montesano et al. 2009; Sexton et al. 2013a; White, Shaw, and Ramsey 2005). Statistically, the IPFC estimates showed greater consistency with the reference data than all of the input data-sets, which varied in their accuracy relative to NLCD2001 (Table 4). All six input data-sets showed fairly high r2 globally (between 0.7 and 0.8) against the reference data, but IPFC achieved a considerably higher r2 of 0.87. MODIS VCF consistently underestimates forest cover in every stratum, whereas GLCC, GLC2000, GlobCover, MODIS LC, and UMD LC consistently overestimate forest cover in the lowest stratum but underestimate in moderate and high strata. IPFC has a slightly greater bias than MODIS LC and UMD LC in the lowest stratum but has the lowest bias

12

X.-P. Song et al.

Table 4. Evaluation of the IPFC product and the six global land cover products for forest cover characterization. MBE (%) Product GLCC GLC2000 GlobCover MODIS LC MODIS VCF UMD LC IPFC 030 3.56 1.74 3.95 0.32 1.76 0.55 1.26 3160 2.92 5.98 8.51 8.92 16.88 6.73 0.73 61100 9.64 14.75 23.13 12.59 27.88 13.92 5.53 0100 0.59 3.38 4.42 4.20 10.33 4.02 0.66 030 13.45 14.41 11.81 10.28 7.55 8.97 8.78 RMSE (%) 3160 24.28 23.57 18.98 21.00 22.82 20.22 16.10 61100 20.34 20.10 27.70 20.32 32.77 22.12 14.53 0100 17.37 17.61 17.96 15.23 19.25 15.15 11.75 r2 0100 0.72 0.72 0.72 0.79 0.77 0.79 0.87

Downloaded by [George Mason University] at 12:17 28 March 2014

Note: The test sample was stratified into three subsets representing low moderate and high forest cover (i.e. 030%, 3160%, 61100%), respectively. Whereas r2 was calculated using the entire test sample (0100%), MBE and RMSE were calculated using these subsets as well as the entire sample.

among all the products in moderate and high strata. The RMSE was twice as great in the moderate (3160%) and high (61100%) ranges than in the low (030%) stratum for every product, implying great uncertainties in characterizing medium- to high-density forests using coarse-resolution satellite data. At the low end of forest cover, the integrated data-set was slightly less accurate than MODIS VCF, but at moderate to high cover range, IPFC was considerably more accurate than any of the six input data-sets. The overall RMSE of the integrated product against reference was 11.75%, significantly lower than a 17.37% of GLCC, 17.61% of GLC2000, 17.96% of GlobCover, 15.23% of MODIS LC, 19.25% of MODIS VCF, and 15.15% of UMD LC, respectively.

5. Discussion Differences in the existing global land cover products stem from multiple sources, including different legends, satellite sensor systems, acquisition dates, classification algorithms, as well as image misregistration (e.g. Defries and Townshend 1994; Fritz and See 2008; Giri, Zhu, and Reed 2005; Hansen and Reed 2000; Herold et al. 2008; Jung et al. 2006; Pflugmacher et al. 2011; Song et al. 2011; Townshend et al. 1992; Townshend et al. 1994). While this variation contributes uncertainty to inferences based on any individual map, the central tendency of multiple data-sets reduces the impacts on an integrated product. Major semantic differences in forest definition can be minimized by weighting categorical classes to approximate the actual forest canopy cover in coarse pixels. Also, spatial aggregation has a positive impact on both positional and thematic accuracy, as the ratio of pixels positional error to pixel size declines through aggregation, and random misclassification errors between neighboring pixels cancel one another when products are coarsened spatially (Marceau, Howarth, and Gratton 1994; Moody and Woodcock 1995). Thus, spatial aggregation is a key precursor to increase comparability between different land cover products. Third, conditioning the integration of multiple data-sets on their agreement enhances information convergence, since there is a greater probability that the maps represent the actual land surface type in places where all independently generated products agree than where they disagree. Lastly, supervised training and classification approach rely heavily on the quality of reference data. Thus,

International Journal of Digital Earth

13

collecting accurate and representative training data is critically important in generating a land cover product. Future high-quality Landsat-resolution data-sets such as Food and Agriculture Organizations remote sensing survey Landsat sampling blocks (Potapov et al. 2011) would be candidate reference sources to the integration approach proposed here. 6. Summary Global land cover products show substantial discrepancies in their representation of land surface type, including forests. We present a data fusion method to integrate existing multiresolution (e.g. 300-m to 1-km) multisource global land cover maps to derive a new hybrid product for the forest class, and we demonstrate the approach over North America. Different from previous data fusion methodologies by Jung et al. (2006) and Fritz et al. (2011), which mainly rely on agreement between different land cover products, our approach also uses a large sample of higher-resolution land cover data as reference to integrate coarser data-sets in a supervised modeling framework. Compatible with previous work, land cover characterization is greatly improved by combing various sources of existing data-sets. Assessment of errors relative to a withheld test sample suggests that the integrated forest map has an overestimation in low forest cover stratum (i.e. 030%) and a slight underestimation in moderate (i.e. 3160%) to high forest cover strata (i.e. 61100%). Compared to the existing individual global maps of forest cover, a considerable improvement is achieved through data integration with an overall RMSE of 11.75% against Landsat reference and, the greatest improvements are achieved in moderate to high forest cover regions. This data-set is freely available for download at the Global Land Cover Facility (www.landcover.org). Acknowledgments
We thank the anonymous reviewers for their valuable comments on the paper.

Downloaded by [George Mason University] at 12:17 28 March 2014

Funding
This study is a contribution to the Global Forest Cover Change project funded by NASAs Making Earth System Data Records for Use in Research Environments (MEaSUREs) Program [NNX08AP33A]. Additional support is provided by the NASA Earth and Space Science Fellowship (NESSF) Program [NNX12AN92H]; the Land-Cover/Land-Use Change Program [NNH07ZDA001N]; the Earth System Science from EOS Program [NNH06ZDA001N]; and the MODIS Science Team.

References
Bartholom, E., and A. S. Belward. 2005. GLC2000: A New Approach to Global Land Cover Mapping from Earth Observation Data. International Journal of Remote Sensing 26 (9): 1959 1977. doi:10.1080/01431160412331291297. Belward, A. S. 1996. The IGBP-DIS Global 1 km Land Cover Data Set DISCover: Proposal and Implementation Plans. Report of the Land Cover Working Group of IGBP-DIS. Toulouse: IGBPDIS Office. Bicheron, P., P. Defourny, C. Brockmann, L. Schouten, C. Vancutsem, M. Huc, S. Bontemps, et al. 2008. GlobCover: Products Description and Validation Report. Toulouse: Medias France. Breiman, L., J. H. Friedman, R. A. Olshen, and C. J. Stone. 1984. Classification and Regression Trees. Boca Raton, FL: Chapman & Hall/CRC. Broich, M., M. C. Hansen, P. Potapov, B. Adusei, E. Lindquist, and S. V. Stehman. 2011. Timeseries Analysis of Multi-resolution Optical Imagery for Quantifying Forest Cover Loss in

14

X.-P. Song et al.

Sumatra and Kalimantan, Indonesia. International Journal of Applied Earth Observation and Geoinformation 13 (2): 277291. doi:10.1016/j.jag.2010.11.004. Defries, R. S., and J. R. G. Townshend. 1994. NDVI-Derived Land-Cover Classifications at a Global-Scale. International Journal of Remote Sensing 15 (17): 35673586. doi:10.1080/0143 1169408954345. Friedl, M. A., D. K. McIver, J. C. F. Hodges, X. Y. Zhang, D. Muchoney, A. H. Strahler, C. E. Woodcock, et al. 2002. Global Land Cover Mapping from MODIS: Algorithms and Early Results. Remote Sensing of Environment 83 (12): 287302. doi:10.1016/S0034-4257(02) 00078-0. Friedl, M. A., D. Sulla-Menashe, B. Tan, A. Schneider, N. Ramankutty, A. Sibley, and X. Huang. 2010. MODIS Collection 5 Global Land Cover: Algorithm Refinements and Characterization of New Datasets. Remote Sensing of Environment 114 (1): 168182. doi:10.1016/j. rse.2009.08.016. Fritz, S., and L. See. 2008. Identifying and Quantifying Uncertainty and Spatial Disagreement in the Comparison of Global Land Cover for Different Applications. Global Change Biology 14 (5): 10571075. doi:10.1111/j.1365-2486.2007.01519.x. Fritz, S., L. You, A. Bun, L. See, I. McCallum, C. Schill, C. Perger, J. Liu, M. Hansen, and M. Obersteiner. 2011. Cropland for Sub-Saharan Africa: A Synergistic Approach Using Five Land Cover Data Sets. Geophysical Research Letters 38 (4): L04404. doi:10.1029/2010GL046213. Giri, C., Z. Zhu, and B. Reed. 2005. A Comparative Analysis of the Global Land Cover 2000 and MODIS Land Cover Data Sets. Remote Sensing of Environment 94 (1): 123132. doi:10.1016/j. rse.2004.09.005. Hansen, M. C., and T. R. Loveland. 2012. A Review of Large Area Monitoring of Land Cover Change Using Landsat Data. Remote Sensing of Environment 122: 6666. doi:10.1016/j. rse.2011.08.024. Hansen, M. C., and B. Reed. 2000. A Comparison of the IGBP DISCover and University of Maryland 1 Km Global Land Cover Products. International Journal of Remote Sensing 21 (67): 13651373. doi:10.1080/014311600210218. Hansen, M. C., R. S. DeFries, J. R. G. Townshend, M. Carroll, C. Dimiceli, and R. A. Sohlberg 2003. Global Percent Tree Cover at a Spatial Resolution of 500 Meters: First Results of the MODIS Vegetation Continuous Fields Algorithm. Earth Interactions 7 (10): 115. doi:10.1175/ 1087-3562(2003)007<0001:GPTCAA>2.0.CO;2. Hansen, M. C., R. S. DeFries, J. R. G. Townshend, and R. Sohlberg. 2000. Global Land Cover Classification at 1 km Spatial Resolution Using a Classification Tree Approach. International Journal of Remote Sensing 21 (67): 13311364. doi:10.1080/014311600210209. Hansen, M. C., J. R. G. Townshend, R. S. DeFries, and M. Carroll. 2005. Estimation of Tree Cover Using MODIS Data at Global, Continental and Regional/Local Scales. International Journal of Remote Sensing 26 (19): 43594380. doi:10.1080/01431160500113435. Herold, M., P. Mayaux, C. E. Woodcock, A. Baccini, and C. Schmullius. 2008. Some Challenges in Global Land Cover Mapping: An Assessment of Agreement and Accuracy in Existing 1 Km Datasets. Remote Sensing of Environment 112 (5): 25382556. doi:10.1016/j.rse.2007.11.013. Homer, C., C. Huang, L. Yang, B. Wylie, and M. Coan. 2004. Development of a 2001 National Land-Cover Database for the United States. Photogrammetric Engineering & Remote Sensing 70: 829840. INPE. 2013. Prodes Digital. Accessed August 23. http://www.obt.inpe.br/prodes/index.php Jung, M., K. Henkel, M. Herold, and G. Churkina. 2006. Exploiting Synergies of Global Land Cover Products for Carbon Cycle Modeling. Remote Sensing of Environment 101 (4): 534553. doi:10.1016/j.rse.2006.01.020. Lawrence, P. J., and T. N. Chase. 2007. Representing a New MODIS Consistent Land Surface in the Community Land Model (CLM 3.0). Journal of Geophysical Research 112: 117. Loveland, T. R., B. C. Reed, J. F. Brown, D. O. Ohlen, Z. Zhu, L. Yang, and J. W. Merchant. 2000. Development of a Global Land Cover Characteristics Database and IGBP DISCover from 1 Km AVHRR Data. International Journal of Remote Sensing 21 (67): 13031330. doi:10.1080/ 014311600210191. Marceau, D. J., P. J. Howarth, and D. J. Gratton. 1994. Remote Sensing and the Measurement of Geographical Entities in a Forested Environment. 1. The Scale and Spatial Aggregation Problem. Remote Sensing of Environment 49 (2): 93104. doi:10.1016/0034-4257(94)90046-9.

Downloaded by [George Mason University] at 12:17 28 March 2014

International Journal of Digital Earth

15

Masek, J. G., C. Huang, R. Wolfe, W. Cohen, F. Hall, J. Kutler, and P. Nelson. 2008. North American Forest Disturbance Mapped from a Decadal Landsat Record. Remote Sensing of Environment 112 (6): 29142926. doi:10.1016/j.rse.2008.02.010. Montesano, P. M., R. Nelson, G. Sun, H. Margolis, A. Kerber, and K. J. Ranson. 2009. MODIS Tree Cover Validation for the Circumpolar TaigaTundra Transition Zone. Remote Sensing of Environment 113 (10): 21302141. doi:10.1016/j.rse.2009.05.021. Moody, A., and C. E. Woodcock. 1995. The Influence of Scale and the Spatial Characteristics of Landscapes on Land-Cover Mapping Using Remote Sensing. Landscape Ecology 10 (6): 363379. doi:10.1007/BF00130213. Pedelty, J., S. Devadiga, E. Masuoka, M. Brown, J. Pinzon, C. Tucker, D. J. J. Roy, et al. 2007. Generating a Long-term Land Data Record from the AVHRR and MODIS Instruments. In Proceedings of the Geoscience and Remote Sensing Symposium (IGARSS), 10211025. Barcelona: IEEE International. Pflugmacher, D., O. N. Krankina, W. B. Cohen, M. A. Friedl, D. Sulla-Menashe, R. E. Kennedy, P. Nelson, et al. 2011. Comparison and Assessment of Coarse Resolution Land Cover Maps for Northern Eurasia. Remote Sensing of Environment 115 (12): 35393553. doi:10.1016/j.rse.2011. 08.016. Potapov, P., M. C. Hansen, A. M. Gerrand, E. J. Lindquist, K. Pittman, S. Turubanova, and M. Lyche Wilkie. 2011. The Global Landsat Imagery Database for the FAO FRA Remote Sensing Survey. International Journal of Digital Earth 4 (1): 221. doi:10.1080/17538947. 2010.492244. Potapov, P., S. Turubanova, and M. C. Hansen. 2011. Regional-scale Boreal Forest Cover and Change Mapping Using Landsat Data Composites for European Russia. Remote Sensing of Environment 115 (2): 548561. doi:10.1016/j.rse.2010.10.001. Potapov, P. V., S. A. Turubanova, M. C. Hansen, B. Adusei, M. Broich, A. Altstatt, L. Mane, and C. O. Justice. 2012. Quantifying Forest Cover Loss in Democratic Republic of the Congo, 20002010, with Landsat ETM+ data. Remote Sensing of Environment 122: 106116. doi:10.1016/j.rse.2011.08.027. Schepaschenko, D., I. McCallum, A. Shvidenko, S. Fritz, F. Kraxner, and M. Obersteiner. 2011. A New Hybrid Land Cover Dataset for Russia: A Methodology for Integrating Statistics, Remote Sensing and in Situ Information. Journal of Land Use Science 6 (4): 245259. doi:10.1080/ 1747423X.2010.511681. Sexton, J. O., X.-P. Song, M. Feng, P. Noojipady, A. Anand, C. Huang, D. Kim, et al. 2013a. Global, 30-m Resolution Continuous Fields of Tree Cover: Landsat-based Rescaling of MODIS Vegetation Continuous Fields with Lidar-based Estimates of Error. International Journal of Digital Earth 6 (5): 427448. Sexton, J. O., X.-P. Song, C. Huang, S. Channan, M. E. Baker, and J. R. Townshend. 2013b. Urban Growth of the Washington, DCBaltimore, MD Metropolitan Region from 1984 to 2010 by Annual, Landsat-based Estimates of Impervious Cover. Remote Sensing of Environment 129: 4253. doi:10.1016/j.rse.2012.10.025. Song, X.-P., C. Huang, J. O. Sexton, M. Feng, R. Narasimhan, S. Channan, and J. R. Townshend. 2011. An Assessment of Global Forest Cover Maps Using Regional Higher-resolution Reference Data Sets. In Proceedings of the Geoscience and Remote Sensing Symposium (IGARSS), 752755. Vancouver: IEEE International. Strahler, A. H., L. Boschetti, G. M. Foody, M. A. Friedl, M. C. Hansen, M. Herold, P. Mayaux, J. T. Morisette, S. V. Stehman, and C. E. Woodcock. 2006. Global Land Cover Validation: Recommendations for Evaluation and Accuracy Assessment of Global Land Cover Maps. Luxembourg: Office for Official Publications of the European Communities. Townshend, J. R. G., C. O. Justice, C. Gurney, and J. McManus. 1992. The Impact of Misregistration on Change Detection. IEEE Transactions on Geoscience and Remote Sensing 30 (5): 10541060. doi:10.1109/36.175340. Townshend, J. R. G., C. O. Justice, D. Skole, J.-P. Malingreau, J. Cihlar, P. Teillet, F. Sadowski, and S. Ruttenberg. 1994. The 1 km Resolution Global Data Set: Needs of the International Geosphere Biosphere Programme. International Journal of Remote Sensing 15 (17): 34173441. doi:10.1080/01431169408954338. USDA. 2013. USDA National Agricultural Statistics Services. Accessed August 23. http://www. nass.usda.gov/research/Cropland/SARS1a.htm

Downloaded by [George Mason University] at 12:17 28 March 2014

16

X.-P. Song et al.

White, M. A., J. D. Shaw, and R. D. Ramsey. 2005. Accuracy Assessment of the Vegetation Continuous Field Tree Cover Product Using 3954 Ground Plots in the South-Western USA. International Journal of Remote Sensing 26 (12): 26992704. doi:10.1080/01431160500080626. Wickham, J. D., S. V. Stehman, J. A. Fry, J. H. Smith, and C. G. Homer. 2010. Thematic Accuracy of the NLCD 2001 Land Cover for the Conterminous United States. Remote Sensing of Environment 114 (6): 12861296. doi:10.1016/j.rse.2010.01.018. Willmott, C. J. 1982. Some Comments on the Evaluation of Model Performance. Bulletin of the American Meteorological Society 63 (11): 13091313. doi:10.1175/1520-0477(1982)063<1309: SCOTEO>2.0.CO;2. Xian, G., and M. Crane. 2005. Assessments of Urban Growth in the Tampa Bay Watershed Using Remote Sensing Data. Remote Sensing of Environment 97 (2): 203215. doi:10.1016/j.rse.2005. 04.017. Xian, G., C. Homer, and J. Fry. 2009. Updating the 2001 National Land Cover Database Land Cover Classification to 2006 by Using Landsat Imagery Change Detection Methods. Remote Sensing of Environment 113 (6): L 11331147. doi:10.1016/j.rse.2009.02.004.

Downloaded by [George Mason University] at 12:17 28 March 2014