mapping land-cover modifications over large areas: a comparison of machine learning algorithms

12
Mapping land-cover modifications over large areas: A comparison of machine learning algorithms John Rogan b, , Janet Franklin a , Doug Stow b , Jennifer Miller c , Curtis Woodcock d , Dar Roberts e a Department of Biology, San Diego State University, San Diego CA 92182, United States b Department of Geography, San Diego State University, San Diego CA 92182, United States c Department of Geology and Geography and Geology, West Virginia University, Morgantown, WV 26506, United States d Department of Geography, Boston University, Boston, MA 02215, United States e Department of Geography, University of California Santa Barbara, Santa Barbara, CA 93106, United States Received 13 January 2006; received in revised form 7 October 2007; accepted 13 October 2007 Abstract Large area land-cover monitoring scenarios, involving large volumes of data, are becoming more prevalent in remote sensing applications. Thus, there is a pressing need for increased automation in the change mapping process. The objective of this research is to compare the performance of three machine learning algorithms (MLAs); two classification tree software routines (S-plus and C4.5) and an artificial neural network (ARTMAP), in the context of mapping land-cover modifications in northern and southern California study sites between 1990/91 and 1996. Comparisons were based on several criteria: overall accuracy, sensitivity to data set size and variation, and noise. ARTMAP produced the most accurate maps overall (84%), for two study areas in southern and northern California, and was most resistant to training data deficiencies. The change map generated using ARTMAP has similar accuracies to a human-interpreted map produced by the U.S. Forest Service in the southern study area. ARTMAP appears to be robust and accurate for automated, large area change monitoring as it performed equally well across the diverse study areas with minimal human intervention in the classification process. © 2007 Elsevier Inc. All rights reserved. Keywords: Land-cover change; Machine learning; Large area monitoring 1. Introduction The Holy Grail of (digital) change detection is still total automation and high accuracy.(Loveland et al., 2002, p. 1098). Over the coming decades, the global effects of land-cover/use change may be as significant, or more so, than those associated with potential climate change (IPCC, 2000). In spite of this there is a lack of comprehensive information on the types and rates of land-cover/use change, and even less evidence of natural and anthropogenic causes and consequences of such change (Turner et al., 1999). As a result, several large area land-cover monitoring programs have been established over the past five years to comprehensively address this issue (Wulder et al., 2004). Monitoring programs, unlike most research-oriented studies, employ change mapping methods that require processing and interpretation of large volumes of in situ, remotely sensed and ancillary data (Cilhar 2000; Franklin & Wulder, 2002). Very large data volumes and time-consuming data processing, integration and interpretation make automated and accurate methods of change mapping highly desirable (Aspinall, 2002; Dobson & Bright, 1994; Hansen et al., 2002; Rogan & Chen, 2004). Complex land change processes are of particular interest to researchers involved in large area monitoring (Roberts et al., 2002), where many different types of land-cover changes can occur and must be characterized (e.g., forest pest infestation, logging, wildfire, and suburbanization) (Rogan & Miller 2006). Thus, increased automation can ensure that the classification process is objective and repeatable in processing large volumes of data over complex and phenologically diverse landscapes (DeFries & Chan 2000; Gong & Xu 2003). Consequently, classification algorithm selection and performance have become Available online at www.sciencedirect.com Remote Sensing of Environment 112 (2008) 2272 2283 www.elsevier.com/locate/rse Corresponding author. Current address: Department of Geography, Clark University, 950 Main Street, Worcester, MA 01610, United States. Tel.: +1 508 793 7562. E-mail address: [email protected] (J. Rogan). 0034-4257/$ - see front matter © 2007 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2007.10.004

Upload: john-rogan

Post on 29-Jun-2016

216 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

Available online at www.sciencedirect.com

112 (2008) 2272–2283www.elsevier.com/locate/rse

Remote Sensing of Environment

Mapping land-cover modifications over large areas: A comparison ofmachine learning algorithms

John Rogan b,⁎, Janet Franklin a, Doug Stow b, Jennifer Miller c, Curtis Woodcock d, Dar Roberts e

a Department of Biology, San Diego State University, San Diego CA 92182, United Statesb Department of Geography, San Diego State University, San Diego CA 92182, United States

c Department of Geology and Geography and Geology, West Virginia University, Morgantown, WV 26506, United Statesd Department of Geography, Boston University, Boston, MA 02215, United States

e Department of Geography, University of California Santa Barbara, Santa Barbara, CA 93106, United States

Received 13 January 2006; received in revised form 7 October 2007; accepted 13 October 2007

Abstract

Large area land-cover monitoring scenarios, involving large volumes of data, are becoming more prevalent in remote sensing applications.Thus, there is a pressing need for increased automation in the change mapping process. The objective of this research is to compare theperformance of three machine learning algorithms (MLAs); two classification tree software routines (S-plus and C4.5) and an artificial neuralnetwork (ARTMAP), in the context of mapping land-cover modifications in northern and southern California study sites between 1990/91 and1996. Comparisons were based on several criteria: overall accuracy, sensitivity to data set size and variation, and noise. ARTMAP produced themost accurate maps overall (∼84%), for two study areas — in southern and northern California, and was most resistant to training datadeficiencies. The change map generated using ARTMAP has similar accuracies to a human-interpreted map produced by the U.S. Forest Servicein the southern study area. ARTMAP appears to be robust and accurate for automated, large area change monitoring as it performed equally wellacross the diverse study areas with minimal human intervention in the classification process.© 2007 Elsevier Inc. All rights reserved.

Keywords: Land-cover change; Machine learning; Large area monitoring

1. Introduction

“The Holy Grail of (digital) change detection is still totalautomation and high accuracy.” (Loveland et al., 2002, p. 1098).Over the coming decades, the global effects of land-cover/usechange may be as significant, or more so, than those associatedwith potential climate change (IPCC, 2000). In spite of this thereis a lack of comprehensive information on the types and rates ofland-cover/use change, and even less evidence of natural andanthropogenic causes and consequences of such change (Turneret al., 1999). As a result, several large area land-cover monitoringprograms have been established over the past five years tocomprehensively address this issue (Wulder et al., 2004).

⁎ Corresponding author. Current address: Department of Geography, ClarkUniversity, 950 Main Street, Worcester, MA 01610, United States. Tel.: +1 508793 7562.

E-mail address: [email protected] (J. Rogan).

0034-4257/$ - see front matter © 2007 Elsevier Inc. All rights reserved.doi:10.1016/j.rse.2007.10.004

Monitoring programs, unlike most research-oriented studies,employ change mapping methods that require processing andinterpretation of large volumes of in situ, remotely sensed andancillary data (Cilhar 2000; Franklin&Wulder, 2002). Very largedata volumes and time-consuming data processing, integrationand interpretation make automated and accurate methods ofchange mapping highly desirable (Aspinall, 2002; Dobson &Bright, 1994; Hansen et al., 2002; Rogan & Chen, 2004).

Complex land change processes are of particular interest toresearchers involved in large area monitoring (Roberts et al.,2002), where many different types of land-cover changes canoccur and must be characterized (e.g., forest pest infestation,logging, wildfire, and suburbanization) (Rogan & Miller 2006).Thus, increased automation can ensure that the classificationprocess is objective and repeatable in processing large volumesof data over complex and phenologically diverse landscapes(DeFries & Chan 2000; Gong & Xu 2003). Consequently,classification algorithm selection and performance have become

Page 2: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

2273J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

particularly important, because large area change monitoringcan only realistically be achieved (i.e., low cost and general-izable results) through techniques that minimize time-consum-ing human interpretation and maximize automated proceduresfor data analysis (Woodcock et al., 2001).

Awide variety of classification algorithms have been used tomap land-cover/use and changes to land-cover/use, from re-motely sensed data. Unsupervised classification and clusterlabeling is the dominant method for large area land-cover map-ping and monitoring (see Wulder et al., 2004). Recently,machine learning algorithms (MLAs) have emerged as moreaccurate and efficient alternatives to conventional parametricalgorithms (e.g., maximum likelihood), when faced with largedata volumes and complex measurement spaces (Foody, 1995;Muchoney & Williamson 2001; Kasischke et al., 2004). Theeffectiveness of MLAs has been demonstrated primarily insingle-date land-cover mapping studies (Friedl et al., 1999; Pal& Mather 2003). However, their effectiveness in change map-ping has also recently been addressed (Gopal & Woodcock1996; Liu & Lathrop 2002; Chan & Chan 2002).

The objective of this researchwas to compare the performanceof three MLAs; two classification tree software routines (S-Plusand C4.5) and an artificial neural network (ARTMAP), in thecontext of mapping land-cover modifications in southern andnorthern California between 1990/91 and 1996. These algorithmswere specifically chosen because they are increasingly used inland-covermapping andmonitoring usingmedium-coarse spatialresolution remotely sensed data (Rogan & Chen 2004), yet havenot been compared with each other in a large area, large datavolume context. This study is based on the US Forest Service andCalifornia Department of Forestry and Fire Protection statewideland-cover mapping and monitoring program (LCMMP) inCalifornia. LCMMP personnel required an evaluation ofavailablemachine learning algorithms to determine the feasibilityof automating their land-cover change detection procedures. Thecurrent LCMMP is based on time-consuming unsupervisedclassification and cluster labeling (Levien et al., 1999).

Algorithm comparison followed two steps. Step one assessedhow algorithm performance was affected by modifications inthe model data. Performance was assessed using classificationaccuracy measures and the model data set was modified to testthe following: (1) effect of partitioning data sets; (2) effect oftraining data set size; and (3) the effect of training data errors(i.e., noise). Step two compared a map of land-cover change,generated using the algorithm deemed best overall from stepone, to a map produced by the LCMMP that was createdthrough unsupervised classification and cluster labeling byhuman interpreters. The results of this work are intended toinform the growing number of resource managers engaged inother operational land-cover monitoring projects.

2. Background

2.1.Machine learning algorithms in land-cover changemapping

Machine learning refers to induction algorithms that analyzeinformation, recognize patterns, and improve prediction accu-

racy through automated, repeated learning from training data(Malerba et al., 2001). There is now a large body of research thatdemonstrates the abilities of machine learning techniques, par-ticularly classification trees and artificial neural networks, todeal effectively with tasks involving high dimensional data(Gahegan 2003). The increased interest in MLAs can be attrib-uted to several factors:

• their non-parametric nature deals well with multi-modal,noisy and missing data (Hastie et al. 2001), but see Simardet al. (2000) in the case of classification trees;

• there is a significant reduction in computational demandswhen data measurement spaces are large and complex (Foody2003);

• they readily accommodate both categorical and continuousancillary data (Lawrence & Wright 2001);

• users can investigate the relative importance of input vari-ables in terms of contribution to classification accuracy(Hansen et al., 1996; Foody & Arora 1997);

• they are flexible and can be adapted to improve performancefor particular problems (Lees & Ritman 1991)

• and multiple subcategories per response variable can beaccommodated (Gopal et al., 1999).

However, while promising, the above assertions have not beentested robustly in the context of land-cover changemapping. As aresult, MLAs have yet to be fully incorporated in large areastudies of land-cover change monitoring (Rogan &Miller 2006).

2.1.1. Classification trees and neural networksClassification trees are a type of MLA used to predict

membership of cases of a categorical dependent variable fromtheir measurements on one or more predictor variables (De'ath& Fabricius 2000). Classification trees are developed usingdifferent measures that recursively split data sets into increas-ingly homogeneous subsets representing class membership. Allclassification tree approaches employ hierarchical, recursivepartitioning of the data, resulting in decision rules that relatevalues or thresholds in the predictor variables with pixel classes(Friedl & Brodley 1997). An important advantage of classifica-tion trees is that they are structurally explicit, allowing for clearinterpretation of the links between the dependent variable ofclass membership and the independent variables of remotesensing and/or ancillary data (Lawrence & Wright 2001).

Generally, a neural network learns a pattern by iterativelyconsidering each training observation and then multiplying theexplanatory variables by a set of weights, applying a set oftransfer functions to their weighted sum, and finally predictingmembership for each desired map class (Franklin et al., 2003).There are many different types of neural network algorithmsavailable including multi-layer perceptrons, learning vectorquantization, Hopfield, and Kohonen Self Organizing Maps.However, a large degree of uncertainty remains as to whichnetwork algorithm works best for a specific application (Borak& Strahler 1999).

MLA applications and comparisons in change mapping arerelatively uncommon. Past studies range in complexity from

Page 3: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

2274 J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

simple change/no change mapping to complex studies invol-ving multiple land-cover change classes (Woodcock et al.,2001). Previous studies in which MLAs were compared withconventional algorithms reported higher overall map accuracyresults for the MLA (see Dai & Khorram 1999; Olthof et al.,2004). As no single MLA has been demonstrated to be superiorfor all applications (Kohavi et al., 1996), it is useful to comparea number of algorithms for a specific application (DeFries &Chan 2000). Further, few studies have compared the perfor-mance of MLAs in the context of large area, operational land-cover monitoring.

3. Methods

3.1. Study areas

Two different study areas were used in this research: (1) SanDiego County, southern California (area 10,970 km2) and; (2)Plumas and Lassen National Forests, northern California (area21,000 km2) (Fig. 1). These study areas were chosen based ontheir contrasting land-cover types and disturbance regimes,which may influence the robustness of the three different MLAs.San Diego County is composed of a variety of heterogeneousland-cover types, including shrub-grassland (60%), conifer and

Fig. 1. Locations of study ar

hardwood forest (12%), agriculture (6%) and urban (18%). ThePlumas/Lassen region is more homogenous and dominated byupland conifer forest (65%) and shrub-grassland (15%). Bothstudy areas are subject to land-cover disturbances includingwildfire, forest pest infestation, logging, and urbanization.

3.2. Data

Four Landsat-5 TM images were used to map land-coverchanges in southern California between 1990 and 1996: 8 Juneand 24 June 1990 (path 39/row 37), and 8 June and 17 June1996 (path 40/row 37). Four Landsat-5 TM images were alsoused to map land-cover changes in northern California between1991 and 1996: 26 August and 9 August 1991(path 44/row 32),and 10 August and 19 August 1996 (path 43/row 33). Allimages were registered to the UTMWGS84 projection using anautomated tie-point, feature-matching geometric registrationalgorithm contained in the DIME software (Positive Systems,Inc) to an independent root mean square error of 0.4 pixels. Anearest neighbor algorithm was used to resample each image to30-m spatial resolution. All images were normalized foratmospheric and illumination differences independently andconverted to reflectance values using a dark-object subtractiontechnique (Chavez, 1996).

eas in California, USA.

Page 4: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

Table 2Land-cover change classification scheme defined by U.S. Forest Service

Quantitative Land-cover Change Class Descriptive Land-cover Change Class

+/−15% canopy change No change/Phenology-related change−71 to −100% canopy change Large decrease in forest cover−41 to −70% canopy change Moderate decrease in forest cover−16 to 40% canopy change Small decrease in forest coverShrub/grass decrease N15% Large decrease in shrub cover+16 to +40% canopy change Moderate increase in forest cover+41 to 100% canopy change Large increase in forest coverShrub/grass increase N15% Large increase in shrub coverChange in developed areas Change in developed areas

Table 1Remote sensing and ancillary data variables used as input to classification (30 m spatial resolution)

Variable Variable name Data range Data units

SpectralΔ Brightness MKT1 0–255 Rescaled reflectanceΔ Greenness MKT2 0–255 Rescaled reflectanceΔ Wetness MKT3 0–255 Rescaled reflectanceStable Brightness MKT4 0–255 Rescaled reflectanceStable Greenness MKT5 0–255 Rescaled reflectanceStable Wetness MKT6 0–255 Rescaled reflectanceTexture Texture 0–255 Rescaled reflectance

Ancillary Southern California Northern CaliforniaElevation Elevation 0–1991 98–3180 MetersSlope Slope 0–66 0–75 DegreesSouthwestness Southwestness −1–1 −1–1 cos(aspect-255)Fire History Fire 0–1 (0=no fire; 1=fire) 0–1 (0=no fire; 1=fire) BinaryExisting Land-cover Land-cover 9 a classes: shrub, hardwood, conifer, mixed,

urban, erbaceous, barren, water, agriculture9 classes: conifer, shrub, herbaceous, barren,mixed, hardwood, agriculture, water, urban

Categorical

a Classes listed in order of areal dominance for each study area.

2275J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

Several ancillary variables were input to each MLAclassification (Table 1). The Multitemporal Kauth Thomas(MKT) linear transformation (Collins & Woodcock 1996) wasused to spectrally enhance the atmospherically corrected TMdata prior to MLA classification. Ancillary variables were cho-sen based on hypothesized relationships between thesevariables and land-cover modifications in California forestsand shrublands. Slope was derived from a USGS digital eleva-tion model (DEM). Southwestness, a measure of deviation fromNorth, was calculated from an aspect grid using the followingtransformation: cos(aspect-255). Fire history and land-coverdata were provided by the US Forest Service Fire Resource andAssessment Program (http://frap.cdf.ca.gov/). Texture variableshave been used previously for monitoring land-cover change inforest areas by providing contextual information about thespatial arrangement of pixels affected by multitemporal change(Riou & Seyler 1997; Chan et al., 2003). A 5×5 pixel variancefilter was applied to Landsat TM band 4 (i.e., near-infrared) foreach image date to derive a texture measure representing spatialvariance.

3.3. Land-cover change categories and reference data collection

The land-cover change classification scheme used in thisstudy was developed by the LCMMP for inclusive mapping of awide range of land-cover modifications in California forests andshrublands (Levien et al., 1999). This scheme was defined to beapplicable to a wide magnitude of land-cover changes in termsof change in percent canopy cover (e.g., discrete intervals offorest canopy cover, shrub cover, and change in developedareas) (Table 2). The +/−15% change class was designed toreduce the confusion between inter-date phenological varia-bility and discrete land-cover disturbances (Rogan et al., 2003).

Each study area was initially stratified into six categories ofKauth Thomas brightness, greenness and wetness changemagnitude (i.e., low-medium-high decrease and low-medium-high increase). The stratified change image was combined with

the land-cover map (see Table 1) and used to prioritize thelocation of field data collection and selection of available aerialphotographs. The two-step landscape stratification procedureensured sampling of likely areas that changed over time whilemaking those samples representative of land-cover change/nochange based on land-cover type.

Ground reference land-cover data for 1990/1991 (first timeperiod) were acquired for both study regions using dot gridsampling of tree canopy/shrub cover by interpretation of 1:15,840-scale true-color forest resource photographs. Ground ref-erence cover proportion data for 1996 (second time period) wereestimated through dot grid sampling of tree canopy/shrub coverby interpretation of 1:15,840 scale (resolution b1 m) color-infrared imagery (digital images were captured using a DCS420Kodak infrared camera), acquired over 400 selected sites in bothregions. Additional data were collected in the field using aprotocol described in detail by Rogan et al. (2003).

The ground reference data sets for both study areas wererandomly divided into 70:30, train:test proportions. This re-sulted in 560 sites for training and 240 sites for testing insouthern California, and 525 sites for training and 225 sites fortesting in northern California, representing nine change classes.The number of training sites per class was kept roughly equal(i.e., approximately 60 training sites and 30 testing sites per

Page 5: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

2276 J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

class). The same training and test data were used to train and testdifferent MLAs, respective to the two study areas. Thus, it waspossible to examine both the relative performance of the differ-ent classifiers and the consistency of these comparisons acrossthe two study areas.

3.4. Machine learning algorithms

3.4.1. S-Plus classification tree (CT)The first classification tree routine tested was the one

incorporated in the S-Plus statistical software package (Clark &Pregibon 1992), one of the most widely used tree algorithms inclassification of land-cover (Hansen et al., 1996, Wessels et al.,2004), which employs a deviance measure to partition a dataset. The reduction in deviance (e.g., increase in subsethomogeneity using and entropy measure) (D) is calculated as:D=Ds−Dt−Du, where s represents the parent node, and t and uare splits from s. When D is maximized, the best split isidentified, and the data are divided at that value. The process isrepeated on the two new nodes of the tree. The deviance fornodes is calculated from

Di ¼ �2X

nik logpik ð1Þwhere n is the number of observations in class k in node i and pis the probability distribution of node i and class k. S-Plusclassification trees were pruned to an optimum size based oncross-validation using ten independent subsets of the trainingdata. This resulted in a parsimonious tree model that did notoverfit the training data, thus leading to more generalizableinformation (Franklin, 1998). While there is generally not asingle solution to pruning for all applications of classificationtrees, and the decision of how much to prune can affect theresults, we did not compare different pruning trials in this work,as has been suggested by Zambon et al. (2006). We refer to S-Plus classification tree as CT.

3.4.2. C4.5 classification tree (MLC)We also tested the C4.5 software package. This algorithm is

witnessing near-operational use by the US Geological Surveyfor creation of the National Land-Cover Dataset (Homer et al.,2004), and global land-cover mapping using data from theModerate Resolution Imaging Spectroradiometer (MODIS)(Friedl et al., 1999). C4.5 uses the ‘information gain ratio’ toestimate splits at each node of a tree, where the informationgain is a measurement of the reduction in entropy (i.e., loss ofinformation) in the data produced by a split (Quinlan 1993).Using this metric, the decision at each node within a tree is madebased on the subset of the data that maximizes the reduction inentropy of the descendent nodes (DeFries & Chan 2000). Theinformation gain ratio is defined as

gain ratio Mð Þ ¼ gain Mð Þ = split info Mð Þ ð2Þ

where gain (M) is the information after an attribute M ischosen as a test for the training samples to divide on, and splitinfo (M) is the information generated when x training samples

are partitioned into n subsets. For more details on this approach,see Quinlan (1993) andDeFries and Chan (2000). The trees weregenerated in C4.5 using the ‘iterative’ mode, which generates aseries of decision trees based on randomly-selected subsets ofthe data (termed the ‘window’). Misclassified objects are addeduntil the trial decision tree correctly classifies all objects not inthe window or until there is no improvement (Quinlan, 1993).

Another distinguishing feature of C4.5 is the method bywhich rule sets are derived. After a tree is generated, pruning isperformed on the splitting rules that define the terminal nodes forgeneralizaton, i.e., one or more conditions can be droppedwithout changing the subset of observations that fall in the node(Hastie et al., 2001). This results in a simplified set of non-hierarchical intuitive classification rules defining each terminalnode (Huang & Jensen 1997). The C4.5 program used in thisresearch was operable within the ERDAS Imagine imageprocessing/GIS, in a module called the ‘Machine LearningClassifier’, (MLC) following the work of Huang and Jensen(1997). Henceforth in this paper, the C4.5 algorithm will bereferred to as MLC.

3.4.3. ARTMAPThe third machine learning algorithm tested in this study was

the fuzzy ARTMAP neural network algorithm (Carpenter et al.,1992). ARTMAP has seen increasing use in land-covermapping (Carpenter et al., 1999) and change detection(Abuelgasim et al., 1999), using Landsat data but has notbeen tested in a large area change mapping context. Neuralnetworks utilizing adaptive resonance theory (ART) have beenshown to minimize the three main drawbacks of backpropaga-tion networks, sensitivity to the choice of initial network pa-rameters, overfitting, and the need for user intervention duringthe training phase (Mannan et al., 1998). In contrast to con-ventional neural network algorithms (i.e., multi-layer percep-trons), ART networks utilize match-based learning. Thispreserves significant past learning while still allowing newinformation to be incorporated in the network structure as itappears in the data input stream and therefore, less userinteraction is required (Carpenter et al., 1999).

The architecture of fuzzy ARTMAP consists of two fuzzyART modules, ARTa and ARTb, connected by an associativelearning network called a ‘map field.’ Each ART module iscomposed of two layers of processing elements known as F1(input signals) and F2 (output categories) that are linked byweighted connections. The other component of the architectureis a controller that uses a minimum learning rule to conjointlyminimize predictive error and maximize predictive general-ization (Carpenter et al., 1997). This enables the system to learnoperationally and determine the number of hidden units(recognition categories) needed to meet the matching standards.The learning rate was set at a maximum 1.0 and the vigilanceparameter was 0.98 and required no further user interaction.

3.5. Classification and evaluation

To ensure consistency in algorithm comparison, the samesets of training and test data were used for CT, MLC, and

Page 6: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

Table 3Change map classification results (mean individual class and overall testaccuracy) for southern California

Category Number Land-cover Change Class CT MLC ARTMAP

1 +/−15% canopy change (nochange)

80.1 83.3 84.0

2 −71 to −100% canopy change 61.7 63.4 67.73 −41 to −70% canopy change 56.0 62.5 67.64 −16 to 40% canopy change 53.4 55.0 58.95 Shrub/grass decreaseN15% 61.4 64.9 71.86 +16 to+40% canopy change 68.3 72.1 79.07 +41 to 100% canopy change 71.6 75.8 77.78 Shrub/grass increaseN15% 75.8 78.5 83.09 Change in developed areas 94.1 94.0 97.2

Mean overall accuracy (%) 70.2 75.4 81.6Mean area-weightedoverall accuracy (%)

76.0 82.3 86.4

Mean kappa 0.69 0.75 0.79Mean kno 0.67 0.75 0.78

2277J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

ARTMAP. The ERDAS Imagine Expert Classifier was used toproduce change maps based on the CT decision rules for bothstudy areas. MLC and ARTMAP automatically generate mapsfrom tree and network decision rules. A 3×3 majority filter wasapplied to the six change maps to provide larger and morerealistic landscape units from an end-user or managementperspective, rather than isolated pixels that are independent oftheir surrounding neighborhood. Overall and per-class accuracymeasures were based on the confusion matrix. The calculationsassociated with these measures are described in (Congalton &Green, 1999). The Kappa coefficient and test Z statistics werealso computed.

3.6. Effect of variations in the training set

In a large area context, change mapping often relies upontraining data that are derived from disparate sources such asfield data and aerial photographs. Thus, an examination of theimpact of training site variation on algorithm performance iswarranted. MLAs and other data-driven algorithms are some-times sensitive to subtle differences in a training set, which canresult in a very different series of splits and altered tree ornetwork structure (Hastie et al., 2001). This “instability” inalgorithm performance has been noted by authors in a numberof practical applications (DeFries & Chan 2000; Rogan et al.,2003; Scull et al., 2005). To test the stability of the three MLAs,different partitions of train and test data were used. To approx-imate training sets with minor variability, the entire referencedata set was repeatedly randomly sampled to obtain ten differentsets of training (70%) and test (30%) sites for each study area.The same train/test sets were applied to each MLA to make theresults comparable.

3.7. Effect of reduction in the training set size

An examination of the impact of training site reduction onalgorithm performance is warranted in light of the logisticalchallenges associated with large area change mapping wheretraining data are often pushed to their operational limit (Roganet al., 2003). Training set size can affect model accuracy. Pre-sumably there is a minimum number of observations necessaryto achieve an acceptable level of model accuracy, as a muchlarger number of observations may be expensive to collect andnot improve model accuracy sufficiently. Additionally, thequality of the data is important — if the observations are notsufficiently representative, accuracy can decrease. Thus, theeffect of training set size on algorithm performance was exam-ined using classification accuracy as a metric for comparison.The training data were randomly divided into subsets for bothstudy areas by reducing the number of training sites by 25% (45sites per class) and 50% (30 sites per class), respectively, withan equal number of sites per land-cover change class.

3.8. Effect of noise in the training set

Training data errors are likely in a large area context giventhe disparate sources of information used and the fact that class

labels can become more confused as landscape heterogeneityincreases. Several authors have noted that MLAs are adverselyaffected by noise, which can yield very different results whenincluded in the training phase by causing intra-class variabilityin the data (Simard et al., 2000; Miller & Franklin 2002).Conversely, Brodley and Friedl (1996) found classification treesto be tolerant of noisy data, and Paola and Schowengerdt (1995)found neural networks robust to training site heterogeneity.Therefore, it is instructive to examine the effect of noise in achange mapping context to provide more information on thistopic, and to examine the level of robustness one may expectfrom MLAs.

The effect of mislabeled training samples as a proxy for data‘noise’ in training accuracy was examined by randomlymislabeling the training set. We randomly assigned an incorrectlabel to each of the nine land-cover change classes for a subsetof sites. This was performed for subsets of 10%, 30%, and 50%of the total training sites. We did not investigate the effects ofnoise in the spectral and ancillary variables that were input tothe MLAs, as performed by DeFries and Chan (2000), becausethat study reported that mislabeling in the training data set had amore pronounced effect on classifier performance than artificialnoise added to the input spectral variables.

3.9. Comparison to LCMMP land-cover change maps

Finally, we compared the results of the “best” machinelearning algorithm, defined as the most stable and accurateoverall, least sensitive to noise, and most robust in the presenceof reduced volumes of training data, to a change map producedby the LCMMP (southern California 1990–1996) (Levien et al.,1999). The change mapping approach of the LCMMP is basedon per-scene unsupervised classification of imagery andancillary variables (similar to those in Table 1), followed bytime-intensive cluster labeling using aerial photography, fireperimeters and a limited set of ground reference data (Levienet al., 1999). The overall accuracy of this map, reduced to theperimeter of San Diego County, was assessed using the same

Page 7: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

Table 4Change map classification results (mean individual and overall class accuracy)for northern California

Category number Land-cover change class CT MLC ARTMAP

1 +/−15% canopy change (nochange)

83.0 83.2 85.2

2 −71 to −100% canopy change 67.5 69.3 73.63 −41 to −70% canopy change 62.8 65.0 67.24 −16 to 40% canopy change 56.1 57.9 62.65 Shrub/grass decreaseN15% 66.9 73.5 77.26 +16 to +40% canopy change 75.0 78.3 80.47 +41 to 100% canopy change 79.5 82.8 84.78 Shrub/grass increase N15% 83.3 87.5 89.29 Change in developed areas 91.3 92.0 94.3

Mean overall accuracy (%) 74.5 79.9 84.3Mean area-weightedoverall accuracy (%)

77.7 83.6 88.8

Mean kappa 0.72 0.77 0.83Mean kno 0.72 0.77 0.81

Fig. 2. Effect of variations in the training and test sites on mean overall accuracyfor (a) southern California and (b) northern California.

2278 J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

test samples used to compare the three MLAs and wascalculated to be 76% (overall accuracy).

4. Results

4.1. Classification accuracy

4.1.1. Southern CaliforniaThe accuracy statistics for southern California represent the

overall and mean per-class accuracy of ten independentlysampled train/test data sets (Table 3). Per-class statistics andoverall accuracy statistics represent the mean accuracy from tenindependent trials of map validation for each MLA. The knostatistic was used here to compensate some of the shortcomingsof the kappa statistic, as kappa is not a chance-corrected mea-sure of agreement and does not make distinctions among var-ious types and sources of disagreement (Pontius 2000).

Overall accuracy results were high for change mappingefforts in shrubland-dominated environments (mean MLAoverall accuracy was 76%). ARTMAP produced the highestoverall accuracy (81.6%), followed by MLC (75.4%) and CT(70.2%). Overall MLA accuracies were significantly higherwhen the area of each land-cover change class weighted theresults, e.g., ARTMAP weighted accuracy was 86.4%. Similarto the overall accuracy results, kappa accuracies were sig-nificantly higher for ARTMAP (0.79), compared to MLC (0.75)and CT (0.69). Examining per-class accuracy, ARTMAPperformed best, followed by MLC and CT. Change in devel-oped areas (Class 9) was the most accurately represented classfor all three MLAs, followed by class 1 (+/−15% canopychange, or no change), classes 6–8 (forest and shrub coverincrease), and finally, classes 2–5 (forest and shrub coverdecrease).

4.1.2. Northern CaliforniaOverall accuracy results for northern California were higher

than southern California, i.e., northern California mean MLAoverall accuracy was 79.5%, and mean kappa accuracy was0.77% (Table 4). Similar to southern California results,

ARTMAP produced the highest overall accuracy (84.5%).The per-class accuracy results followed the same trends as thesouthern California data set, with class 9 (change in developedareas) being the most accurate class for all three MLAs, fol-lowed by class 1 (+/−15% canopy change), classes 6–8 (forestand shrub cover increase), and finally, classes 2–5 (forest andshrub cover decrease). Forest cover change per class accuracieswere approximately 4% higher than results for the southernCalifornia.

4.2. Effect of variations in the training set

The effect of varying train and test data samples for southernCalifornia is shown in Fig. 2(a). Training accuracies were high,especially for ARTMAP (mean overall accuracy was 92%). CTproduced the lowest mean training accuracies (77%). Trainingaccuracies showed relatively little variability associated withdifferent training samples. The greater variation in test accu-racies (indicated by error bars of one standard deviation) pro-vides further evidence that trees and neural networks are betterat recognizing previously learned patterns (resubstitution) thanunseen patterns (test data). The ten MLA test trials exhibited alow degree of variability and results demonstrated a closermatch between train and test accuracy, which is a good indi-cation of model generality and parsimony. ARTMAP testaccuracies deviated most from training accuracies, indicating ahigh degree of overfitting in the training model. Despite this,

Page 8: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

Fig. 4. Effect of noise in the training sites on overall test accuracy for southernCalifornia (a), and northern California (b).

2279J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

ARTMAP test accuracies were significantly highest overall(82%), followed by MLC (75%) and CT (70%).

For northern California, both training and test accuracieswere higher than those produced in the southern California dataset. Mean overall training accuracy of MLAs was 89% andmean test accuracy was 79%. ARTMAP train and test accu-racies were higher than either MLC or CT (mean trainingaccuracy was 97% and mean test accuracy was 84%). Further,all three MLAs were less sensitive to variations in the northernCalifornia data sets than found in the southern California datasets due to lower landscape heterogeneity (Fig. 2).

4.3. Effect of reduction in the training set size

For all MLAs and both study areas, reducing the number oftraining sites produced significantly lower overall map accu-racies (Fig. 3). A 50% reduction in training set size caused up to51% reduction in overall accuracy, when compared to thebaseline (i.e., no reduction). When the training set size wasreduced by 25%, accuracy relative to the baseline decreased by10% (ARTMAP) and 35% (CT). Examining the southernCalifornia data set, a 25% reduction in training sites loweredoverall test accuracy by an average of∼26% for the threeMLAs.ARTMAP was least sensitive to reduction in training samplesize. Similar results were found for northern California. Notably,the level of decrease in overall training accuracy was lower thanthat found in southern California, i.e., a 25% reduction in sitesreduced overall accuracy by 19% for all three MLAs.

4.4. Effect of noise in the training set

Higher proportions of noise in the training sites significantlyreduced overall accuracy for all MLAs when compared to the

Fig. 3. Effect of reduction in training set size on overall test accuracy for (a)southern California, and (b) northern California.

baseline (i.e., MLA accuracy on data set with 0% noise) (Fig. 4).Examining the southern California data, an increase in noise byas little as 10% decreased map accuracies by 6% (ARTMAP),7% (CT), and 20% (MLC). 30% noise reduced map accuracy byas much as 47% (CT). ARTMAP was most resistant to thepresence of noise compared to MLC and CT. However, at 50%noise, the reduction in map accuracy for all MLAs was large.The northern California data set was similarly affected by thepresence of noise, i.e., the percentage accuracy reduction rangedbetween 7% for ARTMAP and 14% for both CT and MLC at10% noise.

4.5. Comparison of LCMMPARTMAP land-cover change maps

Given the superior performance of ARTMAP in the algo-rithm comparisons, the map of land-cover change producedusing ARTMAP was compared to that produced from theLCMMP (based on multitemporal Landsat-5 TM and ancillarydata, including elevation and slope, using unsupervised classi-fication) for southern California (1990–1996) (Fig. 5). Thecomparison focused on western San Diego County. A compar-ison of the percentage of total area per land-cover change classis shown in Fig. 5.

Not surprisingly, no change (+/−15% canopy change) is thedominant class, i.e., LCMMP area total was 97.2%, andARTMAP area total was 96.53%. Class 5 (shrub/grass decreaseN15%) was the largest change class in both maps. However,ARTMAP reported a higher proportion of change in class 5(shrub/grass decreaseN15%). An examination of each mapreveals the cause of this difference. For example, severalpatches of shrub cover decrease are omitted from the LCMMPmap, i.e., in Palomar Ranger District, and on the outskirts of thecoastal cities in San Diego County. This omission resulted

Page 9: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

Fig. 5. Land-cover change maps generated for southern California: (a) LCMMP map, and (b) Fuzzy ARTMAP classification.

2280 J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

because the footprint of wildfires that occurred in grasslandsdisappeared between image acquisition dates (1990–1996)because of rapid post-fire regrowth by herbaceous vegetation.ARTMAP was successful in capturing these disturbances byincluding ancillary input variables, especially the dominantland-cover fire perimeter information.

The areal proportions of forest canopy cover changes weresimilar between the two maps, with ARTMAP producing higherproportions for each change category. The area of forest coverdecrease was higher for ARTMAP because it was moreeffective than the LCMMP approach in detecting low-levelsof canopy disturbance. Areal proportions for class 9 (change indeveloped areas) were marginally higher for the LCMMP map.The key difference between the two maps pertains to increasesand decreases in shrub cover between 1990 and 1996. Fig. 5bshows the location of four patches of shrub cover modification.Patch 1 and 2 are areas of actual shrub cover decrease that wereomitted from the LCMMP map. Patches 3 and 4 are areas ofshrub cover increase that were also omitted from the LCMMPmap.

5. Discussion and conclusions

The purpose of this paper was to compare the performance oftwo classification trees and a neural network to map large arealandscape change in California. Incorporating a suite of multi-

temporal Landsat data and environmental variables the threeclassifiers performed well in the context of large area changedetection with eight discrete change categories. The results ofthis research provide new insights to the performance of MLAsin the context of mapping land-cover change over large areas.The accuracy results for both study areas are promising for largearea change detection scenarios; overall kappa accuracies were0.79 for southern California and 0.84 for northern California.ARTMAP yielded the most accurate classification results in thisanalysis compared to MLC and CT.

One of the biggest differences in the performance of the threeMLAs is likely related to iteration. Both ARTMAP and themode in which we used MLC are iterative in their search forpatterns and their subsequent generation of rules, CT is not. Thislack of iteration could be responsible for the inferior per-formance of CT, especially when training data are reduced orperturbed. Voting classifiers such as bagging and boosting aretwo commonly used methods to achieve higher accuracy withclassification tree algorithms (Chan et al., 2001). However,voting methods were not applied in this analysis to allow for astraight comparison of algorithm performance.

When comparing per-class results, class accuracies werehigher for the northern California (forest dominated) than for thesouthern California (shrub and urban dominated).While ‘changein developed areas' (class 9) had high accuracies for both studyareas, vegetation increase and decrease change classes exhibited

Page 10: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

C

2281J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

notable differences. Canopy cover and change in conifer-dominated areas tended to be more reliably mapped due to morespatially homogeneous canopy closure and a stronger relation-ship between stand properties and spectral response than inhardwoods and shrublands. This result is similar to that found inunitemporal land-cover mapping studies, which have foundconifer-dominated areas to be more accurately mapped thanshrub dominated areas (Franklin et al., 2000).

The effect of training set size on algorithm performanceindicates that large numbers of training and test sites are impor-tant in change mapping using MLAs. Nonetheless, reasonableaccuracies were achieved at certain levels of data reduction,implying that the quality, if not the quantity of sample data wasadequate. Results suggest that below a certain size, a data set isless representative of the conditions it is supposed to represent,resulting in reduced map accuracy. It might also suggest that thedata sets used here have few redundant observations, or elsereducing the size would have less of an effect. Careful attentionshould be paid to training and test set selection in aMLA context.Variations in training and test site sample size had a significanteffect on the behavior of each MLA. ARTMAP produced morestable results overall, followed by MLC and CT. As substan-tiated in unitemporal land-cover mapping studies, the average ofa number of trials should be used to employMLA classificationsin a robust manner (Gahegan 2003). Further, greater error/noisein training data resulted in dramatically lower map accuracy.ARTMAP was least affected by noise while CT and MLC weremost affected. ARTMAP has potential for automated large areamonitoring over diverse landscape types. The comparisonbetween the ARTMAP and LCMMP change maps revealedthat ARTMAP has potential to increase automation in large areachange mapping while achieving reasonable map accuracy.Future work will examine the influence of input variables onclassification accuracy and extend the methodology developedhere to study regions outside California.

Remote sensing-based mapping and monitoring of largeareas has created new challenges for the remote sensing andland change science communities. The spatial scale and datavolumes of large area change monitoring dictate the need forclassification algorithms that minimize requirements for humanintervention and promote automated classification. Attempts tosolve these challenges have led researchers away from acceptedand well-understood parametric algorithms, such as maximumlikelihood, to newer non-parametric algorithms that offer in-creased accuracy and performance over large areas and datavolumes. At the same time, these algorithms are less understoodin the remote sensing context, particularly in land-cover changemapping. To address this issue, this study provided insightstoward improved understanding of non-parametric algorithmsfor large area change mapping and found the ARTMAP neuralnetwork classifier to be most robust and accurate in mappingland-cover modifications within southern and northern Cali-fornia study areas. The supervised classification approachdeveloped in this paper can potentially allow for the rapid re-mapping of the same study area at a different time period using(a) high quality ground reference data and (b) a robust clas-sification algorithm.

References

Abuelgasim, A. A., Ross, W., Gopal, S., & Woodcock, C. E. (1999). Changedetection using adaptive fuzzy neural networks environmental damageassessment after the Gulf War. Remote Sensing of Environment, 70, 208−223.

Aspinall, R. (2002). A land-cover data infrastructure for measurement,modeling, and analysis of land-cover change dynamics. PhotogrammetricEngineering and Remote Sensing, 68(10), 1101−1105.

Borak, J. S., & Strahler, A. H. (1999). Feature selection and land-coverclassification of a MODIS-like data set for a semiarid environment. Inter-national Journal of Remote Sensing, 20(5), 919−938.

Brodley, C., & Friedl, M. A. (1996). Identifying and eliminating mislabeled traininginstances. In Proceedings of Thirteenth National Conference on ArtificialIntelligence, Portland Oregon, August, 1996 (pp. 799−805). : AAAI Press.

Carpenter, G. A., Grossberg, S., Markuzon, N., Reynolds, J. H., & Rosen, D. B.(1992). Fuzzy ARTMAP: A neural network architecture for incrementalsupervised learning of analog multidimensional maps. IEEE Transactions onNeural Networks, 3, 698−713.

Carpenter, G. A., Gjaja, M. N., Gopal, S., & Woodcock, C. E. (1997). ARTneural networks for remote sensing: Vegetation classification from LandsatTM and terrain data. IEEE Transactions on Geoscience and RemoteSensing, 35(2), 308−325.

Carpenter, G. A., Gopal, S., Macomber, S., Martens, S., Woodcock, C. E., &Franklin, J. (1999). A neural network method for efficient vegetationmapping. Remote Sensing of Environment, 70, 326−338.

Chan, J. C-W., & Chan, K-P. (2002). The use of SOM-encoded texture spectrumfor change detection. In L. Bruzzone & P. Smits (Eds.), Proceedings of theFirst Annual Workshop on the Analysis of Multi-temporal Remote SensingImages (pp. 440). : World Scientific Publishing Co., Pte. Ltd.

Chan, J. C-W., Laporte, N., & DeFries, R. S. (2003). Texture classification oflogged forests in tropical Africa using machine-learning algorithms. Inter-national Journal of Remote Sensing, 24(6), 1401−1407.

havez, P. S. (1996). Image-based atmospheric corrections revisited and improved.Photogrammetric Engineering and Remote Sensing, 62, 1025−1036.

Cihlar, J. (2000). Land-cover mapping of large areas from satellites: status andresearch priorities. International Journal of Remote Sensing, 21(6 and 7),1093−1114.

Clark, L. A., & Pregibon, D. (1992). Tree-based models. Pages 377–419. In T. J.Hastie (Ed.), Statistical Models in S Pacific Grove, California: Wadsworth.

Collins, J. B., & Woodcock, C. E. (1996). An assessment of several linearchange detection techniques for mapping forest mortality using multi-temporal Landsat TM data. Remote Sensing of Environment, 56, 66−77.

Congalton, R. G., & Green, K. (1999). Assessing the accuracy of remotelysensed data Boca Raton, FL: CRC Press 137 pp.

Dai, X. L., & Khorram, S. (1999). Remotely sensed change detection based onartificial neural networks. Photogrammetric Engineering and RemoteSensing, 65(10), 1187−1194.

De'ath, G., & Fabricius, K. E. (2000). Classification and regression trees: Apowerful yet simple technique for ecological data analysis. Ecology, 81(11),3178−3192.

DeFries, R. S., & Chan, J. (2000). Multiple criteria for evaluating machinelearning algorithms for land-cover classification from satellite data. RemoteSensing of Environment, 74, 503−515.

Dobson, J., & Bright, E. (1994). Large-area change analysis: The Coast-watchChange analysis Project (C-CAP). Proceedings of Pecora 12, 24–26August 1993, Sioux Falls, South Dakota (pp. 73−81). Bethesda, Maryland:American Society for Photogrammetry and Remote Sensing.

Foody, G. M. (1995). Land-cover classification by an artificial neural networkwith ancillary information. International Journal of Geographical Informa-tion Systems, 9(5), 527−542.

Foody, G. M., & Arora, M. K. (1997). Evaluation of some factors affecting theaccuracy of classification by an artificial neural network. InternationalJournal of Remote Sensing, 18, 799−810.

Foody, G. M. (2003). Uncertainty, knowledge discovery and data mining in GIS.Progress in Physical Geography, 27(1), 113−121.

Franklin, J. (1998). Predicting the distribution of shrub species in southernCalifornia from climate and terrain-derived variables. Journal of VegetationScience, 9, 733−748.

Page 11: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

2282 J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

Franklin, J., Simons, D. K., Beardsley, D., Rogan, J., & Gordon, H. (2000).Evaluating errors in a digital vegetation map with forest inventory data andaccuracy assessment using fuzzy sets. Transactions in GIS, 5(4), 285−304.

Franklin, J., Phinn, S. R., Woodcock, C. E., & Rogan, J. (2003). Rationale andconceptual framework for classification approaches to assess forestresources and properties. In M. Wulder & S.E. Franklin (Eds.), Methodsand applications for remote sensing of forests Concepts and case studies(pp. 279−300). Dordrecht: Kluwer Academic Publishers.

Franklin, S. E., & Wulder, M. A. (2002). Remote sensing methods in mediumspatial resolution satellite data land-cover classification of large areas.Progress in Physical Geography, 26(2), 173−205.

Friedl, M. A., & Brodley, C. E. (1997). Decision tree classification of landcoverfrom remotely sensed data. Remote Sensing of Environment, 61, 399−408.

Friedl, M. A., Brodley, C. E., & Strahler, A. H. (1999). Maximizing land-coverclassification accuracies produced by decision trees at continental to globalscales. IEEE Transactions on Geoscience and Remote Sensing, 37(2),969−977.

Gahegan, M. (2003). Is inductive machine learning just another wild goose (ormight it lay the golden egg)? International Journal of GeographicInformation Science, 17(1), 69−92.

Gong, P., & Xu, B. (2003). Remote sensing of forests over time: Change types,methods, and opportunities. In M. Wulder & S.E. Franklin (Eds.), Methodsand applications for remote sensing of forests: Concepts and case studies(pp. 301−333). Dordrecht: Kluwer Academic Publishers.

Gopal, S., & Woodcock, C. E. (1996). Remote sensing of forest change usingartificial neural networks. IEEE Transactions on Geoscience and RemoteSensing, 34, 398−404.

Gopal, S., Woodcock, C. E., & Strahler, A. H. (1999). Fuzzy neural networkclassification of global land-cover from a 1o AVHRR data set. RemoteSensing of Environment, 67, 230−243.

Hansen, M., Dubayah, R., & DeFries, R. (1996). Classification trees: Analternative to traditional land-coverclassifiers. International Journal ofRemote Sensing, 17(5), 1075−1081.

Hansen, M. C., Chan, J. C. -W., Pagis, J., De Fries, R., & Luo, D. (2002). Longterm change detection using continuous fields of tree cover from 8KMAVHRR data for the years 1982–2000. In L. Bruzzone & P. Smits (Eds.),Proceedings of the First Annual Workshop on the Analysis of Multi-temporalRemote Sensing Images : World Scientific Publishing Co., Pte. Ltd. 440 pp.

Hastie, T., Tibshirani, R., & Friedman, J. (2001). The elements of statisticallearning: Data mining, inference, and prediction (pp. 536). New York:Springer–Verlag.

Homer, C., Huang, C., Yang, L., Wylie, B., & Coan, M. (2004). Development ofa 2001 national land-cover database for the United States. PhotogrammetricEngineering & Remote Sensing, 70, 829−840.

Huang, X., & Jensen, J. R. (1997). A machine-learning approach to automatedknowledge-base building for remote sensing image analysis with GIS data.Photogrammetric Engineering and Remote Sensing, 63, 1185−1194.

Intergovernmental Panel on Climate Change (2000). Land use, land-usechange, and forestry. Special report (pp. 184). Cambridge: CambridgeUniversity Press.

Kasischke, E. S.,Goetz, S.,Hansen,M.C.,Ozdogan,M.,Rogan, J.,Ustin, S. L., et al.(2004). Temperate and boreal forests. In S. Ustin (Ed.), Manual of RemoteSensing Volume 4: Remote Sensing for Natural Resource Management andEnvironmental Monitoring : John Wiley & Sons 848 p.

Kohavi, R., Sommerfield, D., & Dougherty, J. (1996). Data mining using MLC+:A machine learning library in C+. International Journal of ArtificialIntelligence Tools, 6, 537−566.

Lawrence, R. L., & Wright, A. (2001). Rule-based classification systems usingclassification and regression tree (CART) analysis. PhotogrammetricEngineering and Remote Sensing, 67(10), 1137−1142.

Lees, B. G., & Ritman, K. (1991). Decision-tree and rule-induction approach tointegration of remotely sensed andGISdata inmapping vegetation in disturbedor hilly environments. Environmental Management, 15(6), 823−831.

Levien, L. M., Fischer, C. S., Roffers, P. D., Maurizi, B. A., Suero, J., & Huang,X. (1999). A machine-learning approach to change detection using multi-scale imagery. Proceedings of ASPRS Annual Conference Portland, Oregon.

Liu, X., & Lathrop, R. G., Jr (2002). Urban change detection based on an artificialneural network. International Journal of Remote Sensing, 23(12), 2513−2518.

Loveland, T. R., Sohl, T. L., Stehman, S. V., Gallant, A. L., Sayler, K. L. &Napton, D. E. (2002). A strategy for estimating the rates of recent UnitedStates land-cover changes. Photogrammetric Engineering and RemoteSensing, 68, 1091−1099.

Malerba, D., Esposito, F., Lanza, A., & Lisi, F. A. (2001). Machine learning forinformation extraction from topographic maps. In H. J. Miller & J. Han(Eds.), Geographic Data Mining and Knowledge Discovery (pp. 291−314).New York, NY: Taylor and Francis, Inc.

Mannan, B., Roy, J., & Ray, A. K. (1998). Fuzzy ARTMAP supervisedclassification of remotely-sensed images. International Journal of RemoteSensing, 19, 767−774.

Miller, J., & Franklin, J. (2002). Modeling the distribution of four vegetationalliances using generalized linear models and classification trees with spatialdependence. Ecological Modelling, 157, 227−247.

Muchoney, D., & Williamson, J. (2001). A Gaussian adaptive resonance theoryneural network classification algorithm applied to supervised land-cover-mapping using multitemporal vegetation index data. IEEE Transactions onGeoscience and Remote Sensing, 39(9), 1969−1977.

Olthof, I., King, D. J., & Lautenschlager, R. A. (2004). Mapping deciduousstorm damage using Landsat and environmental data. Remote Sensing ofEnvironment, 89, 484−496.

Pal, M., & Mather, P. M. (2003). An assessment of the effectiveness of decisiontree methods for land-cover classification. Remote Sensing of Environment,30, 554−565.

Paola, J. D., & Schowengerdt, R. A. (1995). A review and analysis ofbackpropagation neural networks for classification of remotely sensedmultispectral imagery. International Journal of Remote Sensing, 16(16),3033−3058.

Pontius, R. G., Jr (2000). Quantification error versus location error incomparison of categorical maps. Photogrammetric Engineering and RemoteSensing, 66(8), 1011−1016.

Quinlan, J. R. (1993). C4.5: Programs for machine learning. San Mateo, CA:Morgan Kaufmann.

Riou, R., & Seyler, F. (1997). Texture analysis of tropical rain forest infraredsatellite images. Photogrammetric Engineering and Remote Sensing, 63,515−521.

Roberts, D.A.,Numata, I., Holmes, K.W., Batista, G., Krug, T.,Monteiro, A., et al.(2002). Large area mapping of land-cover change in Rondônia usingmultitemporal spectral mixture analysis and decision tree classifiers.Journalof Geophysical Research, 107(D20), 8073 LBA 40-1 to 40-18.

Rogan, J., Miller, J., Stow, D. A., Franklin, J., Levien, L., & Fischer, C. (2003).Land-cover change mapping in California using classification trees withLandsat TM and ancillary data. Photogrammetric Engineering and RemoteSensing, 69(7), 793−804.

Rogan, J., & Chen, D. (2004). Remote sensing technology for mapping andmonitoring land cover and land use change.Progress in Planning, 61, 301−325.

Rogan, J., & Miller, J. (2006). GIS data in mapping of forest disturbance andchange. In M. Wulder & S. Franklin (Eds.), Forest Disturbance and SpatialPattern : CRC/Lewis.

Scull, P., Franklin, J., & Chadwick, O. A. (2005). The application ofclassification tree analysis to soil type prediction in a desert landscape.Ecological Modeling, 181(1), 1−15.

Simard, M., Saatchi, S. S., & De Grandi, G. (2000). The use of a decision treeand multiscale texture for classification of JERS-1 SAR data over tropicalforest. IEEE Transactions on Geoscience and Remote Sensing, 38(5),2310−2321.

Turner, B. L., II, Skole, D., Sanderson, S., Fischer, G., Fresco, L. O., &Leemans, R. (1999). Land-use and land-cover change science/researchplan. IGBP Report No. 35 and HDP Report No. 7. Stockholm InternationalGeosphere–Biosphere Programme.

Wessels, K. J., DeFries, R. S., Dempewolf, J., Anderson, L. O., Hansen, A. J.,Powell, S. L., et al. (2004).Mapping regional land-cover withMODIS data forbiological conservation: Examples from the Greater Yellowstone Ecosystem,USA and Para State, Brazil. Remote Sensing of Environment, 92, 67−83.

Woodcock, C. E., Macomber, S. A., Pax-Lenney, M., & Cohen, W. B. (2001).Monitoring large areas for forest change using Landsat Generalization acrossspace, time and Landsat sensors. Remote Sensing of Environment, 78,194−203.

Page 12: Mapping land-cover modifications over large areas: A comparison of machine learning algorithms

2283J. Rogan et al. / Remote Sensing of Environment 112 (2008) 2272–2283

Wulder, M. A., Kurz, W. A., & Gillis, M. (2004). National level forestmonitoring and modeling in Canada. Progress in Planning, 365−381.

Zambon, M., Lawrence, R., Bunn, A., & Powell, S. (2006). Effect of alternativesplitting rules on image processing using classification tree analysis. Pho-togrammetric Engineering and Remote Sensing, 72(1), 25–3.

Further reading

Biging, G. S., Colby, D. R., & Congalton, R. G. (1998). Sampling systems forchange detection accuracy assessment. In R. S. Lunetta & C.D. Elvidge (Eds.),Remote Sensing Change Detection: Environmental Monitoring Methods andApplications. (pp. 281−308). Chelsea, Michigan: Ann Arbor Press.

Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classifi-cation and Regression Trees, The Wadsworth Statistics and ProbabilitySeries, Wadsworth International Group, Belmont California (pp. 356).

Cohen, W. B., Spies, T., & Fiorella, M. (1995). Estimating the age and structureof forests in a multi-ownership landscape of western Oregon, U.S.A. Inter-national Journal of Remote Sensing, 16, 721−746.

Cohen, W. B., & Goward, S. N. (2004). Landsat's role in ecological applicationsof remote sensing. Bioscience, 54(6), 2004.

Coppin, P., Nackaerts, K., Queen, L., & Brewer, K. (2001). Operationalmonitoring of green biomass change for forest management. Photogram-metric Engineering and Remote Sensing, 67, 603−611.

Jakubauskas, M. E. (1996). Thematic Mapper characterization of lodgepole pineseral stages in Yellowstone National Park, USA. Remote Sensing ofEnvironment, 56, 118−132.

Rubin, M. A. (1995). Issues in automatic target recognition from radar rangeprofiles using Fuzzy ARTMAP. The 1995 World Congress on NeuralNetworks, vol. 1. (pp. 197−202)Mahwah, NJ: Lawrence ErlbaumAssociates.