Hierarchical object-based classiﬁcation of ultra-high ...

Hierarchical object-based classification of ultra-high-resolution digital mapping

camera (DMC) imagery for rangeland mapping and assessment

A.S. Lalibertea*, D.M. Browningb, J.E. Herrickb and P. Gronemeyera

aJornada Experimental Range, New Mexico State University, PO Box 30003, MSC3JER, NMSU, LasCruces, NM 88003-8003, USA; bUSDA Agricultural Research Service, Jornada Experimental Range, PO

Box 30003, MSC3JER, NMSU, Las Cruces, NM 88003-8003, USA

Ultra-high-resolution digital aerial imagery has great potential to complement or replaceground measurements of vegetation cover for rangeland monitoring and assessment. Thisresearch investigated object-based image analysis (OBIA) techniques for classifyingvegetation in southwestern USA arid rangelands with 4 cm resolution digital aerialimagery. We obtained high r-square values for the regressions relating ground- to image-based measures of percent cover (r-square values: 0.82–0.92). OBIA enabled us toautomate the classification process and demonstrated potential for quantifying fine-scaleland cover attributes with ultra-high-resolution imagery. This approach exhibits promisefor nationwide application for monitoring grazing lands.

Keywords: aerial photography; object-based image analysis; very high resolution;rangelands; vegetation classification

1. Introduction

In the United States, land managementagencies such as the Bureau of LandManagement (BLM) and the Natural Re-sources Conservation Service (NRCS) arerequired to monitor and assess vegetationconditions across millions of acres of grass-lands and savannas (i.e. rangelands). Field-based assessments are costly and inefficientover large areas; remote sensing offers thepotential to increase the number of mon-itoring locations, automate the image clas-sification process, and reduce costs of themonitoring. Remote sensing approacheshave been tested and implemented forrangeland applications at various spatialresolutions (Weber 2006; Ludwig et al.2007; Laliberte et al. 2007; Booth & Cox2008; Sankey et al. 2008). However, the

challenge that remains is developing aremote sensing-based approach that isrepeatable, potentially applicable to variousvegetation communities, and adapted toimagery of sufficiently high resolution toyield high correlations with ground-basedmeasurements commonly used in nationalmonitoring efforts.

This study is part of a larger researcheffort focused on developing and testingremote sensing acquisition and analysistechniques for potential integration intothe National Resources Inventory (NRI)of grazing lands to enhance assessmentof conservation effects. The NRI is astatistical survey of land use and naturalresource conditions and trends on U.S.non-federal lands conducted by the Natio-nal Resources Conservation Service

*Corresponding author. Email: [email protected]

Journal of Spatial Science

Vol. 55, No. 1, June 2010, 101–115

ISSN 1449-8596 print/ISSN 1836-5655 online

� 2010 Surveying and Spatial Sciences Institute and Mapping Sciences Institute, Australia

DOI: 10.1080/14498596.2010.487853

http://www.informaworld.com

(NRCS) (Nusser & Goebel 1997). The NRIprovides one part of the scientific frame-work for the Conservation Effects Assess-ment Project (CEAP), an inter-agency effortto quantify natural resource benefits deliv-ered through conservation actions on pri-vate land (Duriancik et al. 2008).

As part of NRI, NRCS acquires over70,000 aerial photos over the USA annuallyand is in the process of changing from filmto digital aerial image acquisition. Cur-rently, only broad land use/land coverclasses are determined through photo inter-pretation. In some cases, such as the2003 study initiated in southwestern range-lands (Godinez-Alvarez et al. 2009), moredetailed information is collected throughfield measurements and observations. How-ever, ground- and image-based measurescurrently have not been related. There isa keen interest to integrate remote sen-sing techniques to a greater extent andbetter correlate ground- and image-derivedmeasurements.

Digital aerial mapping cameras, suchas the large format UltraCam, DMC andADS40 are used to a greater extent thanfilm cameras for aerial photo acquisitionstoday (Neumann 2008). Compared to film-based products, images acquired with digi-tal mapping cameras have greater radio-metric resolution, lower noise levels and arebetter suited for quantitative remote sensing(Honkavaara & Markelin 2008). The tech-nology has seen increasing use in the lastfive years, and there is a growing literatureof remote sensing applications with digitalaerial imagery (Green & Lopez 2007;Coulter & Stow 2008; Rosso et al. 2008).

Compared to pixel-based approaches,object-based image analysis (OBIA) withhigh- or very-high-resolution (51 m) aerialphotography has been shown to yieldlower errors and better regression modelsfor urban impervious mapping (Hodgsonet al. 2003), and higher classification ac-curacies for detailed vegetation mapping at

the alliance level in California (Yu et al.2006). The ability to segment an image atmultiple scales allows for retaining fine-scale vegetation patches within a coarserlandscape element (Laliberte et al. 2004),and aids in determining appropriate scalesfor analysis (Laliberte & Rango 2009).Because individual pixels are grouped intohomogenous objects, the classificationdoes not suffer from the salt and peppereffect of pixel-based classification, and theimage objects are ecologically meaningful,whereas individual pixels may not be. Theability to incorporate spatial and contextualfeatures combined with expert knowledge atthe object-level allows for greater flexibilityin the analysis and improves the classifica-tion output (Platt & Rapoza 2008).

In this study, we investigated the use ofsub-decimetre resolution (4 cm ground re-solved distance) digital aerial imagery forestimating percent cover for vegetation andbare ground using OBIA. The objectiveswere (1) to compare image-based andground-based estimates of cover, and (2)to assess the viability and efficiency ofapplying the image-based method to abroad range of vegetation communitiesin the region. Future studies will extendthese techniques to additional vegetationcommunities.

2. Methods

Study area

Our research was conducted at the JornadaExperimental Range (JER) and the Chi-huahuan Desert Rangeland Research Cen-ter (CDRRC) in the Jornada Basin ofsouthwestern New Mexico in the northernChihuahuan Desert (latitude 328340110 N,longitude 1068490440 W) (Figure 1). Thearea is situated at about 1200 m elevationbetween the Rio Grande Valley to the westand the San Andres Mountains to the east.

Average monthly maximum tempera-tures for the JER range from 138C in

102 A.S. Laliberte et al.

January to 368C in June, and mean annualprecipitation is 241 mm, of which morethan 50 percent occurs in July, August andSeptember (Wainwright 2006). Rainfallamount and distribution is highly variable.Historically, a large part of the study areawas semidesert grassland, but shrub en-croachment has led to the conversion toshrubland over the last century (Gibbenset al. 2005). Common species include honeymesquite (Prosopis glandulosa Torr.), creo-sotebush (Larrea tridentata (Sess. & Moc.ex DC) Cov.), tarbush (Flourensia cernuaDC.), four-wing saltbush (Atriplex canes-cens (Pursh) Nutt.), soap-tree yucca (Yuccaelata Engleman.), mormon tea (Ephedratorreyana Wats.), broom snakeweed (Gu-tierrezia sarothrae (Pursh) Britt. & Rusby),black grama (Bouteloua eriopoda Torrey),tobosa (Pleuraphis mutica Buckley), drop-seed (Sporobolus spp.), threeawn (Aristidaspp.), and burrograss (Scleropogon brevifo-lius Phil.). Tobosa grasses are more likely tooccur in pure stands, while black grama,dropseed and threeawn tend to grow inmixed stands.

The fifteen plots (70 m 6 70 m) used inthis study were established as long-termresearch plots as part of the Jornada Basinlong-term ecological research (LTER) pro-gram. The plots were located in five differentvegetation communities (mesquite, tarbush,creosote, grassland, playa) to capture a widevariety of vegetation cover, density, andpattern (Figure 2). The mesquite, tarbushand creosote sites represent shrub-domi-nated communities, while the grasslandand playa sites are grass-dominated.

Digital imagery

Aerial imagery was acquired over the studyplots (one image per plot) on 19 and 21June 2007 with a large-format digitalmapping camera, the Intergraph Z/I Ima-ging1 Digital Mapping Camera (DMC).The dynamic range of the DMC imagery is12 bits, and the image size is 13,824 6 7680pixels. The multi-head sensor acquiresimagery in the red (590–675 nm), green(500–650 nm), blue (400–580 nm) and nearinfrared (675–950 nm) bands at a coarser

Figure 1. Map of study area in southwestern New Mexico at the Jornada Experimental Range (JER)and the Chihuahuan Desert Rangeland Research Center (CDRRC), showing the locations of the 15sites (black dots), over which the digital aerial imagery was acquired.

Journal of Spatial Science 103

resolution, and panchromatic data (400–950 nm) at finer resolution, with a groundsampling distance reduction factor of 4.8.The individual bands are then merged into apan-sharpened multispectral image (Hinzet al. 2000) using a proprietary method. Theimagery for this project was acquired at a

flying height of approximately 300 m aboveground, which resulted in a ground resolveddistance of 4 cm. Imagery was delivered asa three-band colour infrared image (nearinfrared (NIR), red (R), green (G)).

A subset of each image that encom-passed a 90 m 6 90 m area centred on the70 m 6 70 m plot was georectified to anorthorectified QuickBird image (60 cm re-solution). For ground control points, differ-entially corrected GPS coordinates of thecorner fence posts surrounding eachplot were used as well as other visibleobjects, such as large rocks and prominentshrubs. The root mean square error forthe polynomial model fit to the GCPsused to calibrate the transformation was52 cm.

Field measurements

Within three weeks of image acquisition,line-point intercept (LPI) data were col-lected following a standard rangeland mon-itoring protocol (Herrick et al. 2005). LPIdata collection was chosen because it is anintegral part of the NRI assessment. In eachplot, LPI data were collected along fourparallel 70 m transects for a total of 280points/plot (Godinez-Alvarez et al. 2009).For each point at a specified distance(i.e., 1 m) along the line transect, a pin isdropped to the ground, and all plant speciesintercepted by the pin are recorded. Inaddition, soil surface conditions (e.g., bare,litter, rock) are recorded. Field-based LPIsurveys are based on multiple hits per point,but in order to compare results withremotely sensed data, only the top or firstintercept of vegetation or soil were used forthe analysis. Percent cover by species wasderived by dividing the number of hitsfor each species by the total number ofpoints/plot. Vegetation cover at the specieslevel was aggregated into structure groups(shrubs, grasses, forbs), and then furtheraggregated into vegetation/non-vegetation

Figure 2. Photographs (left column) and digitalaerial imagery (right column) of the fivevegetation communities assessed in this study.The digital aerial imagery is clipped to the70 m 6 70 m plots.


for regression analysis. Species-specific datawere used to assess detection limits.

Digital image analysis

The object-based image analysis programDefiniens Developer 7.0.9 (Definiens 2007)was used for image classification. In a firststep, an image is segmented into homo-geneous areas, and the second step consistsof classification of these segments to deriveimage objects. The segmentation is abottom-up region merging approach,whereby smaller image objects are mergedinto larger ones depending on heterogeneitybetween adjacent image objects. Unit-lessparameters (scale, colour and shape) con-trol the segmentation. Colour and shape areweighted from 0 to 1, and within the shapeparameter, smoothness and compactnessare also weighted from 0 to 1. The scaleparameter controls the relative size of theimage segments (Baatz & Schaepe 2000;Benz et al. 2004).

The images were segmented at 2 levels:(1) a finer multiresolution segmentation(level 1) with scale parameter 50, colour/shape of 0.9/0.1, and smoothness/compact-ness of 0.5/0.5, and (2) a coarser spectraldifference segmentation (level 2) using amaximum spectral difference of 150. Thespectral difference segmentation aggregatesadjacent image objects with similar spectralresponses up to the maximum spectraldifference. This allows for combining ad-jacent image objects with similar spectralproperties into larger objects (such as largebare areas) while maintaining small spec-trally distinct objects (such as small patchesof vegetation) within the bare areas. Weused expert judgment and visual interpreta-tion to determine the segmentation para-meters. The classifications were performedon the coarser level 2 segmentation.

The objective of this research was todevelop a classification approach suitablefor multiple images and different vegetation

communities, therefore classes at the top ofthe hierarchy (left in Figure 3) were com-mon to many sites and generally easier todefine, and classes at the bottom of thehierarchy (right in Figure 3) were specific toeach site (species-level) and generally moredifficult to define. The scheme combinedrule-based classification at the top of thehierarchy, and nearest-neighbour classifica-tion techniques at the bottom. Shadow/non-shadow, vegetation/non-vegetation,and structure group classes (shrubs, grasses,forbs) were defined with rules and user-defined thresholds.

The suitability of the threshold valuesthat separate two classes in question wasassessed with visual interpretation. Thefeature view tool was used, which allowedfor visualizing object values for a givenfeature as a grey-scale image. Classificationof litter, bare ground, rocks, grass and shrubspecies was performed using image-derivedsamples and nearest-neighbour classifica-tion. In sites where litter and/or rocks were

Figure 3. The hierarchical classificationapproach used in the study. Boxes with lightbackground show classes (i.e., land cover types) inbold that were defined using rules with thresholds.Features used for the rules are shown in boxeswith dashed outlines (Brightness and NormalizedDifference Vegetation Index [NDVI]). Boxes withgrey background show classes in bold that weredefined using samples and a nearest neighbourclassification approach. Please see Table 1 forfurther description of features.


absent or could not be detected, the non-vegetation class represented bare ground.Likewise, grasses and forbs were grouped ifthey could not be distinguished separately.The brightness of digital numbers (DNs)and the normalized difference vegetationindex (NDVI) (Rouse et al. 1974) were usedas rule-based features. For the nearest-neighbour classification, the means andratios of the three bands and NDVI wereused. The spectral features were used,because they showed the most discrimina-tion for the vegetation of interest. Thefeature roundness was included for siteswhere broom snakeweed was detected be-cause of the distinctly round growth form ofthis sub-shrub. The features used in theanalysis are described further in Table 1.

In Definiens Developer, the structureand flow control of the entire image analysisprocedure (i.e., segmentation, rule-basedevelopment, features and samples used,classification) is defined in a process tree.The process tree can be saved and appliedto another image, and parameters can bemodified easily. This allows for consistencyin the analysis and rapid application tomultiple images, while still allowingfor customization to specific sites. Wedeveloped an initial process tree on the firstimage of a vegetation community andapplied it to the next two images in thesame vegetation community. The sameinitial process tree was then applied to the

next vegetation community and so on. Thiswas done to save time and take advantageof similarities in vegetation within a com-munity as well as across different commu-nities, although the process tree requiredmore changes across different vegetationcommunities compared to within the samecommunity. For each image, the followingparameters could be changed, although thesame threshold values were used for differ-ent images if the results were deemedsuitable:

. threshold value for brightness todifferentiate shadow and non-shadow

. threshold value for NDVI to differ-entiate vegetation and non-vegetation

. threshold value for NDVI to differ-entiate shrubs, grasses, and forbs

. selection of samples for shrub species,grass species, litter, bare ground, androcks which were unique to eachimage.

Statistical analysis

Ground-based LPI measurements and im-age-derived measurements were expressedas percent cover. For the image analysisproduct, we summed the pixels for eachland cover class and expressed the pro-portion of each class as percent cover.Regression analysis was used to quantifythe relationship between image- and

Table 1. Object-based features derived from Definiens Developer and used in the analysis. Thecolumn ‘Use’ describes use of the feature in rule-based (RB) or nearest neighbor classification (NN).Features were calculated for each image object at the level 2 segmentation and are based on digitalnumber (DN) values.

Feature Use Description

Mean NN Mean of pixels in image object. Computed for NIR, R, and G bandsRatio NN Band mean value of image object divided by sum of all band mean values.

Computed for NIR, R, and G bandsBrightness RB Sum of mean values of NIR, R, G divided by 3NDVI RB/NN (Mean NIR – mean R)/(mean NIR þ mean R)Roundness NN Radius of largest enclosing ellipse minus radius of smallest enclosing ellipse


ground-based measurements of percentcover at the structure group level. Inaddition, paired t-tests (n ¼ 15 plots) wereperformed to evaluate whether there werestatistically distinguishable differences be-tween estimates of land cover from LPI(field) and classified imagery. Species leveldata were summarized in tabular format.

3. Results

Segmentation results

Mean object sizes and standard deviationswere relatively consistent for the level 1segmentation (Table 2). The level 2 or spec-tral difference segmentation showed a great-er range in mean object size as well asstandard deviation due to the nature of thespectral difference segmentation. Vegeta-tion heterogeneity across sites resulted inretention of smaller, spectrally distinctobjects, such as small shrubs, while aggre-gating larger bare areas.

The classification approach

Using the same process tree on multipleimages proved to be an efficient approachfor classifying vegetation with high-resolu-tion digital aerial imagery. Approximately 8hours was spent developing the initialprocess tree. This task involved testingsegmentation parameters and determiningsuitable features for the analysis. On sub-sequent images, we executed the seg-mentation and initial classification to thestructure group level by using the samethresholds used on the previous image. Thisinitial run was completed in a few seconds.Upon visual inspection, the threshold levelsfor shadow/non-shadow, vegetation/non-vegetation, and shrubs/grasses/forbs wereadjusted. On average, adjusting thresholdlevels took approximately 20 minutes perimage. Further classification to the specieslevel involved selection of samples for eachspecies. Species-specific sample selection

Table 2. Number of objects, mean object sizes (in m2) and standard deviations (SD) (in m2) derivedfrom segmentations at two scales for fifteen images of 70 m 6 70 m plots. The first letter of each plotdenotes the vegetation community (mesquite [M], tarbush [T], creosote [C], grassland [G], playa [P]).The segmentations included a multiresolution segmentation (level 1) with scale parameter 50, colour/shape of 0.9/0.1, and smoothness/compactness of 0.5/0.5, and a spectral difference segmentation (level2) using a maximum spectral difference of 150.

Plot

Level 1 Level 2

No. of objectsMean object

size Object SD No. of objectsMean object

size Object SD

MNORT 79823 0.07 0.07 59842 0.09 0.77MRABB 97841 0.08 0.09 84412 0.12 0.90MWELL 72584 0.05 0.07 48106 0.14 1.05TTAYL 50849 0.08 0.08 24325 0.13 1.11TEAST 49010 0.07 0.07 33886 0.12 0.95TWEST 50236 0.06 0.07 32599 0.11 0.88CCALI 24222 0.08 0.09 20924 0.13 0.78CGRAV 44543 0.07 0.08 35967 0.12 1.13CSAND 28117 0.08 0.09 23451 0.12 0.86PCOLL 78520 0.05 0.07 58456 0.13 0.99PSMAL 72598 0.05 0.06 53258 0.11 1.04PTOBO 65230 0.06 0.07 42051 0.13 0.79GSUMM 33369 0.07 0.07 21436 0.11 0.60GIBPE 67812 0.08 0.08 44490 0.12 1.30GBASN 85581 0.07 0.07 66129 0.09 0.39


was most time-consuming, taking from halfan hour to an hour depending on vegetationcomplexity and diversity. The forb class wasmerged with the grass class, because forbsand grasses were often intermixed and mostforbs were too small to be consistentlydetected in the imagery.

In order to assess the consistency inthreshold values from image to image, therange of values (in DN) of a feature (e.g.,brightness, NDVI) were plotted and thethreshold values (chosen by visual inter-pretation) used to separate shadow fromnon-shadow (Figure 4A) and vegetationfrom non-vegetation (Figure 4B). While theplot-based sample size was small (threeplots per vegetation community), it wasobserved that the threshold values wereconsistent for each vegetation community.For example, the same brightness thresholdvalue was used to separate shadow fromnon-shadow for all plots in the creosote-dominated communities (CCALI, CGRAV,CSAND). In addition, communities withno shrubs (all playa sites) had a consider-ably higher brightness threshold than theshrub-dominated communities.

Classification to structure group level

The r-square values for image- and ground-based measures of percent cover wererelatively high, ranging from 0.823 to0.924 for the shrub, non-vegetation, totalvegetation and grass/forb classes (Figure 5).Paired t-tests indicated statistical differencesbetween the two methods in which theimage classification resulted in higher coverestimates for grass/forb (4.5%, p ¼ 0.012)and shrubs (2.9%, p ¼ 0.038). However,cover estimates for non-vegetation andtotal vegetation cover were statisticallycomparable (p ¼ 0.302 and p ¼ 0.618, re-spectively). Coefficients of variation werecomparable, indicating a similar variabilityfor image- and ground based estimates ofcover (Table 3).

Classification to species level

Classification to the species level yieldedmixed results. We were able to classifydominant shrubs to the species level inevery plot where they occurred, but onlytwo prominent grass species, black grama

Figure 4. Choice of thresholds (black and whitedots) in relation to range of DN values (verticallines) for rule-based classification for the 15plots. (A) shows the feature brightness used toseparate shadow from non-shadow (Shadow/non-shadow). (B) shows the feature NDVI usedto separate vegetation from non-vegetation(Vegetation/non-vegetation). All thresholdswere chosen by visual interpretation. The threeplots in each of the five vegetation communities(identified below the plot names) are groupedtogether (alternating groups of three black andthree white dots). Brightness and NDVI valuesare derived from digital numbers (DN).


Table 3. Results for paired t-tests of differences in percent cover estimates derived from linepoint intercept (LPI) and image analysis (DEF) methods for 70 m 6 70 m plots (n¼12 for shrubs,n¼15 for all others [three plots had no shrubs]). The mean differences, 95 percent confidenceintervals (CI), the means for LPI and DEF, and the coefficients of variation (CV) are shown inpercent.

Mean CV

t-statistic p-value Mean diff. 95% CI DEF LPI DEF LPI

Non-vegetation 71.07 0.302 72.0 75.7–1.7 43.2 45.2 42.0 39.3Total vegetation 0.51 0.618 0.9 72.7–4.5 55.7 54.8 32.8 32.4Grass/forb 2.89 0.012 4.5 1.4–7.5 28.3 23.8 73.1 75.5Shrubs 2.36 0.038 2.9 0.5–5.2 17.1 14.2 57.9 56.2

Figure 5. Scatter plots for estimates of percent cover obtained with the ground-based line pointintercept method (LPI), and from object-based image analysis (DEF) for fifteen 70 m 6 70 m plots.Non-vegetation consists of bare ground, rocks and litter. Total vegetation includes litter that isintermixed with vegetation and cannot be detected.


and tobosa, were adequately represented inthe imagery. Estimates of percent coverfrom LPI and image analysis for shrub andshrub-like species are compared in Table 4.While there was some variability in theresults, the following general observationscan be made: (1) shrubs had higherestimates with image analysis, especiallylarge shrubs such as mesquite. With LPI,every gap in the shrub canopy is recorded,while with image analysis, the entire canopyis generally classified as shrub; (2) small

shrubs, such as snakeweed and some salt-bush had lower estimates, probably becausesome of them were missed due to their size(shrubs 512 cm 6 12 cm); (3) shrub spe-cies with percent cover of 1–2 percent basedon LPI were generally not detected withimage analysis, although there were someexceptions in the grass sites (denotedwith G as first letter of the plot name inTable 4). The classification of the grasslandplot GIBPE is shown in Figure 6. Thelegend shows the hierarchical classification

Table 4. Cover estimates from line point intercept (LPI) and image analysis (DEF) and differences inpercent cover between the methods (Diff. ¼ DEF-LPI) for shrubs and shrub-like species for 12 plots (incapital letters). The first letter of each plot denotes the vegetation community (mesquite [M], tarbush[T], creosote [C], grassland [G]).

LPI DEF Diff. LPI DEF Diff.

MNORT TTAYLShrubs total 23.20 31.41 þ8.21 Shrubs total 14.64 19.89 þ5.25Mesquite 18.57 26.60 þ8.03 Tarbush 14.28 19.89 þ5.61Saltbush 4.27 3.97 70.30 Wolfberry 0.36Snakeweed 0.36 0.85 þ0.49MWELL TEASTShrubs total 21.43 23.82 þ2.39 Shrubs total 10.71 13.94 þ3.23Mesquite 15.00 21.65 þ6.65 Tarbush 9.28 13.94 þ4.66Snakeweed 5.71 0.85 74.86 Wolfberry 1.07Saltbush 0.36 1.32 þ0.96 Mesquite 0.36Ephedra 0.36

MRABB TWESTShrubs total 20.36 27.23 þ6.87 Shrubs total 12.86 9.34 73.52Mesquite 11.79 22.96 þ11.17 Tarbush 12.50 9.34 73.16Saltbush 5.00 2.51 72.49 Wolfberry 0.36Snakeweed 3.57 1.76 71.81

CSAND GBASNShrubs total 23.93 24.29 þ0.36 Shrubs total 3.21 2.58 70.63Creosote 21.43 24.29 þ2.86 Mesquite 1.07 1.27 þ0.20Mesquite 2.14 Snakeweed 1.07 0.94 70.13Ephedra 0.36 Yucca 1.07 0.37 70.70

CGRAV GIBPEShrubs total 19.64 19.75 þ0.11 Shrubs total 9.29 8.29 71.00Creosote 16.07 17.87 þ1.80 Mesquite 3.57 4.94 þ1.37Mesquite 1.79 1.88 þ0.09 Snakeweed 4.29 2.59 71.70Sumac 1.07 Yucca 0.36 0.76 þ0.40Tarbush 0.71 Ephedra 1.07

CCALI GSUMMShrubs total 21.43 20.11 71.32 Shrubs total 2.14 2.82 þ0.68Creosote 21.43 20.11 71.32 Creosote 1.42 2.16 þ0.74

Yucca 0.36 0.66 þ0.30Ephedra 0.36


approach, which allows for collapsing theclasses to the structure group level or to thebare/vegetation level.

4. Discussion and conclusions

In this study, we compared image andground-based measures of percent coverand assessed the viability and efficiency ofapplying the image-based method to siteswith a wide range in vegetation compositionand structure. The OBIA approach per-formed well using sub-decimetre resolutionimagery. Generalizing groups of pixels intosegments is nearly essential for meaningfulclassification of this imagery due to the highdegree of spectral variability within a singleshrub canopy or patch of vegetation. Highcorrelations were obtained with ground-based measurements for total vegetation,shrubs, grasses and non-vegetated surfaces.At the species level, grasses were difficult todelineate and large shrub species exhibitedthe greatest disparity between estimatedcover for field- and image-based methods.Image estimates were consistently higherthan LPI field estimates presumably due todifferences in field perception of gap frac-tion and nadir depictions of vegetation onwhich the image analysis was based.

While the radiometric and spatial reso-lution of the DMC digital imagery washigh, difficulties were still encountered inmapping certain species. Grasses wereespecially difficult to separate by speciesfor several reasons. First, grass species suchas dropseed and threeawn are bunchgrassesand individual plants or patches are difficultto identify due to their small size. Second,many grass species grow in mixed standswhich are not spectrally distinct. Third, inthis semi-arid environment, many patchescontain both senescent and photosynthetic(i.e., green) components of the same species.Because we did not have spatially explicittraining sites to delineate patches by spe-cies, we had to rely on information thatcould be extracted visually from the ima-gery; only black grama and tobosa grassesexhibited unique spectral and texturalproperties that permitted identification byspecies.

Due to the relatively recent introductionof digital aerial mapping cameras, very fewstudies are available for comparison, andthose available tend to employ coarserresolution imagery. Rosso et al. (2008)compared four digital airborne sensors forland surface mapping with imagery rangingfrom 10 to 40 cm spatial resolution, but

Figure 6. DMC colour-infrared image (left) and classification (right) of the grassland plot GIBPE.


their objective was to evaluate and comparethe performance of the sensors for landcover mapping, not to evaluate transfer-ability of non-traditional image classifica-tion algorithms. Stow et al. (2008)quantified change in a 1 km2 shrublandhabitat preserve in southern Californiaover 8 years using 1 m resolution digitalaerial imagery. Our objective was to developand evaluate a semi-automated image clas-sification protocol with sub-decimetre reso-lution digital imagery to permit analysis ofmultiple sites across a range of land coverconditions.

This research determined that the appli-cation of a well-developed process treeiteratively to subsequent images had severaladvantages. Those included faster proces-sing of multiple images, ease of comparisonof results from different sites, and the abilityto expand or collapse the hierarchicalclassification scheme for management deci-sions at different levels of detail (i.e.vegetation/non-vegetation, structure groupor species level). For example, an estimateof the amount of non-vegetation at the plotlevel is related to bare ground, which is animportant indicator of rangeland healthbecause of the potential for wind and watererosion in areas of exposed soil (Pellantet al. 2005).

A potential disadvantage of transferringan image analysis approach to multipleimages is that the procedure may notperform consistently across all vegetationcommunities. For example, the separationof vegetation and non-vegetation usingNDVI may not be suitable if the proportionof senescent vegetation is high. A differentsegmentation scale may also be necessaryfor certain sites. Because all vegetationcommunities in this study were located inthe Chihuahuan desert consisting of differ-ent arrangements of similar species withcomparable spectral properties, the segmen-tation and workflow parameters were broadenough to apply to all images. With this

high-resolution imagery, the appropriatesegmentation scale was relatively easy toidentify visually. If this technique wasapplied to sites in other eco-regions, weanticipate that different process trees foreach of these sites would probably have tobe implemented.

A prime motivation for this study wasto develop an approach for the NRI forimage-based assessments of plots overspatially extensive and often remote areasthat may not be visited in the field. There-fore, we had to rely on visual interpretationof structure groups/species in the plot.While the species list from LPI was avail-able to us for image analysis, LPI data arenot designed for validating remotely sensedimagery but represent a repeatable methodto quantify changes in land cover condi-tions. Derivation of image sample objectsfor object-based classification is not possi-ble using LPI data that correspond to adifferent scale of observation. LPI datarepresent very fine-scale measurements;LPI measurements are derived by droppinga pin at specified intervals along a linetransect to record multiple levels of vegeta-tion and surface components that the pinhits. If the pin hits a single blade of grassover a bare patch, it is recorded as grass. InOBIA, the same location would in alllikelihood be classified as a bare patch.This discrepancy highlights existing chal-lenges relating LPI information to remotelysensed imagery.

We believe additional ground measure-ments are needed and suggest the collec-tion of spatially explicit training samples,consisting of polygons depicting specificspecies or vegetation patches. These poly-gons are more suitable to OBIA thanpoint-level field measurements (Laliberteet al. 2007). The issue of identifyingappropriate image samples is confoundedby the challenge of co-registering spatiallyexplicit field data to sub-decimetre imagerywith sufficient location accuracy. Even


differentially corrected GPS data have sub-metre levels of error, which is relativelylarge when those data are overlaid on 4 cmresolution imagery. However, given theassumed errors, it is easier to determinewhether a polygon is off with regard to theimage than determining the location of apoint. It is planned to incorporate GPS-based training samples in on-going andfuture mapping studies.

This investigation of threshold cutoffvalues across images and vegetation com-munities proved informative for this studyand will be tested further with nationwideimagery. Because these images were rathersmall in spatial extent (70 m 6 70 m), itwas relatively easy to assess the suitabilityof a threshold value by roaming over theentire image of the plot. For an image witha larger extent this task would be moredifficult. Therefore, we believe that thebroad guidelines for threshold cutoffvalues for certain vegetation communities(Figure 4) could be applied to larger images.The high radiometric resolution of digitalaerial imagery makes such guidelines useful,because there is greater difficulty in deter-mining a threshold value with 12- or 16-bitdigital imagery than with 8-bit film-baseddigital aerial photography.

It has been demostrated that this OBIAtechnique represents a viable approach forquantifying vegetation cover in differentvegetation communities in the NorthernChihuahuan desert. While the level of detailis less than that of ground-based measure-ments, an image analysis approach resultsin complete classification of the plot,providing baseline information for subse-quent studies of land cover change. Due tothe transferability of the process tree toadditional images or sites, the image analy-sis approach is more efficient than ground-based measures, especially at the structuregroup level. While initial algorithm devel-opment requires time, subsequent applica-tions of the process tree are relatively fast.

In a previous study, we determined that thebreakeven point in time allocated to obtain-ing ground-based versus image-based mea-sures was eight plots (Laliberte et al., inpress). Proximity of plots and time spent onorthorectification of imagery was similar inthis study, therefore the time estimates arecomparable. The cost of imagery should notbe included in a comparison, becauseimagery is being acquired already for NRIapplications.

The ability to classify multiple imagesefficiently offers the potential to increase theprecision of national level inventories byincreasing sample locations and to reducecosts by requiring fewer personnel to obtainground measurements. We are currentlyimplementing the same OBIA approach fora nationwide study in an even broaderrange of vegetation communities in grazinglands, and are developing additional toolsand techniques for potential integration ofthese techniques into the NIR CEAPGrazing Lands.

Acknowledgements

This research was funded by the USDANatural Resources Conservation Service insupport of the Conservation Effects AssessmentProject, and by the USDA Agricultural Re-search Service and the National Science Foun-dation Long-Term Ecological ResearchProgram, Jornada Basin V: Landscape Lin-kages in Arid and Semiarid Ecosystems. Theauthors would like to thank Debra Peters andJohn Anderson for access to the LTER NPPsites, and Hector Godinez-Alvarez, MichelleMattocks, David Toledo and Justin Van Zeefor providing assistance with field data collec-tion. Additional thanks go to the three anon-ymous reviewers.

References

Baatz, M., & Schaepe, A. (2000) Multiresolutionsegmentation: an optimization approachfor high quality multi-scale image segmen-tation. In: Strobl, J. & Blaschke, T., eds.Angewandte Geographische Informationsver-arbeitung, vol. XII, Heidelberg, Wichmann,pp. 12–3.


Benz, U.C., Hoffmann, P., Willhauck, G.,Lingenfelder, I. & Heynen, M. (2004)Multi-resolution, object-oriented fuzzy ana-lysis of remote sensing data for GIS-readyinformation. ISPRS Journal of Photogram-metry and Remote Sensing, vol. 58, pp. 239–258.

Booth, D.T., & Cox, S.E. (2008) Image-basedmonitoring to measure ecological change inrangeland. Frontiers in Ecology and Environ-ment, vol. 6, no. 4, pp. 185–190.

Coulter, L.L., & Stow, D.A. (2008) Assessmentof the spatial co-registration of multi-temporal imagery from large formatdigital cameras in the context of detailedchange detection. Sensors, vol. 8, pp. 2161–2173.

Definiens (2007) Definiens Developer 7.0 UserGuide. Definiens AG, Munich.

Duriancik, L.F., Bucks, D., Dobrowolski, J.P.,Drewes, T., Eckles, S.D., Jolley, L., Kellogg,R.L., Lund, D., Makuch, J.R., O’Neill,M.P., Rewa, C.A., Walbridge, M.R., Parry,R., & Weltz, M.A. (2008) The first five yearsof the Conservation Effects AssessmentProject. Journal of Soil and Water Conserva-tion, vol. 63, no. 6, pp. 185–197.

Gibbens, R.P., McNeely, R.P., Havstad, K.M.,Beck R.F., & Nolen, B. (2005) Vegetationchanges in the Jornada Basin from 1858 to1998. Journal of Arid Environments, vol. 61,pp. 651–668.

Godinez-Alvarez, H., Herrick, J.E., Mattocks,M., Toledo, D., & Van Zee, J. (2009)Comparison of three vegetation monitoringmethods: their relative utility for ecologicalassessment and monitoring. Ecological In-dicators, vol. 9, pp. 1001–1008.

Green, K., & Lopez, C. (2007) Using object-oriented classification of ADS40 to mapbenthic habitats of the state of Texas.Photogrammetric Engineering and RemoteSensing, vol. 73, pp. 861–865.

Herrick, J.E., Van Zee, J.W., Havstad, K.M,Burkett, L.M. & Whitford, W.G. (2005)Monitoring Manual for Grassland, Shrublandand Savanna Ecosystems. Volume I: QuickStart, and Volume II: Design, SupplementaryMethods and Interpretation, Las Cruces,NM, USDA-ARS Jornada ExperimentalRange.

Hinz, A., Doerstel, C., & Heier, H. (2000)Digital modular camera: system conceptsand data processing workflow. InternationalArchives of Photogrammetry and RemoteSensing, vol. 33, no. B2.

Hodgson, M.E., Jensen, J.R., Tullis, J.A., Rior-dan, K.D., & Archer, C.M. (2003) Synergisticuse of lidar and color aerial photography formapping urban parcel imperviousness.Photogrammetric Engineering and RemoteSensing, vol. 69, no. 9, pp. 973–980.

Honkavaara, E., & Markelin, L. (2008) Radio-metric performance of digital image datacollection - a comparison of ADS40/DMC/UltraCam and EmergeDSS. In: Fritsch, D.,ed. Proceedings of the 51st PhotogrammetricWeek 2007, Sept. 2007, Stuttgart, Institutfuer Photogrammetrie, Universitaet Stutt-gart, Germany.

Laliberte, A.S., & Rango, A. (2009) Texture andscale in object-based analysis of sub-deci-meter resolution unmanned aerial vehicle(UAV) imagery. IEEE Transactions ofGeoscience and Remote Sensing, vol. 47, no.3, pp. 761–770.

Laliberte, A.S., Rango, A., Havstad, K.M.,Paris, J.F., Beck, R.F., McNeely R., &Gonzalez, A.L. (2004) Object-oriented imageanalysis for mapping shrub encroachmentfrom 1937-2003 in southern New Mexico.Remote Sensing of Environment, vol. 93, pp.198–210.

Laliberte, A.S., Fredrickson, E.L., & Rango, A.(2007) Combining decision trees with hier-archical object-oriented image analysis formapping arid rangelands. PhotogrammetricEngineering and Remote Sensing, vol. 73, no.2, pp. 197–207.

Laliberte, A.S., Herrick, J.E., Rango, A., &Winters, C. Acquisition, orthorectification,and object-based classification of unmannedaerial vehicle (UAV) imagery for rangelandmonitoring.PhotogrammetricEngineering andRemote Sensing, vol. 76, no. 6, pp. 661–672.

Ludwig, J.A., Bastin, G.N., Chewings, V.H.,Eager, R.W., & Liedloff, A.C. (2007) Leaki-ness: a new index for monitoring the healthof arid and semiarid landscapes using remo-tely sensed vegetation cover and elevationdata. Ecological Indicators, vol. 7, no. 2, pp.442–454.

Neumann, K.J. (2008) Trends for digital aerialmapping cameras. International Archives ofthe Photogrammetry, Remote Sensing andSpatial Information Sciences, vol. XXXVIIPart B1, pp. 551–554.

Nusser, S.M., & Goebel, J.J. (1997) TheNational Resources Inventory: a long termmulti resource monitoring programme. En-vironmental and Ecological Statistics, vol. 4,pp. 181–204.


Pellant, M., Shaver, P., Pyke, D., & Herrick, J.E.(2005) Interpreting Indicators of RangelandHealth, version 4, Interagency TechnicalReference, Bureau of Land Management,Denver, CO.

Platt, R.V., & Rapoza, L. (2008) An evaluationof an object-oriented paradigm for land use/land cover classification. Professional Geo-grapher, vol. 60, pp. 87–100.

Rosso, P.H., Klonus, S., Ehlers, M., & Tschach,E. (2008) Comparative properties of fourairborne sensors and their applicability toland surface interpretation. InternationalArchives of Photogrammetry, Remote Sensingand Spatial Information Sciences, vol.XXXVII, Part B1, pp. 545–550.

Rouse, J.W., Haas, R.H., Schell, J.A., &Deering, D.W. (1974) Monitoring vegetationsystems in the Great Plains with ERTS. In:Freden, S.C., Mercanti, E.P., & Becker,M.A., eds. Third Earth Resource TechnologySatellite (ERTS) Symposium, NASA, Wa-shington DC, pp. 309–317.

Sankey, T.T., Moffet, C., & Weber, K. (2008)Postfire recovery of sagebrush communities:assessment using SPOT-5 and very large-scale aerial imagery. Rangeland Ecology andManagement, vol. 61, pp. 598–604.

Stow, D., Hamada, Y., Coulter, L., & Zlatina,A. (2008) Monitoring shrubland habitatchanges through object-based change identi-fication with airborne multispectral imagery.Remote Sensing of Environment, vol. 112, pp.1051–1061.

Wainwright, J. (2006) Climate and Climatolo-gical Variations in the Jornada Basin. In:Havstad, K.M., Huennecke, L.F., & Schle-singer, W.H., eds. Structure and Function ofa Chihuahuan Desert Ecosystem. The Jorna-da Basin Long-Term Ecological ResearchSite, Oxford, Oxford University Press, pp.44–80.

Weber, K.T. (2006) Challenges of integratinggeospatial technologies into rangelandresearch and management. RangelandEcology and Management, vol. 59, pp.38–43.

Yu, Q., Gong, P., Clinton, N., Biging, G., Kelly,M., & Schirokauer, D. (2006) Object-based detailed vegetation classification withairborne high spatial resolution remotesensing imagery. Photogrammetric Engineer-ing and Remote Sensing, vol. 72, no. 7, pp.799–811.


Copyright of Journal of Spatial Science is the property of Taylor & Francis Ltd and its content may not be

copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written

permission. However, users may print, download, or email articles for individual use.

Hierarchical object-based classiﬁcation of ultra-high ...

Documents

Transcript of Hierarchical object-based classiﬁcation of ultra-high ...