9 theory and applications of object-based image …...ping wetlands using remotely sensed and...

175

9 Theory and Applications of Object-Based Image Analysis and Emerging Methods in Wetland Mapping

Joseph F. Knight, Jennifer M. Corcoran, Lian P. Rampi, and Keith C. Pelletier

IntroductIon

This chapter addresses object-based image analysis (OBIA) and other emerging methods for map-ping wetlands using remotely sensed and ancillary data. This treatment includes an introduction to OBIA and decision tree techniques, a history of the development of OBIA methods in wetland map-ping, and four case studies describing recent wetland mapping research conducted in the Remote Sensing and Geospatial Analysis Laboratory at the University of Minnesota. The chapter ends with a discussion of lessons learned and possibilities for future applications.

oBIA BAckground

Traditional approaches to mapping land cover are photointerpretation and semiautomated per-pixel classifiers. Human analysts use the elements of image interpretation (EII) for delineating and label-ing features such as buildings, transportation infrastructure, agricultural fields, forests, wetlands, soils, and grasslands (Olson 1960). While these maps can be detailed and accurate, the process is labor intensive, which limits the frequency of updates (Goetz et al. 2003; Van and Bao 2010). Per-pixel approaches rely on spectral information to either aggregate statistically similar pixels for class labels (e.g., unsupervised) or use training data to group similar pixels into classes reflecting the training data (e.g., supervised).

Unlike per-pixel approaches that focus on classification of individual pixels in an image, object-based analyses delineate and classify homogeneous groups of pixels or objects using spectra, shape, size, texture, and spatial context, producing land cover classes that better approximate real-world fea-tures (Benz et al. 2004; Blaschke 2003, 2010; Blaschke and Strobl 2001; Kettig and Landgrebe 1976).

AQ1

contents

Introduction .................................................................................................................................... 175OBIA Background ......................................................................................................................... 175Decision Tree Classifier Background ............................................................................................. 176History of OBIA Methods in Wetland Mapping ............................................................................ 177Case Studies ................................................................................................................................... 180Discussion and Conclusion ............................................................................................................ 188References ...................................................................................................................................... 189

K23165_C009.indd 175 10/28/2014 7:27:53 PM

176 Remote Sensing of Wetlands: Applications and Advances

These efforts integrate geospatial data from different sensors with varying resolutions and vector data (Benz et al. 2004). The ability to incorporate disparate such data (e.g., multiple resolutions and GIS layers) into object-based classification often results in higher mapping accuracy compared to traditional pixel-based classification techniques (Baatz and Schape 2000; Benz et al. 2004; Blaschke 2010; Bock et al. 2005; Haralick and Shapiro 1985).

An object-based approach starts with the segmentation process, in which an image is partitioned into image objects based on color, shape, and size. These image object primitives contain attributes based on spectra, geometry, texture, and context. These attributes are similar to the information employed by human interpreters using the EII. Classes are defined by creating fuzzy ranges or thresholds, which are applied to the image objects. The process is iterative, in which some or all of the image object primitives are classified and then merged or further segmented to delineate addi-tional classes. An iterative approach to segmentation and classification reduces oversegmentation and undersegmentation, resulting in objects that correspond to the boundaries of land cover features (Witharana et al. 2014). Object-based approaches have been successfully used for mapping different land cover features such as transportation infrastructure (Nobrega et al. 2008; Repaka and Truax 2004), urban infrastructure and vegetation (Frauman and Wolff 2005; Harayama and Jaquet, 2004; O’Neil-Dunne et al. 2012; Syed et al. 2005; Zhou and Troy 2008), wildlife habitat (Blaschke 2010; Bock et al. 2005), land use (Kressler and Steinnocher 2008), buildings (Chen et al. 2006; Meng and Peng 2009), and forests (Kim et al. 2009).

Context can be used for classifying features based on landscape patterns or distances between objects. As additional classes are identified, the contextual information increases such that areas can be defined based on pattern; that is, an area with few buildings and roads suggests a rural landscape, while an area with numerous buildings and roads suggests an urban landscape (MacFaden et al. 2012). The areas can be classified as urban and not urban, while the objects within these classes are reclassified based on spectral geometry and additional context. For example, a recently plowed or harvested field can have similar spectral and geometric properties as impervious surfaces. However, given the landscape context (e.g., rural), the objects are classified as agricultural fields rather than large parking areas.

Context can be also used to classify land cover features based on topological relationships or shared boundaries between objects. The relative border or percentage of common boundary between classes can be used to separate or aggregate neighboring features. Impervious surfaces found in urban infrastructure often have similar spectral and geometric characteristics as exposed soils or barren areas. For example, both features often have high reflectance with long and narrow shape. The spectral and geometric similarities could result in stream banks or shorelines being misclassified as impervious surfaces. The spatial context or shared border with objects classified as water allows for reclassifying these features as exposed soils rather than impervious features.

Wetland classes are challenging to classify given the need to characterize vegetation, hydrol-ogy, and soils for accurately identifying wetland boundaries and differentiating wetland features from similar upland features. Traditional wetland maps have been produced using photointerpre-tation (e.g., Tiner 1990) or pixel-based classifiers (e.g., Ozesmi and Bauer 2002). Object-based approaches for mapping wetlands have integrated optical data, LiDAR-derived surface layers, and vector data. Recent efforts using LiDAR or integrating LiDAR and high-resolution optical data have been shown to improve identification of wetlands (e.g., Corcoran et al. 2012; Kim et al. 2011; Knight et al. 2013; Lang and McCarty 2009; Lang et al. 2012, 2013; Maxa and Bolstad 2009; Rampi et al. 2014a).

decIsIon tree clAssIfIer BAckground

Many researchers report that decision tree classifiers can improve wetland mapping accuracy over traditional approaches such as maximum likelihood classifier (MLC; Baker et al. 2006; Hogg and Todd 2007; Liu et al. 2008; Rodriguez-Galiano et al. 2012; Rover et al. 2011). A decision tree

AQ2

K23165_C009.indd 176 10/28/2014 7:27:53 PM

177Theory and Applications of Object-Based Image Analysis

(e.g., Random Forest [RF], CART, and See5) is a supervised classification algorithm that works by creating decision points, or nodes, based on the diagnostic features of the input variables. Combining these decision points in a set of sequential nodes results in the “growing” of a “tree” of nodes that can be used to partition the input data into classes of interest (leaves). RF extends this approach by creating a “forest” of hundreds of trees, each of which has a “vote” as to the correct class for a given case (e.g., a pixel). In this ensemble approach, the class chosen by the majority of the trees in the forest is selected. RF uses out-of-bag (OOB) sampling to compute a measure of internal cross-validation accuracy, in addition to any independent reference data–based accuracy assessment. Decision tree classifiers are attractive for land cover mapping, in general, and wetland mapping, in particular, because they require no assumptions about the underlying distributions of the input datasets, are able to use both continuous and categorical data, and can be resistant to overfitting (especially RF). Also, decision trees may help to reduce the subjectivity involved in defining OBIA parameters for wetland mapping (Breiman 2001; Friedl and Brodley 1997; Quinlan 1993; Smith 2010).

A prominent example of a decision tree classification approach is that used to create the 2001 version of the National Land Cover Data (NLCD; Homer et al. 2007). The decision tree used inputs a wide variety of data types: multidate Landsat imagery, topographic derivatives, impervious sur-face maps, wetland maps, and other ancillary datasets. Though decision tree use in land cover map-ping is now common, use in the OBIA context is developing but shows great promise (Corcoran et al. 2013, 2011; Smith 2010).

HIstory of oBIA MetHods In WetlAnd MAppIng

The history of OBIA methods used in wetland and aquatic vegetation mapping is relatively brief. Though image segmentation based on pixel spectral response was a technique introduced by Kettig and Landgrebe (1976) in the form of their “extraction and classification of homogeneous objects” (ECHO) algorithm, computing speed, software availability, and algorithmic sophistication did not allow for widespread adoption of what we now call OBIA techniques until the early 2000s. The first peer-reviewed study using OBIA to map wetlands was Costa et al. (2002), which used a spec-tral value segmentation and subsequent region-based classification of Radarsat and JERS-1 radar images to map Amazon floodplain communities. The authors reported overall accuracies of 97% for classification of aquatic vegetation, flooded forest, and unflooded forest. However, the use of image segmentation was not the focus of the study. The authors concerned themselves more with the per-formance of radar imagery for vegetation mapping. Their innovative use of image segmentation for mapping wetland vegetation was not discussed.

In a similar study, Hess et al. (2003) used spectral-based image segmentation of JERS-1 radar imagery to map wetlands in the Amazon basin. Once derived, the segments were classified using an unsupervised approach based on Mahalanobis distance followed by manual editing of “outli-ers.” The authors reported 95% wetland/nonwetland mapping accuracy and suggested that, with a combination of automated classifiers and interpreter editing, image segmentation is well suited for mapping wetlands at high spatial resolution.

In a pair of 2004 studies, Wang et al. (2004a,b) provided the first peer-reviewed reports of wetland mapping using image segmentation applied to high-resolution optical satellite imagery. The authors used IKONOS and QuickBird imagery to map mangroves on the Caribbean coast of Panama. The 2004a study describes the use of eCognition for object-based segmentation and classi-fication using MLC, but since the focus of the research was on comparing IKONOS and QuickBird for mangrove mapping, the authors did not provide much detail on the OBIA approach or its ben-efits. In contrast, Wang et al. (2004a) used IKONOS imagery to compare the suitability of three classifiers for mangrove mapping: pixel-based MLC, object-based nearest neighbor (NN), and a hybrid pixel and object-based classifier (MLCNN). Segmentation was done using the Bhattacharyya distance measure to group similar pixels. The hybrid MLCNN classifier outperformed the others

AQ4

AQ5

AQ3

K23165_C009.indd 177 10/28/2014 7:27:53 PM


(91% overall accuracy versus 89% for MLC and 80% for NN). The authors report a number of use-ful conclusions. The scale parameter is an important component of the segmentation process but is subjective. Using the Bhattacharya distance helped to optimize the selection of the scale parameter and subsequently confirm that separability of mangroves is improved at the object level over the pixel level. However, a pixel-based classifier may preserve “rich spectral information” in the image, while OBIA methods tend to generalize such information at the object level. Thus, when mapping areas with fine spatial reflectance variability, objects may contain mixed classes. The authors report that such mixed objects were unavoidable, particularly near object edges and class boundaries. The better performance of the hybrid classifier indicates that it is important to maintain pixel-based spectral information when spectral variability is high.

Harken and Sugumaran (2005) is the first peer-reviewed wetland mapping paper using OBIA with hyperspectral imagery and also the first to explicitly mention using textural measures in addi-tion to spectral information in the segmentation process. The authors used 0.6 m CASI airborne hyperspectral imagery in a comparison of two classification approaches: the pixel-based spectral angle mapper (SAM) and an OBIA approach using eCognition v3. The authors describe using a texture image to assist in segmenting the hyperspectral imagery, though little detail is provided on that process. Nine wetland and upland land cover classes were mapped, with the OBIA approach achieving 92% overall mapping accuracy versus 65% for the SAM. In anticipation of new directions in wetland mapping, the authors recommended evaluation of decision tree classifiers such as CART and multitemporal imagery as future work.

Fournier et al. (2007) and Grenier et al. (2007, 2008) jointly described the approach used in the Canadian Wetland Inventory (CWI). Fournier et al. laid out the general case for using OBIA meth-ods. Grenier et al. (2007, 2008) followed with case studies of OBIA methods for mapping CWI wet-land classes. The 2008 study used Landsat-7 multispectral and panchromatic bands with Radarsat-1 imagery in a multiscale segmentation approach—the first mention of multiscale segmentation in a wetland mapping paper. Supervised NN and fuzzy membership function (MF) approaches incor-porating spectral, textural, and shape measures were used. The MF approach is described as being superior due to its greater flexibility. The authors reported overall accuracies for their seven-class wetland maps as 68%–76%, depending on study area. Li and Chen (2011) described a more devel-oped rule-based version of a CWI mapping protocol.

Gilmore et al. (2008) provided the first study of LiDAR data in combination with optical imagery to map wetlands using OBIA. This study was also the first to attempt to map an invasive wetland spe-cies (Phragmites) with OBIA. The authors used multidate QuickBird imagery with LiDAR-derived vegetation height data to map marsh plant communities in Connecticut. The authors described the segmentation parameters used, but did not provide detail on the subsequent classification approach. Reported accuracies ranged from 63% to 97%, depending on species, with the higher accuracy applying to Phragmites. Seasonal variability was described as an important consideration when mapping marsh plant communities, so the authors recommended multitemporal imagery for such inventories.

By 2008, the basic elements of effective OBIA wetland mapping methods were in place: (1) segmentation based not just on spectral attributes but also on size, shape, texture, and context, (2) classification using object-derived measures, and (3) the use of multidate high-resolution optical imagery, radar, and LiDAR derivatives. Several subsequent studies have confirmed or extended these results. Conchedda et al. (2008) used multiresolution segmentation to map mangroves in Senegal using multidate (1986 and 2006) SPOT XS imagery. The object classification was per-formed using an NN classifier, with an overall accuracy estimate of 86% and a wetland class user’s accuracy of 97%. Shen et al. (2008) used multiresolution segmentation and MF classification of ENVISAT ASAR imagery to map inundated areas in the Poyang Lake, China, area. The authors compared their OBIA approach to an MLC and reported that accuracy estimates of inundation increased from 86% to as high as 95%. Dissanska et al. (2009) used multiresolution segmentation and MF classification of QuickBird panchromatic images to map peatlands in James Bay, Quebec,

AQ6

K23165_C009.indd 178 10/28/2014 7:27:53 PM


Canada. Peatland accuracy estimates were 95% for producer’s accuracy and 88% for users. The overall classification accuracy of their five classes was 88%.

Frohn et al. (2009) provide several recommendations for OBIA-based wetland mapping. The authors claimed to be the first to attempt to map isolated wetlands using satellite imagery. They used Landsat-7 imagery and ancillary data to map isolated wetlands of various sizes in Florida (United States) with an MF classification approach. Reported accuracy estimates for isolated wetlands varied depending on their size. For small wetlands (>0.5 acres), the user’s and producer’s accuracies were 88% and 89%, respectively. For larger wetlands (>2 acres), the user’s and producer’s accuracies were 97% and 95%, respectively. The authors noted that Landsat imagery is suitable to meet the U.S. Federal Geographic Data Committee (FGDC) wetland mapping accuracy standards. They recommended using the wettest Landsat scene of the year, reported that rainfall may not be a useful input to the classification process and suggested using Landsat Band 5 to determine moisture, and concluded that image segmentation of Landsat data has great potential for mapping isolated wetlands.

Several recent studies have compared pixel-based versus OBIA methods or assessed ways to optimize the performance of OBIA approaches. Laba et al. (2010) mapped invasive wetland plants in the Hudson River National Estuarine Research Reserve using IKONOS imagery. They reported that OBIA textural methods combined with per-pixel spectral information resulted in slight increases in accuracy over maximum likelihood alone (78% vs. 76%), though they suggested that their field-based method might be biased in favor of pixel-based techniques. Dronova et al. (2011, 2012) assessed the effects of segmentation scale, thematic detail, and classification methods on the accuracy of wetland plant functional type maps in Poyang Lake, China. The authors reported that the highest classifica-tion accuracies were obtained using (1) scales that were substantially coarser than pixels, (2) classi-fication schemes that match the spatial and spectral hierarchy of the landscape, and (3) classification methods that operate at multiple scales appropriate to the plant community patch sizes.

Kamal and Phinn (2011) assessed pixel and OBIA methods for mapping mangroves using CASI-2 hyperspectral imagery. They compared SAM and linear spectral unmixing (LSU) pixel-based approaches with multiscale segmentation. The mapping results showed that the SAM produced accurate class polygons with only a few unclassified pixels (overall accuracy of 69%), LSU resulted in a patchy polygon pattern with many unclassified pixels (overall accuracy of 56%), and the object-based mapping produced the most accurate results (overall accuracy of 76%). Kamal et al. (2014) extended their study of mangroves by describing the relationship between mangrove patch size and pixel size in WorldView-2 imagery. They concluded that semivariograms can be used to optimize scale parameters for more accurate multiscale mapping of mangroves.

Continuing the exploration of optimal OBIA methods, Kim et al. (2011) provided the first wetland-focused study of 0.3 m aerial imagery in an OBIA context. The authors examined three issues: a comparison of single- and multiscale segmentation, the benefits of incorporating the gray-level co-occurrence matrix (GLCM) textural measure in classification, and the effects of GLCM quantization level in OBIA. Salt marsh classes mapped included vegetation, channel, and mud. The single-scale approach was found to be the least accurate (76% overall). The multiscale method using only spectral information performed better (82%). Adding the GLCM texture measure increased the accuracy to 89%. GLCM quantization level was found to affect the value of including texture in the classification. Ouyang et al. (2011) suggested that multiscale segmentation provides greater accuracy benefits than texture and shape features.

Building on work such as Kim et al. (2011) and the preceding technical exploration studies, Moffett and Gorelick (2013) performed a sensitivity analysis to determine optimal data and method choices for wetland vegetation and channel mapping in San Francisco Bay (California, United States). They concluded that, overall, optimal choices depend on the data used, the wetland features to be mapped, and “conceptual models of wetland organization.” Specifically, the authors recom-mended high-resolution CIR or RGB imagery, multiresolution segmentation, and scale parameters appropriate for the patch size of the landscape and the data used. They noted that high-resolution

K23165_C009.indd 179 10/28/2014 7:27:53 PM


LiDAR data were not beneficial in distinguishing channels or other features. Concerns raised included the following: feature shape affects the quality of segmentation results; guidance in using OBIA is difficult to generalize from the literature due to inherent customization of the approach for particular study areas and input data; and even optimal OBIA methods may not be reproducible because of the human interpretation factor.

In the first OBIA-specific wetland mapping study to evaluate a decision tree classifier, Smith (2010) used the RF algorithm to determine an optimal scale parameter and then to classify SPOT 4/5 images into wetland classes. The study was conducted in North and South Dakota (United States). The overall accuracy for the seven mapped classes was estimated at 85%, with wetland veg-etation producer’s and user’s accuracies of 73% and 81%, respectively. The author suggested that, given the substantially increased numbers of data-derived variables allowed by OBIA techniques, using an algorithmic approach such as RF may reduce the subjectivity suggested by Moffett and Gorelick (2013) and others. Despite challenges in the use of OBIA techniques, Antonellini et al. (2014) suggested that such methods may be especially appropriate for making land use suitability maps for assessment and management of freshwater resources.

cAse studIes

The following are four case studies based on work completed at the University of Minnesota’s Remote Sensing and Geospatial Analysis Laboratory and published in peer-reviewed journals between 2011 and 2014. These case studies are intended to provide an overview of recent research in the context of broader wetland mapping efforts. The cases progress from data selection, to data preprocessing, to mapping methods and results. Both OBIA and decision tree classifiers were used.

The first case is an assessment of various data types for their usefulness in mapping wetlands. The second examines topographic preprocessing options for computing the Compound Topographic Index (CTI), a commonly used topographic derivative for wetland mapping. Third is a study of wetland mapping of northern Midwest wetlands using multisource and multitemporal data. The last case presents a study of integration of high-resolution optical and LiDAR data using an OBIA approach in different types of Midwest wetlands.

case study 1 data type suitability for Mapping Midwest Wetlands (Based on knight et al. 2013)

Choosing the most useful input data is one of the most important steps to ensuring a high-quality classification result—regardless of the method used. Given that wetlands are often more difficult to accurately map than other land cover classes, the use of optimal input data is neces-sary (Ozesmi and Bauer 2002). Many studies have been done of wetland mapping approaches, OBIA and other, examining both appropriate input data and classification algorithms. Most of those studies have focused on radar or Landsat-like sensors (e.g., Baker et al. 2006; Lunetta and Balogh 1999; Wright and Gallant, 2007). However, in recent years, the increasing availability of high-spatial-resolution imagery and LiDAR is allowing for new approaches to mapping wet-lands. Though several studies address this question (e.g., Bowen et al. 2010; Halabinski et al. 2011; Laba et al. 2008; Lang and McCarty 2009; Maxa and Bolstad 2009), there is a need for broader study of the available high-resolution data types to determine which potentially provide the largest increases in mapping accuracy. The objective of this research was to examine the effects of data type selection on wetland mapping accuracy.

This research question was evaluated in two different physiographic regions in Minnesota (United States): (1) an urban–suburban area in the Twin Cities Metropolitan Area and (2) a northern boreal forest. The data used included high- and low-resolution topographic deriva-tives (digital elevation model [DEM], CTI, curvature), National Agricultural Imagery Program

K23165_C009.indd 180 10/28/2014 7:27:53 PM


(NAIP) 1 m optical imagery with infrared, 0.6 m spring leaf-off aerial imagery, SSURGO soil maps, and C-band Radarsat-2 imagery (forested area only). The CTI (Beven and Kirkby 1979) uses the local upslope contributing area, also known as specific catchment area (SCA), and local slope to indicate the relative potential wetness of landscape elements. Experimental trials were conducted by varying input data to the classifier, for example, all available data, optical but no topography, radar included or excluded, high- versus low-resolution topographic data, and topo-graphic derivatives included or excluded. The decision tree classifier See5 was used to evaluate all combinations of input data. The classifier was trained with image- and ground-based train-ing data. Independent image- and ground-based reference data were used to evaluate the accu-racies of the various trials. Classes used were a simple wetland/nonwetland determination and the Cowardin classification system for wetland types (Cowardin and Myers 1974).

The results of this study reinforced the conclusions of existing research and suggested inter-esting cost–benefit questions for future projects. First, the importance of topographic data to accurate wetland classification results cannot be overstated. In both study areas, when topog-raphy layers were removed from the input data stack, accuracies declined significantly. In par-ticular, when NAIP imagery only was used, accuracy fell by 15%–31%, depending on whether the simple wetland/nonwetland or Cowardin wetland-type classification schemes were used. However, the source of the topography data was not a significant factor. Mapping accuracies were statistically indistinguishable, no matter whether the topographic variables were derived from the 10 m National Elevation Dataset (NED) or the 3 m LiDAR data. Thus, the presence of topographic information was more important than its source. This result suggested that if cost is a significant issue, and high spatial resolution is not a critical specification of a given project, relatively low-resolution topographic data, when used with imagery, may provide suf-ficient wetland mapping accuracy. The comparisons between CTI and curvature also indicated no significant difference in accuracy estimates, regardless of spatial resolution. Thus, in areas similar to those used in this study, which are not topographically diverse, the simpler curvature derivative may be a suitable choice over the more computationally intensive CTI.

Second, the choice of classification thematic detail had a significant effect on mapping accu-racy estimates. Wetland/nonwetland accuracies ranged from 93% to 79% for the suburban and forested study areas, respectively. The more specific Cowardin wetland-type mapping accura-cies were lower in both areas: 84% and 58%, respectively. Accuracy estimates were lower in the forested area because of canopy occlusion of ground features and the often-subtle reflectance differences between upland and wetland tree species. Thus, as expected, using the simplest clas-sification scheme that meets the needs of a particular project is recommended.

Finally, the inclusion of Radarsat-2 or leaf-off imagery did not result in higher mapping accu-racies. C-band radar is not ideally suited for wetland mapping in forested areas due to relatively low canopy penetration compared with other radar wavelengths, so the lack of improvement in accuracy was unsurprising. However, spring leaf-off imagery was expected to provide sub-stantial benefits, due to an enhanced ability to see ground features and moisture and because of its suitability for discriminating coniferous and deciduous vegetation. Possible explanations for these results include the particularly dry conditions existing during the radar and leaf-off acquisitions and canopy illumination differences in the high-spatial-resolution leaf-off imagery.

case study 2 topographic data preparation for Wetland Mapping (Based on rampi et al. 2014b)

Topography layers and derivatives are often used as inputs to wetland mapping methods because of their suitability for showing where water might stand on the landscape and thus where wetlands might exist. The topographic derivative CTI is based on local slope and the SCA.

AQ7

K23165_C009.indd 181 10/28/2014 7:27:53 PM


Because slope is a simple, intuitive calculation, the accuracy of the CTI depends much more on the quality of the SCA. The SCA is computed using a flow accumulation layer that is normally derived from a flow direction layer. A flow direction layer can be computed in several ways that are broadly separated into single flow direction (SFD) and multiple flow direction (MFD) algorithms (Erskine et al. 2006; Gruber and Peckham 2008; Wilson et al. 2008). This case study examines the impacts of flow direction algorithm choice on the CTI and resulting wetland mapping accuracy.

SFD algorithms route flow from a DEM grid cell to only one adjacent cell. The most promi-nent examples of SFD algorithms are the Deterministic 8 (D8; O’Callaghan and Mark 1984) and Rho8 (Fairfield and Leymarie 1991). In contrast, MFD algorithms allow water to flow to multiple neighboring grid cells. Tested MFDs included DEMON (Costa-Cabral and Burges 1994), Deterministic Infinite (D∞; Tarboton 1997), Mass Flux (MF; Gruber and Peckham 2008), Triangular Multiple Flow (MD∞; Siebert and McGlynn 2007), and Divergent Flow (FD8; Freeman 1991).

This study was conducted in watersheds in three ecoregions in Minnesota: Northern Glaciated Plains, Central Hardwood Forest, and Northern Lakes and Forest. The base datasets for flow direction model and CTI creation were 3 m LiDAR DEMs acquired in 2010 and 2011. For each area, the LiDAR DEM was processed with each SFD and MFD algorithm. The result-ing flow direction models were used with slope layers to create CTIs. The CTI layers were then thresholded to establish two simple classes: upland and wetland. The accuracies of these classes were assessed with respect to field- and image-based reference data and compared to an existing National Wetlands Inventory (NWI) map.

The results of this study suggested two conclusions about the importance of flow direction in wetland mapping. First, the MFD algorithms generally performed better than SFD. The differ-ences in these classes of algorithm were particularly evident in visual comparison. The MFD-derived wetland maps appear much smoother, clearer, and more realistic in terms of expected water flow over the DEM (Figure 9.1). MFD results also had fewer visual artifacts, such as linear features and speckle. Quantitatively, MDF algorithms were generally better in terms of wetland mapping accuracy. Though the differences were not large and varied per area, the best-performing algorithms were D∞, MD∞, and MF.

Second, SFD algorithms may be appropriate when software availability and computation speed are factors or when visual appearance of the flow direction layer is not a major concern. The SFD algorithms are implemented in a wider array of commercial and open-source software packages and generally have lighter computational requirements. MFD algorithms require sig-nificantly more computing resources and are available in a more limited number of software packages. Thus, an ease-of-use argument can be made in favor of SFD. If the flow direction model is intended only as in input to the creation of a derived product such as the CTI, which is then generalized to a relatively coarse wetland map, then the generally poorer appearance of SFD layers may not be a limiting factor. However, if a particular use relies on detailed analysis of water movement across the landscape, or requires the most realistic possible depiction of flow, then an MFD algorithm is recommended.

Further research is recommended to determine the effects of using CTI layers derived from various flow direction models in the following areas: (1) in more complex studies, such as those using OBIA methods, where the visual artifacts resulting from using SFD approaches may affect segmentation boundaries; (2) when using DEM resolutions other than 3 m, for coarser or finer DEMs, the flow direction algorithm choice may be more important due to variations in the flow due to resolution; and (3) in areas where nondepressional wetlands are common, for example, organic flat wetlands, groundwater discharge-fed wetlands, and some types of fens. Traditional flow direction models are unlikely to capture the factors leading to wetland forma-tion in such areas.

K23165_C009.indd 182 10/28/2014 7:27:53 PM


case study 3 Multisource, Multitemporal Mapping of northern Midwest Wetlands (Based on corcoran et al. 2011, 2013)

Selection of input data types and their acquisition times is an important determining factor of the accuracy of wetland classifications from geospatial data. Capturing diagnostic features in a wet-land’s hydroperiod, or seasonal changes in water storage capacity (“hydrologic signature”), has been described as the most important aspect of its biodiversity and phenology (Wissinger 1999).

(a) (b) (c)

(e)(d) (f )

(g) (h) (i)

fIgure 9.1 Visual comparison of (a) the NWI polygons, (b) CIR aerial imagery 2011, (c) D8 CTI, (d) Rho8 CTI, (e) DEMON CTI, (f) FD8 CTI, (g) D∞ CTI, (h) MD∞ CTI, and (i) Mass Flux CTI for the Central Hardwood Forest study area. Bluer colors represent higher CTI values or more water accumulation. Redder colors represent lower CTI values or dryer areas. The SFD models contain more pixilation and erroneous linear features. (With kind permission from Springer Science + Business Media: Rampi et al. 2014, Figure 9.4; reprinted.)

AQ8

K23165_C009.indd 183 10/28/2014 7:28:08 PM


The hydroperiod of a wetland dictates the optimal acquisition timing of geospatial data for clas-sification and, thus, indirectly affects data availability. Changes in hydroperiod can change a wetland’s classification. This case study examined the impacts of input data type and timing on thematic accuracy of wetland maps derived from high-resolution aerial imagery, optical imag-ery, SAR imagery, DEM-derived topographic measures, and soil maps.

These studies were conducted in the Northern Lakes and Forest ecoregion in Minnesota. The input data were incorporated into various RF classification trials to determine which data types and acquisition times were most influential on the accuracy of wetland maps. Thematic accura-cies were assessed using both image- and field-based reference data.

Corcoran et al. (2011) reported wetland/nonwetland accuracy of 75% when using a combina-tion of optical, topographic, and radar data (Figure 9.2). The accuracy decreased to 72% when

Water

Emergentwetlands

Legend

ForestedwetlandsScrub/shrubwetlands

(a) (c)

(b) (d)

Upland

Confidence

High

Low

N1

km

fIgure 9.2 Subset area showing the upland/water/wetland determinant classification results from an RF decision tree classification: (a) using optical and topographical data alone and (b) using optical, topographic, and radar imagery combined. The original NWI data are shown in panel (c), where land cover classes were consolidated for ease in comparison. The relative confidence of the decision tree classification using all opti-cal, topographic, and radar data is shown in panel (d). The red circle indicates an area with substantial dif-ferences in emergent wetland classification between the datasets. (Reprinted with permission from Corcoran, J.M. et al., Can. J. Remote Sens., 27(5), 564, 2011.)

K23165_C009.indd 184 10/28/2014 7:28:14 PM


radar imagery was excluded. Wetland-type accuracies were significantly lower at 63%. No sig-nificant accuracy differences due to acquisition timing were identified. Corcoran et al. (2013) reported Level 1 (upland, water, and wetland) mapping accuracy of 85% when using Landsat-5, topographic derivatives, PALSAR radar imagery, and soil data. Level 2 (Cowardin wetland types) accuracy was 69% (Figure 9.3). Wetland class user’s and producer’s accuracies were 93% and 79%, respectively. Spring optical and radar imagery was consistently more influential in the RF classification process for mapping Level 1 classes; however, spring imagery was no more valuable than summer imagery for Level 2 classes.

The results of these studies suggested several conclusions related to input data selection and timing. First, the most influential variables for Level 1 mapping were the red, near-infrared, and middle-infrared optical bands, NDVI, elevation, curvature, hydric soils, and L-band H-V polarization SAR. For Level 2 mapping, additional important variables were Landsat Band 5, tasseled cap greenness and wetness bands, L-band HH polarization SAR. Second, SAR imagery can increase mapping accuracies, but care must be taken to select imagery that is appropriate for the landscape being mapped. In this study area, PALSAR L-band imagery resulted in higher accuracies than Radarsat-2 C-band imagery for Level 1 mapping—an expected result for a forested area. However, C-band radar did provide some benefits in Level 2 mapping. Third, variation in precipitation affected wetland mapping accu-racy. Imagery acquired in above normal precipitation conditions was more influential than that acquired after below normal precipitation. Fourth, reducing the number of input variables

Upland

Cowardinclassification

TM, aerial, Rsat-2,PALSAR, Soil

all season, all data

WaterEmergent wetlandForested wetlandScrub/shrub wetland

ConfidenceHigh: 1Low: 0

0

0 1 2 4

1 2 km

46°4

0'30''

N

46°4

4'0''N

46

°47'3

0''N

92°28'30''W92°33'0''W92°37'30''W

km

N

fIgure 9.3 Output classification of the most accurate full season RF model for the Level 2 land cover clas-sification using all available Landsat 5 TM, aerial orthophoto, topographic, RADARSAT-2, PALSAR, and soil data. (From Corcoran, J.M. et al., Remote Sens., 5(7), 3212, 2013.)

K23165_C009.indd 185 10/28/2014 7:28:20 PM


provided to the RF classifier can result in lower computational requirements, with only small decreases in accuracy. Thus, as long as the most influential data types and timing (e.g., spring with average or above precipitation) are used, there is no need to include additional input data in an RF approach.

Future research should examine several aspects of RF-based wetland mapping: (1) the use of RF confidence maps to select areas for additional training and reference data collection, (2) including additional LiDAR-derived topographic and vegetation structure metrics (e.g., number of returns, standard deviation of returns, and vegetation height), (3) including context and tex-ture in an OBIA approach, and (4) more in-depth study of the benefits of the available bands, polarizations, and decompositions of radar imagery.

case study 4 Mapping upper Midwest Wetlands using oBIA (Based on rampi et al. 2014a)

Integration of multiple high-resolution data types has the potential to improve the accuracy of wetland maps. OBIA techniques are well suited for high-resolution mapping efforts because they incorporate spectral, spatial, textural, and contextual information and human knowledge to delineate and characterize image segments. This case examines the effectiveness of using an OBIA approach to integrate two types of high-resolution data, multispectral leaf-off aerial imagery and LiDAR elevation data, across three Midwest physiographic provinces.

This research was carried out in three study areas located in different ecoregions in Minnesota: Northern Glaciated Plains, Western Corn Belt Plains, and Northern Lakes and Forest. Data used in this project included (1) multispectral leaf-off imagery with four spectral bands (near-infrared, red, green, and blue) with a spatial resolution of 0.5 and acquired in 2009 and 2011; (2) the Green Ratio Vegetation Index (GRVI); (3) LiDAR data collected in 2010 and 2011; and (4) five LiDAR-derived products: 3 m DEM, 1 m digital surface model (DSM), a 1 m normalized DSM (nDSM), 1 m intensity layer, and 3 m CTI. For each study area, an OBIA approach was developed through the creation of a customized rule set using the Cognition Network Language (CNL) in Definiens eCognition Developer version 8.8.0. eCognition was used to develop the three rule sets with a divide and conquer approach, which is a multiscale iterative method that partitions image objects based on size, shape, and spec-tral attributes. Four land cover classes were mapped using the OBIA technique: wetlands, agriculture, forest, and urban. The accuracies of these classes were evaluated against field- and image-based reference data. Four standard accuracy estimators were used: overall accu-racy, producer’s accuracy, user’s accuracy, and the kappa coefficient of agreement (Congalton and Green 2009). In addition, the wetland class was compared to an existing NWI map in each study area.

The results of this study suggested several conclusions regarding the importance of integrat-ing high-resolution imagery and LiDAR data for mapping wetlands using OBIA. First, the com-bination of high-resolution optical imagery and different LiDAR-derived products including the CTI significantly improved the accuracy of wetland classification compared to traditional pixel-based approaches (Corcoran et al. 2011; Knight et al. 2013; Rampi et al. 2014; Sader et al. 1995). The overall percent accuracy estimates for the three study areas ranged from 90% to 93%. The wetland-specific user’s and producer’s accuracies ranged from 91% to 96%. Kappa coefficients were 0.84–0.90.

Second, while many existing OBIA studies focus primarily on segmentation techniques, this study focused on developing customized rule sets for multiple areas using a multiscale iterative

AQ9

K23165_C009.indd 186 10/28/2014 7:28:20 PM


approach that combined LiDAR and imagery layers suitable for discriminating wetlands from upland classes. This OBIA-based method used information about the input data layers, includ-ing size, shape, spectral, textural, and contextual information, and further refined the approach using human knowledge and experience relevant to each area. The customized rule sets devel-oped in this study may be appropriate for wetlands located in areas with similar characteristics such as topography, soils, vegetation, and land use.

Finally, the OBIA results were significantly improved over the NWI map (Figure 9.4), espe-cially in terms of wetland omission errors. This result was expected, given that the NWI is decades old in parts of the state; however, the methods used in this study would be appropriate as a base classification for a new version of the NWI. In fact, the state of Minnesota is currently using such an approach in its ongoing statewide update of the NWI, in which OBIA segmenta-tion and classification are used to provide a base classification from which image analysts work to produce a final refined product.

These OBIA results for wetland mapping boundaries are encouraging and useful as an initial classification of these complex and diverse ecosystems. Further research is needed to classify the wetlands into types (e.g., Cowardin) using high-resolution data and OBIA techniques.

OBIA wetland polygonsNWI polygons

fIgure 9.4 Comparison map of the original NWI polygons and OBIA polygons for a small portion of the Minnesota River Headwaters with background aerial imagery. The OBIA polygons much more closely match actual wetland boundaries than does the NWI. (Reprinted with permission from Rampi, L.P. et al., Photogram. Eng. Remote Sens., 80(5), 53, 2014a.)

K23165_C009.indd 187 10/28/2014 7:28:22 PM


dIscussIon And conclusIon

Wetland classification using geospatial data will likely always be among the more difficult tasks facing remote sensing mapping professionals. However, significant improvements have been made as a result of new and emerging techniques such as OBIA and decision tree classifiers. Though these methods require new ways of thinking about remote sensing–based mapping as opposed to traditional techniques such as maximum likelihood supervised classification, the extensive research presented in this chapter provides guidance as to how to work within this new paradigm.

Data selection is of paramount importance to producing a high-quality wetland classification. The use of high-resolution optical imagery has become much more common as its availability increases. Though such imagery is undeniably useful, care must be taken to select a spatial reso-lution that is appropriate for mapping the patch size and shape of the landscape, since these fac-tors affect the accuracy of image segmentation. In addition, aerial imagery often does not provide middle-infrared bands that have been shown to be useful for identification of moisture. Thus, the mapping analyst must balance the desire to use higher-resolution imagery with the disadvantages of not having middle-infrared and sometimes even near-infrared bands (e.g., many NAIP datasets).

Given the increasing availability of LiDAR datasets, high priority should be assigned to acquir-ing such data due to the substantial benefits they offer for characterizing wetlands. LiDAR deriva-tives like elevation, curvature, CTI, vegetation height, and intensity provide landscape structural information that is difficult or impossible to obtain from optical imagery. If LiDAR data are not available, coarser elevation-only datasets such as the US NED or Shuttle Radar Topography Mission may be acceptable substitutes. Elevation derivatives like the CTI provide information about water flow and storage on the landscape, which is helpful in wetland classification. When using metrics such as CTI, consideration should be given to the method by which flow is accounted for, with MFD algorithms providing a more realistic depiction of water movement through the landscape. Like LiDAR, radar imagery provides structural and also textural information. While radar can improve the accuracy of wetland classification, especially in inundated areas, the limited availability and often high cost of radar imagery limit its broad use.

Wetlands are 4D (space and time) processes in 3D landscapes. A wetland’s hydroperiod regulates not only how it might appear on a single remotely sensed image but also its extent, water depth, vegetation composition, biological productivity, and, over time, soil type. Given the significant tem-poral variability of many wetlands, data acquisition timing is a critical issue. Ideally many images showing the intra- and interannual variability of a wetland landscape would be used in classifica-tion, but rarely is such a rich dataset available. Thus, the analyst should seek to maximize the extent to which temporal variability is represented in the data that are available. Diagnostic times, for example, in the spring when areas are often at their wettest, should be prioritized for data acquisi-tion. A contrasting dataset, for example, in the summer, may be used to highlight the changes in wetland extent and water level throughout the year and to identify nonpersistent vegetation (e.g., aquatic beds). It is important to note the prevailing long-term precipitation regime at the time of data collection, since especially wet or dry intervals may significantly affect the apparent extent of wetlands on the landscape—a phenomenon that might not be anticipated based only on relatively short-term weather information.

For classification of wetlands, OBIA and decision trees have emerged as high-performing addi-tions to the mapping analyst’s toolset. The additional information gained from examining image segments in addition to individual pixels is powerful. The inclusion of texture, shape, size, and con-text variables in wetland classification has been shown to result in significantly higher accuracies than without that information. Multiscale segmentation has become the standard in creating seg-ments for subsequent classification. Viewing a landscape at multiple scales allows for hierarchical processing of its features. This is especially important given the influence patch size and shape have on both segmentation and classification. Wetland areas having relatively consistent patch sizes and shapes are likely to be much easier to classify than those with more complex patch characteristics.

K23165_C009.indd 188 10/28/2014 7:28:22 PM


A challenge to the broad adoption of OBIA techniques is the subjective nature of segmentation parameter setting and classification rule set creation. More research is needed to describe optimal OBIA approaches for different wetland types and input data.

Decision trees have several benefits that make them attractive for wetland classification. They are able to use continuous and categorical data, do not rely on assumptions about the distributions of input data, and can be resistant to overfitting. Potentially the biggest advantage of decision trees in the OBIA wetland mapping context may be that they can help to reduce the subjectivity in both the segmentation and rule set creation processes. Emerging research suggests that using informa-tion from OBIA image segments within an RF or other tree-based classifier may be advantageous. Analyst-driven OBIA segmentation and classification will likely perform better when study area sizes are conducive to such relatively labor-intensive efforts. However, when study areas are large or highly variable, making rule set customization onerous, a decision tree classifier used within the OBIA context may represent a good trade of accuracy for efficiency.

Though the use of OBIA and decision trees has resulted in significant advances in the accuracy of wetland maps, there remain many issues that require future research. Perhaps most prominent of these is the lack of guidance, in the relatively nascent literature in this area, for how best to imple-ment OBIA methods across a wide range of wetland types and using the many different data sources available. A large experimental study or an integrative analysis of widely varying wetlands and the data types that are most suitable for mapping them would be a welcome addition to the wetland mapping literature.

references

Antonellini, M., T. Dentinho, A. Khattabi, E. Masson, P.N. Mollema, V. Silva, and P. Silveira. 2014. An inte-grated methodology to assess future water resources under land use and climate change: An application to the Tahadart drainage basin (Morocco). Environmental Earth Sciences 71(4):1839–1853.

Baatz, M. and A. Schape. 2000. Multiresolution segmentation : An optimization approach for high quality multi-scale image segmentation. Journal of Photogrammetry and Remote Sensing 58:12–23.

Baker, C., R. Lawrence, C. Montagne, and D. Patten. 2006. Mapping wetlands and riparian areas using Landsat ETM+ imagery and decision-tree-based models. Wetlands 26(2):465–474.

Benz, U.C., P. Hofmann, G. Willhauck, I. Lingenfelder, and M. Heynen. 2004. Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS Journal of Photogrammetry and Remote Sensing 58(3–4):239–258.

Beven, K.J. and M.J. Kirkby. 1979. A physically based, variable contributing area model of basin hydrology. Hydrological Sciences Journal 24:43–69.

Blaschke, T. 2003. Object-based contextual image classification built on image segmentation. In Proceedings of the 2003 IEEE Workshop on Advances in Techniques for Analysis of Remotely Sensed Data, 27–28 October, Washington, DC, CD-ROM.

Blaschke, T. 2010. Object based image analysis for remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing 65(1):2–16.

Blaschke, T. and J. Strobl. 2001. What’s wrong with pixels? Some recent developments interfacing remote sens-ing and GIS. GIS—Zeitschrift fur Geoinformationssysteme 6:12–17.Bock, M., P. Xofis, J. Mitchley, G. Rossner, and M. Wissen. 2005. Object-oriented methods for habitat mapping at multiple scales—Case studies from Northern Germany and Wye Downs, UK. Journal for Nature Conservation 13(2–3):75–89.

Bowen, M.W., W.C. Johnson, S.L. Egbert, and S.T. Klopfenstein. 2010. A GIS-based approach to identify and map playa wetlands on the high plains, Kansas, USA. Wetlands 30:675–684.

Breiman, L. 2001. Random forests. Machine Learning 45:5–32.Chen, L.-C., T.-A. Teo, C.-H. Hsieh, and J.-Y. Rau. 2006. Reconstruction of building models with curvilinear

boundaries from laser scanner and aerial imagery. In L.-W. Chang, W.-N. Lie, and R. Chiang (Eds.), Advances in Image and Video Technology (pp. 24–33). Springer, Berlin, Germany.

Conchedda, G., L. Durieux, and P. Mayaux. 2008. An object-based method for mapping and change analysis in mangrove ecosystems. ISPRS Journal of Photogrammetry and Remote Sensing 63(5):578–589.

Congalton, R.G. and K. Green. 2009. Assessing the Accuracy of Remotely Sensed Data: Principles and Practices, 2nd edn. CRC Press/Taylor & Francis, Boca Raton, FL.

K23165_C009.indd 189 10/28/2014 7:28:22 PM


Corcoran, J.M., J.F. Knight, B. Brisco, S. Kaya, A. Cull, and K. Murhnaghan. 2011. The integration of optical, topographic, and radar data for wetland mapping in northern Minnesota. Canadian Journal of Remote Sensing 27(5):564–582.

Corcoran, J.M., J.F. Knight, and A.L. Gallant. 2013. Influence of multi-source and multi-temporal remotely sensed and ancillary data on the accuracy of random forest classification of wetlands in northern Minnesota. Remote Sensing 5(7):3212–3238.

Costa, M.P.F., O. Niemann, E. Novo, and F. Ahern. 2002. Biophysical properties and mapping of aquatic veg-etation during the hydrological cycle of the Amazon floodplain using JERS-1 and Radarsat. International Journal of Remote Sensing 23:1401–26.

Costa-Cabral, M. and S.J. Burges. 1994. Digital elevation model networks (DEMON): A model of flow over hillslopes for computation of contributing and dispersal areas. Water Resources Research 30:1681–1692.

Cowardin, L.M. and V.I. Myers. 1974. Remote sensing for identification and classification of wetland vegeta-tion. The Journal of Wildlife Management 38(2):308–314.

Dissanska, M., M. Bernier, and S. Payette. 2009. Object-based classification of very high resolution panchro-matic images for evaluating recent change in the structure of patterned peatlands. Canadian Journal of Remote Sensing 35(2):189–215.

Dronova, I., P. Gong, N.E. Clinton, L. Wang, W. Fu, S. Qi, and Y. Liu. 2012. Landscape analysis of wetland plant functional types: The effects of image segmentation scale, vegetation classes and classification methods. Remote Sensing of the Environment 127:357–369.

Dronova, I., P. Gong, and L. Wang. 2011. Object-based analysis and change detection of major wetland cover types and their classification uncertainty during the low water period at Poyang Lake, China. Remote Sensing of Environment 115(12):3220–3236.

Erskine, R.H., T.R. Green, J.A. Ramirez, and L.H. MacDonald. 2006. Comparison of grid-based algorithms for computing upslope contributing area. Water Resources Research 42:W09416.

Fairfield J. and P. Leymarie. 1991. Drainage networks from grid digital elevation models. Water Resources Research 27:709–717.

Fournier, R.A., M. Grenier, A. Lavoie, and R. Helie. 2007. Towards a strategy to implement the Canadian Wetland Inventory using satellite remote sensing. Canadian Journal of Remote Sensing 33(S1):S1–S16.

Frauman, E. and E. Wolff. 2005. Segmentation of very high spatial resolution satellite images in urban areas for segments-based classification. In ISPRS WG VII/1 “Human Settlements and Impact Analysis” Third International Symposium Remote Sensing and Data Fusion Over Urban Areas (URBAN 2005) and Fifth International Symposium Remote Sensing of Urban Areas (URS 2005), Tempe, AZ.

Freeman, T.G. 1991. Calculating catchment area with divergent flow based on a regular grid. Computers and Geosciences 17(3):413–422.

Friedl, M.A. and C.E. Brodley. 1997. Decision tree classification of land cover from remotely sensed data. Remote Sensing of the Environment 61(3):399–409.

Frohn, R.C., M. Reif, C.R. Lane, and B. Autrey. 2009. Satellite remote sensing of isolated wetlands using object-oriented classification of Landsat 7 data. Wetlands 29(3):931–941.

Gilmore, M.S., E.H. Wilson, N. Barrett, D.L. Civco, S. Pirsloe, J.D. Hurd, and C. Chadwick. 2008. Integrating multi-temporal spectral and structural information to map wetland vegetation in a lower Connecticut River tidal marsh. Remote Sensing of the Environment 112(11):4048–4060.

Goetz, S.J., R.K. Wright, A.J. Smith, E. Zinecker, and E. Schaub. 2003. IKONOS imagery for resource manage-ment: Tree cover, impervious surfaces, and riparian buffer analyses in the mid-Atlantic region. Remote Sensing of the Environment 88(1–2):195–208.

Grenier, M., A.-M. Demers, S. Labrecque, M. Benoit, R.A. Fournier, and B. Drolet, 2007. An object-based method to map wetland using RADARSAT-1 and Landsat ETM images: Test case on two sites in Quebec, Canada. Canadian Journal of Remote Sensing 33(S1):S28–S45.

Grenier, M., S. Labrecque, M. Garneau, and A. Tremblay. 2008. Object-based classification of a SPOT-4 image for mapping wetlands in the context of greenhouse gases emissions: The case of the Eastmain region, Québec, Canada. Canadian Journal of Remote Sensing 34(S2):S398–S413.

Gruber S. and S. Peckham. 2008. Land-surface parameters and objects in hydrology. In: Hengl T. and Reuter H.I. (Eds.), Geomorphometry: Concepts, Software, Applications (pp. 171–194). Elsevier, Amsterdam, the Netherlands.

Halabisky, M., L.M. Moskal, and S.A. Hall. 2011. Object-based classification of semi-arid wetlands. Journal of Applied Remote Sensing 5:053511-1–053511-13.

Haralick, R.M. and Shapiro, L.G. 1985. Image segmentation techniques. Computer Vision, Graphics, and Image Processing 29(1):100–132.

K23165_C009.indd 190 10/28/2014 7:28:22 PM


Harayama, A. and J.M. Jaquet. 2004. Multi-source object-oriented classification of landcover using very high resolution imagery and digital elevation model. In Enviroinfo, Geneva, Switzerland.

Harken, J. and Sugumaran, R. 2005. Classification of Iowa wetlands using an airborne hyperspectral image: A comparison of the spectral angle mapper classifier and an object-oriented approach. Canadian Journal of Remote Sensing 31(2):167–174.

Hess, L.L., J.M. Melack, E.M.L.M. Novo, C.C.F. Barbosa, and M. Gastil. 2003. Dual-season mapping of wetland inundation and vegetation for the central Amazon basin. Remote Sensing of the Environment 87:404–428.

Hogg, A.R. and K.W. Todd, 2007. Automated discrimination of upland and wetland using terrain derivatives. Canadian Journal of Remote Sensing, 33(1):68–83.

Homer, C., J. Dewitz, J. Fry, M. Coan, N. Hossain, and C. Larson. 2007. Completion of the 2001 national land cover database for the conterminous United States, Photogrammetric Engineering and Remote Sensing 73(4):337–341.

Kamal, M. and S. Phinn. 2011. Hyperspectral data for mangrove species mapping: A comparison of pixel-based and object-based approach. Remote Sensing 3(10):2222–2242.

Kamal, M., S. Phinn, and K. Johansen. 2014. Characterizing the spatial structure of mangrove features for optimizing image-based mangrove mapping. Remote Sensing 6(2):984–1006.

Kettig, R.L. and D.A. Landgrebe. 1976. Classification of multispectral image data by extraction and classifica-tion of homogeneous objects. IEEE Transactions on Geoscience Electronics GE-14(1):19–26.

Kim, M., M. Madden, and T.A. Warner. 2009. Forest type mapping using object-specific texture measures from multispectral Ikonos imagery: Segmentation quality and image classification issues. Photogrammetric Engineering and Remote Sensing 75(7):819–829.

Kim, M., T.A. Warner, M. Madden, and D.S. Atkinson. 2011. Multi-scale GEOBIA with very high spatial reso-lution digital aerial imagery: Scale, texture and image objects. International Journal of Remote Sensing 32(10):2825–2850.

Knight, J.F., B.P. Tolcser, J.M. Corcoran, and L.P. Rampi. 2013. The effects of data selection and thematic detail on the accuracy of high spatial resolution wetland classifications. Photogrammetric Engineering and Remote Sensing 79(7):613–623.

Kressler, F. and K. Steinnocher. 2008. Object-oriented analysis of image and LiDAR data and its potential for a dasymetric mapping application. In T. Blaschke, S. Lang, and G. J. Hay (Eds.), Object-Based Image Analysis—Spatial Concepts for Knowledge-Driven Remote Sensing Applications (pp. 611–624). Springer, Berlin, Germany.

Laba, M., B. Blair, R. Downs, B. Monger, W. Philpot, S. Smith, P. Sullivan, and P.C. Baveye. 2010. Use of textural measurements to map invasive wetland plants in the Hudson River National Estuarine Research Reserve with IKONOS satellite imagery. Remote Sensing of the Environment 114(4):876–886.

Laba, M., R. Downs, S. Smith, S. Welsh, C. Neider, S. White, M. Richmond, W. Philpot, and P. Baveye. 2008. Mapping invasive wetland plants in the Hudson River National Estuarine Research Reserve using QuickBird satellite imagery. Remote Sensing of the Environment 112:286–300.

Lang, M.W. and G.W. McCarty. 2009. Lidar intensity for improved detection of inundation below the forest canopy. Wetlands 29(4):1166–1178.

Lang, M.W., G.W. McCarty, R. Oesterling, and I. Yeo. 2013. Topographic metrics for improved mapping of forested wetlands. Wetlands 33:141–155.

Lang, M.W., O. McDonough, G.W. McCarty, R. Oesterling, and B. Wilen. 2012. Enhanced detection of wetland-stream connectivity using LiDAR. Wetlands 32:461–473.

Li, J. and W. Chen. 2005. A rule-based method for mapping Canada’s wetlands using optical, radar and DEM data. International Journal of Remote Sensing 26(22):5051–5069.

Lunetta, R. and M. Balogh. 1999. Application of multi-temporal Landsat-5 TM imagery for wetland identifica-tion. Photogrammetric Engineering and Remote Sensing 65(12):1303–1310.

MacFaden, S. W., J.P. O’Neil-Dunne, A.R. Royar, J.W. Lu, and A.G. Rundle. 2012. High-resolution tree canopy mapping for New York City using LIDAR and object-based image analysis. Journal of Applied Remote Sensing 6(1):063567-1.

Maxa, M. and P. Bolstad. 2009. Mapping northern wetlands with high resolution satellite imagers and lidar. Wetlands 29(1):248–260.

Meng, Y. and S. Peng. 2009. Object-oriented building extraction from high-resolution imagery based on fuzzy SVM. In International Conference on Information Engineering and Computer Science. IEEE, 2009.

Moffett, K.B. and S.M. Gorelick. 2013. Distinguishing wetland vegetation and channel features with object-based image segmentation. International Journal of Remote Sensing 34(4):1332–1354.

AQ10

K23165_C009.indd 191 10/28/2014 7:28:22 PM


Nobrega, R.A.A., C.G. O’Hara, and J. A. Quintanilha. 2008. An object-based approach to detect road fea-tures for informal settlements near Sao Paulo, Brazil. In Object-Based Image Analysis (pp. 589–607). Springer, Berlin, Germany.

O’Callaghan, J.F. and D.M. Mark. 1984. The extraction of drainage networks from digital elevation data. Computer Vision, Graphic and Image Processing 28:328–344

Olson, C.E. 1960. Elements of photographic interpretation common to several sensors. Photogrammetric Engineering 26(4):651–656.

O’Neil-Dunne, J.P.M., S.W. MacFaden, A.R. Royar, and K.C. Pelletier. 2012. An object-based system for LiDAR data fusion and feature extraction. Geocarto International 1–16.

Ouyang, Z.-T., M.-Q. Zhang, X. Xie, Q. Shen, H.-Q. Guo, and B. Zhao. 2011. A comparison of pixel-based and object-oriented approaches to VHR imagery for mapping saltmarsh plants. Ecological Informatics 6(2):136–146.

Ozesmi, S.L. and M.E. Bauer. 2002. Satellite remote sensing of wetlands. Wetlands Ecology and Management 10(5):381–402.

Quinlan, J.R.1993. C4.5: Programs for Machine Learning. Morgan Kaufmann Publishers, San Mateo, CA.Rampi, L.P., J.F. Knight, and C.F. Lenhart. February 2014a. Comparison of flow direction algorithms in the

application of the CTI for mapping wetlands in Minnesota. Wetlands 34(3):513–525.Rampi, L.P., J.F. Knight, and K.C. Pelletier. 2014b. Wetland mapping in the Upper Midwest United States: An

object-based approach integrating Lidar and imagery data. Photogrammetric Engineering and Remote Sensing 80(5):53–62.

Repaka, S.R. and D.D. Truax. 2004. Comparing spectral and object based approaches for classification and transportation feature extraction from high resolution multispectral imagery. In Proceedings of the ASPRS 2004 Annual Conference, Denver, CO.

Rodriguez-Galiano, V.F., B. Ghimire, J. Rogan, M. Chica-Olmo, and J.P. Rigol-Sanchez. 2012. An assess-ment of the effectiveness of a random forest classifier for land-cover classification. ISPRS Journal of Photogrammetry and Remote Sensing 67:93–104.

Rover, J., C.K. Write, N.H. Euliss, D.M. Mushet, and B.K. Wylie. 2011. Classifying the hydrologic function of prairie potholes with remote sensing and GIS. Wetlands 31:319–332.

Sader, S.A., D. Ahl, and W.S. Liou. 1995. Accuracy of Landsat- TM and GIS rule-based methods for forest wetland classification in Maine. Remote Sensing of the Environment 53(3):133–144.

Seibert, J. and B. McGlynn. 2007. A new triangular multiple flow direction algorithm for computing upslope areas from gridded digital elevation models. Water Resources Research 43:1–8.

Shen, G., H. Guo, and J. Liao. 2008. Object oriented method for detection of inundation extent using multi-polarized synthetic aperture radar image. Journal of Applied Remote Sensing 2(1):023512.

Smith, A. 2010. Image segmentation scale parameter optimization and land cover classification using the Random Forest algorithm. Journal of Spatial Science 55(1):69–79.

Syed, S., P. Dare, and J. Simon. 2005. Automatic classification of land cover features with high resolution imagery and lidar data: An object-oriented approach. In Proceedings of SSC2005 Spatial Intelligence, Innovation and Praxis: The National Biennial Conference of the Spatial Sciences Institute, pp. 512–522.

Tarboton, D.G. 1997. A new method for the determination of flow directions and upslope areas in grid digital elevation models. Water Resources Research 33:309–319.

Teo, T.-A. and L.-C. Chen. 2004. Object-based building detection from LiDAR data and high resolution satellite imagery. In Proceedings of Asian Conference on Remote Sensing, Ching-Mai, Thailand (on CD-ROM).

Tiner, R.W. 1990. Use of high-altitude aerial photography for inventorying forested wetlands in the United States. Forest Ecology and Management 33–34:593–604.

Van, T.T. and H.D.X. Bao. 2010. Study of the impact of urban development on surface temperature using remote sensing in Ho Chi Minh City, Northern Vietnam. Geographical Research 48(1):86–96.

Wang, L., W.P. Sousa, and P. Gong. 2004a. Integration of object-based and pixel-based classification for map-ping mangroves with IKONOS imagery. International Journal of Remote Sensing 25(24):5655–5668.

Wang, L., W.P. Sousa, P. Gong, and G.S. Biging. 2004b. Comparison of IKONOS and QuickBird images for mapping mangrove species on the Caribbean coast of Panama. Remote Sensing of the Environment 91(3–4):432–440.

Wilson, J.P., G. Aggett, Y.X Deng, and C.S. Lam. 2008. Water in the landscape: A review of contemporary flow routing algorithms. In: Zhou, Q., Lees, B., and Tang, G. (Eds.), Advances in Digital Terrain Analysis (pp. 213–236). Springer, Berlin, Germany.

Wissinger, S.A. 1999. Ecology of wetland invertebrates. In Batzer, D.P., Rader, R.B., and Wissinger, S.A. (Eds.), Invertebrates in Freshwater Wetlands of North America: Ecology and Management (pp. 1043–1086). Wiley, New York.

AQ11

AQ12

AQ13

K23165_C009.indd 192 10/28/2014 7:28:22 PM


Witharana, C., D.L. Civco, and T.H. Meyer. 2014. Evaluation of data fusion and image segmentation in earth observation based rapid mapping workflows. ISPRS Journal of Photogrammetry and Remote Sensing 87:1–18.

Wright, C. and A. Gallant. 2007. Improved wetland remote sensing in Yellowstone National Park using classification trees to combine TM imagery and ancillary environmental data. Remote Sensing of the Environment 107(4):582–605.

Zhou, W. and A. Troy. 2008. An object-oriented approach for analysing and characterizing urban landscape at the parcel level. International Journal of Remote Sensing 29(11):3119–3135.

AUTHOR QUERIES[AQ1] Please check the identified head level for correctness.[AQ2] Please provide complete details for references “Corcoran et al. (2012) and Liu et al. (2008).”[AQ3] Please check if the shortened running head is okay.[AQ4] Please provide the expansion of “CART,” if appropriate.[AQ5] References “Friedl and Brodley (1996) and Moffett and Gorelick (2012)” have been

changed to “Friedl and Brodley (1997) and Moffett and Gorelick (2013)” as per reference list. Please check.

[AQ6] Please provide the expansion of “CASI,” if appropriate.[AQ7] Please check the word “See5” starting in the sentence “The decision tree classifier...” for

correctness.[AQ8] Please fix “a” or “b” for Rampi et al. (2014).[AQ9] Please fix “a” or “b” for reference citation “Rampi et al. (2014).”[AQ10] Please provide publisher location for references “Meng and Peng (2009) and Syed et al.

(2005).”[AQ11] Please provide editor(s) name for reference “Nobrega et al. (2008).”[AQ12] Please provide volume number for reference “O’Neil-Dunne et al. (2012).”[AQ13] Please provide in-text citation for reference “Teo and Chen (2004).”

K23165_C009.indd 193 10/28/2014 7:28:22 PM

K23165_C009.indd 194 10/28/2014 7:28:22 PM

9 theory and applications of object-based image …...ping wetlands using remotely sensed and...

Documents