UNIVERSITY OF CINCINNATI October 27, 2006 Date:___________________ Navendu Chaudhary I, _________________________________________________________, hereby submit this work as part of the requirements for the degree of:

Doctorate of Philosophy in:

Geography It is entitled: An Object Oriented Approach to Land Cover Classification for State of Ohio.

This work and its defense approved by:

Dr. Robert C. Frohn Chair: _______________________________ Dr. Nicholas P. Dunning _______________________________ Dr. Susanna Tong _______________________________ Dr. Alan P. Sullivan _______________________________

_______________________________

An Object Oriented Approach to Land Cover Classification for State of Ohio.

A dissertation submitted to the

Division of Research and Advanced Studies of the University of Cincinnati in partial fulfillment of the requirements for the degree of

DOCTORATE OF PHILOSOPHY (Ph.D.)

in the Department of Geography of the McMicken College of Arts and Sciences 2006 by Navendu Chaudhary M.S., University of Cincinnati, 2002 M.Sc., University of Pune, 1997 B.Sc., University of Pune, 1992

Committee Chair: Dr. Robert C. Frohn

Abstarct

The purpose of this research was to develop an object oriented approach to land cover analysis and evaluate this approach along with five other classifiers for accuracy in classifying Level II land-cover categories in Ohio. These methods consist of (1) USGS National Land Cover Data; (2) the spectral angle mapper; (3) the maximum likelihood classifier; (4) the maximum likelihood classifier with texture analysis; and (5) a recently introduced hybrid artificial neural network; The segmentation object-oriented processing (SOOP) classifier outperformed all others with an overall accuracy of 93.8% and Kappa of 0.93. SOOP was the only classifier to have by-class producer and user accuracies of 90% or higher for all categories. An artificial neural network (ANN) classifier had the second highest overall accuracy of 87.6% and Kappa of 0.85. The four remaining classifiers had overall accuracies less than 85%. The SOOP classifier has been applied to Landsat-7 data to perform a level II land-cover classification for the state of Ohio.

i

ii

Acknowledgements

I would like to take this opportunity to thank my parents Avinash Chaudhary, Maya Chaudhary for their care and support throughtout my life. I am also thankful for receiving love and support form my wife Shalaka. I would like to express my gratitude to my advisor Dr. Robert C. Frohn whose generous support and expertise has helped me through my tenure here at Department of Geography. I am very grateful to my other committee members, Dr. Susanna Tong, Dr. Nicholas P. Dunning in geography, and Dr. Alan P. Sullivan in anthropology, for their advice and support over the course of the study. I would like to thank all my classmates Andrew Miller, John Baker, Olimpia Neri, Yongping Hao for making my experience at the department joyous. I thank the faculty members within the University of Cincinnati , Dr. Lin Liu, Dr. Robert B. South, Dr. Howard A. Stafford, and Dr. Roger Selya, Dr. Wolf Roder, Dr. Richard Beck for their support and inspiration that made this study possible. I would also like to thank my friends , my brother, sister-in law and especially my niece for their help and encouragement.

iii

Table of Contents Abstract…………………………………………..……………….…………………….i Acknowledgements…………………………………………………………………iii Table of Contents………………………………..…………………………………..Iv List of Tables and Figures…………………………………………………………vi Chapter 1: Introduction………………………………..…………...…….…..……..1 1.1 Research objective……………………………………………….……….……….2 1.2 Significance of Research …………..……………………………………………..2 1.3 Organization of the Dissertation……………………………………….…………4 Chapter 2: Background and Literature Review……………………….……..…..5 2.1 Land-cover mapping………………………………………………….……………5 2.2 Classification Scheme……………………………………………….….…………6 2.3 Classification criteria…………………………….……………….……….……….6 2.4 Traditional classifiers……………………………………………….……………..9 2.5 Object oriented approach…………………………….………………...………..14 2.6 Accuracy Assessment……………………………………………………………21 Chapter 3: Methodology..…………………..……….…….……….…………….....23 3.1 Data Collection……………………..……….…….…….....................................23 3.2 Study Area………………………………….………….……….………………….26 3.3 Image processing………………………………………………….……………...27 3.3.1 USGS National Land Cover Data (NLCD)………………. ………….27 3.3.2 Spectral angle mapper (SAM) ………………………….………..……29 3.3.3 Maximum likelihood classifier (MLC)………………………………….31 3.3.4 Maximum likelihood classifier with texture………………..………….33 3.3.5 Artificial neural network (ANN)………………………………………...35 3.4. Object oriented segmentation …………………………………………………37 3.4.1 Segmentation and object-oriented processing (SOOP)…...………. 37 3.4.2 Image segmentation…………………………………………………….43 3.5 Object based classification……………………………………………………….47 3.6 Anderson Level II classification ………………………………………………...50 3.6.1 Evergreen Forest………………………………………………………..50 3.6.2 Deciduous forest………………………………………………………...52 3.6.3 Commercial Industrial and Transportation.......................................54 3.6.4 Residential………………………….……………………………………57 3.6.5 Row Crops……………………………………………………………….59 3.6.6 Pasture …………………………………………………………………..60 3.6.7 Water……………………………………………………………………..61 3.6.8 Urban Recreational Grasses…………………………………………..62 3.7 Accuracy Assessment …………………………………………..……………….64 Chapter 4: Results and Discussion……………………………………………....68 4.1 Overall classification accuracy …………………….……………………………68

iv

4.2Producer accuracy…………………………………………………………………68 4.3 User accuracy……………………………………………………………………..69 4.4 Class accuracies…………………………………………………………………..70 4.4.1 Water…………………………………………………………………......70 4.4.2 Forest Categories………………………………………………………71 4.4.3 Agricultural Categories ………………………………………………..73 4.4.4 Urban categories………………………………………………………..76 Chapter 5: Conclusions and Contributions……….………..………………... ..99 References…………………………………………………………………………...105

v

List of Tables and Figures Table 2.1 Definitions of Land -Cover Categories ……………………………………………4 Table 4.1 Overall Accuracy and Kappa Coefficients for the six classifiers……………....82 Table 4.2 Producer accuracies of each class for the six classification methods...........................................................................................................83 Table 4.3 User accuracies of each class for the six classification methods ……........…84 Table 4.4 Contingency matrix for accuracy assessment of the Segmentation Object-Oriented Processing (SOOP) Classification .... ………………………..85 Table 4.5 Contingency matrix for accuracy assessment of the Artificial Neural Network (ANN) Classification. …………………………………86 Table 4.6 ANN Contingency matrix for accuracy assessment of the Maximum Likelihood with Texture Classification (MLTC). ……………….......87 Table 4.7 Contingency matrix for accuracy assessment of the Maximum Likelihood Classification (MLC). ………………………..…………..88 Table 4.8 Contingency matrix for accuracy assessment of the National Land Cover Data (NLCD) Classification. ……………………………………....89 Table 4.9 Contingency matrix for accuracy assessment of the Spectral Angle Mapper (SAM) Classification. ………………………………………..…..90 Figure 3.1 Map Showing location of study area.…………………………………………....26 Figure 3.2 NLCD classification ……………………………………………………………....28 Figure 3.3 Spectral Angle Mapper……………………………………………………………30 Figure 3.4 Maximum Likelihood Classifier…………………………………………………..32 Figure 3.5 Maximum Likelihood Classifier with Texture …………………………………..34 Figure 3.6 Artificial Neural Network…………………………………………………………..36 Figure 3.7 eCognition: Multiresolution segmentation……………………………………….38 Figure 3.8 eCognition showing scale parameter and composition of Homogeneity criterion. ………………………………………………….………..39 Figure 3.9 eCognition : Level 3 Objects ………………………………………….………….40 Figure 3.10 eCongition: Level 1 Objects …………………………………………………....41 Figure 3.11 eCognition : Single object for evergreen category …………………………..41 Figure 3.12 a eCognition showing Scale parameter and corresponding no. of object ……………………………………………………………..……….46 Figure 3.12 b eCognition showing Scale parameter and no. of objects eCongition: Level 1 Objects ………………………………………...……...47 Figure 3.13 eCognition showing image band combination ……………………………….48 Figure 3.14 Segmentation and Object Oriented Programming…………………………..49 Figure 3.15 Evergreen trees: Pine & Spruce ……………………………………………….51 Figure 3.16 eCognition: Distribution of Evergreen category….......................................52 Figure 3.17 eCognition: Object created for evergreen category …………………………52 Figure 3.18 Deciduous trees: Pin Oak, Sugar Maple ……………………………………..53 Figure 3.19 eCognition: Object created for deciduous category ………………...............54 Figure 3.20 Transportation (a) Highway (b) IKONOS image showing highway………………………………………………………………….55 Figure 3.21 eCognition: Level 1 object for CIT category …………………………............55 Figure 3.22 eCognition: Level 2 object for CIT category…………………………………...56 Figure 3.23 eCognition : Showing Level 5 object for CIT category …………..………….56

vi

Figure 3.24 Residential: (a) Housing development (b) View from IKONOS image……………………………………………………………………57 Figure 3.25 eCognition: Level 4 object for Residential category …………………….…..58 Figure 3.26 eCognition: Level 2 object for Residential category………………..………..58 Figure 3.27 Row Crops ……………………………………………………………….………59 Figure 3.28 eCognition: Level 3 object for Row Crop category…...................................59 Figure 3.29 Pastures…...................................................................................................60 Figure 3.30 eCognition: Level 5 object for Pasture category……………………………..60 Figure 3.31 eCognition: Level 1 object for Water category ………………………………61 Figure 3.32 eCognition: Lavel 4 object for Water category………………………............62 Figure 3.33 URG: Golf course ……………………………………………………………....63 Figure 3.34 eCognition: Object showing Urban Recreational Grasses category………63 Figure 3.35 Landsat ETM leaf-on data…………………………………………………….. 65 Figure 3.36 Landsat ETM leaf-off data………………………………………………………65 Figure 3.37 High resolution IKONOS image…................................................................66 Figure 3.38 Design of Contingency matrix …………………………………………….......67 Figure 4.1 Comparison of six classifiers for categories : Row crops and Pasture.……………………………………………………………………………..91 Figure 4.2 Comparison of six classifiers for categories : CIT, Residential and URG …………………………………………………………………………...93

vii

Chapter 1 Introduction

Land cover mapping has many applications such as environmental monitoring, hydrological modeling and resource management (Stehman 1998, Candille 2003, Goetz et al. 2003). Land cover maps rely on data collected from the earth’s surface. Many researchers have utilized remotely sensed data in the form of satellite images, especially for large scale maps (Gitas et al 2004). The collected data are then subjected to various classification methods to generate land cover maps. The usability of such maps depends on their accuracy. According to the United States Geological Survey (USGS), a land-cover classification system should have accuracy for land cover categories of at least 85% (Smits et al. 1999). Past attempts to produce large scale land cover maps of Anderson Level II categories have been less than satisfactory according to these standards of accuracy (Reese 2002, Stehman 2003). This study is an attempt to improve the accuracy for land cover classification of Anderson level II categories by applying image segmentation and object-oriented processing to remotely sensed data.

1

1.1 Research Objectives

The goal of this research is to improve on land cover classification by applying image segmentation and object oriented processing to remotely sensed data as follows, i. Apply an object oriented approach for classification to a study area in southwest Ohio using segmentation techniques. ii. Perform accuracy assessment of the object oriented classification. iii. Improve or modify the object oriented approach if necessary. iv. Compare NLCD and pixel-based traditional classifications with the object oriented classification. v. Apply the object oriented classification approach to classify the entire state of Ohio.

1.2 Significance of Research

Use of remotely sensed data has clear advantages in land-cover classification (Reese et al. 2002). The accuracy of such datasets, however depends on the accurate extraction of information from the remotely sensed data. Traditional classification techniques rely on pixel-based spectral information available in remotely sensed images (Lillesand et a.l 2004). These traditional classifiers 2

therefore have limited success when classifying land-cover categories, like urban, which require additional spatial information (Shackelford 2003). Another limitation is that these spectral per-pixel classifiers fail to distinguish between spectrally similar categories such as between pasture and a small parcel of grass in a residential area. This problem gives rise to high local spectral variation and creates less homogenous and noisier classification (Hill 1998, Lobo et al. 1998).

Image segmentation methods are well known and have been applied in many disciplines especially for pattern recognition ( Schiewe 2003). Image segmentation techniques use low-level pixel based information to create higher level regions or image objects. The use of such objects instead of pixels gives image segmentation techniques the ability to incorporate spatial, spectral, textural, and contextual as well as shape information (Shackelford and Davis 2003). Since the smallest unit is a higher level homogeneous object and not a pixel, image segmentation results in reduced local spectral variation (Hill 1999). The resulting classification is thus comprised of more visually recognizable objects that are more similar to landscape patches than are pixels and thus make more sense ecologically intuitive (Liliberte et al 2004). Since the image segmentation bears the same spatial hierarchical structure displayed by landscape systems it results in a reduction of the modifiable area unit problem (Hay 2003). Bearing these advantages in mind it is the hypothesis of this study

3

that image segmentation will be able to improve the classification accuracy of land-cover using the remotely sensed data.

1.3 Organization of Dissertation

The dissertation is divided into five chapters. Chapter one includes clearly defined research objectives and discussion on significance of this research. Chapter two sets up the background for the study. It reviews literature on the key components of the research that includes the mapping concept, classification scheme, previously researched classifiers as well as new approach and accuracy assessment. Chapter three comprises the methodology, which describes in detail data collection, evaluated classifiers, development of the new SOOP approach and accuracy assessment. Chapter four is a discussion of results. It includes results and comparison of accuracy assessment performed on all six classifiers. Finally chapter five concludes the dissertation by summarizing the research and future work.

4

Chapter 2 Background and Literature Review

2.1. Land-cover Mapping

Recently, use of remotely sensed data for land-cover mapping has gained prominence. The growing interest can be attributed to the fact that advances in computer technology as well as availability of satellite data has made land-cover mapping relatively easy and cheap. The advantages of using satellite data for land-cover mapping can be summarized as follows:

i. Availability of large coverage at relatively high resolution. A typical Landsat scene covers 12190 sq. mi. at a resolution of 30m.

ii. Availability of multiple spectral bands which are useful in identifying different classes.

iii. Easy access to satellite images. The data are available in digital format and often can be downloaded directly on to a computer.

iv. Availability of high resolution satellite images. Although expensive, it can serve as a template for accuracy assessment.

5

v. Maps can be revisited and updated easily.

Before the use of remotely sensed data various issues regarding classification scheme, choice of classification technique and accuracy assessment methodology needs to be resolved. 2.2 Classification Scheme Finding an ideal classification system for land-cover mapping can be a challenging task. There are different ways of looking at classification of landcover and the process is quite subjective in nature. The classification scheme is often determined by the needs of the user depending upon the purpose served. Thus it is reasonable to expect that the classification scheme developed to be used with remotely sensed data is not perfect as it is designed to satisfy the majority of its users. For the purpose of this research a land use and land-cover classification system was chosen which can effectively employ remotely sensed data and meets the following criteria as proposed by Anderson (1971). 2.3 Classification Criteria A land use and land cover classification system which can effectively employ orbital and high-altitude remotely sensed data should meet the following criteria as proposed by Anderson, 1971:

6

1. The minimum level of interpretation accuracy in the identification of land use and land cover categories from remote sensor data should be at least 85 percent. 2. The accuracy of interpretation for the several categories should be about equal. 3. Repeatable or repetitive results should be obtainable from one interpreter to another and from one time of sensing to another. 4. The classification system should be applicable over extensive areas. 5. The categorization should permit vegetation and other types of land cover to be used as surrogates for activity. 6. The classification system should be suitable for use with remote sensor data obtained at different times of the year. 7. Effective use of subcategories that can be obtained from ground surveys or from the use of larger scale or enhanced remote sensor data should be possible. 8. Aggregation of categories must be possible. 9. Comparison with future land use data should be possible. 10. Multiple uses of land should be recognized when possible. In my research a multilevel, hierarchical land use classification was developed based on the priori knowledge of the study area and is roughly based upon an Anderson level II classification (Table 2.1).

7

Table2. 1 Definitions of Land -Cover Categories(NLCD 2001) Class Forest Categories

Definition Areas characterized by tree cover tree canopy accounts for 25-100 percent of the cover.

Evergreen Forest

Areas characterized by trees where 75 percent or more of the tree species maintain their leaves all year. Canopy is never without green foliage Areas dominated by trees where 75 percent or more of the tree species shed foliage simultaneously in response to seasonal change.

Deciduous Forest

Agricultural Categories

Areas characterized by herbaceous vegetation that has been planted or maintained in developed settings for specific purposes.

Pasture/Hay

Areas of grasses, legumes, or grasslegume mixtures planted for livestock grazing or the production of seed or hay crops. Areas used for the production of crops, such as corn, soybeans, vegetables, tobacco, and cotton. Areas of open water, generally with 25 percent or greater cover of water (per pixel).

Crops Water

Open Water

All areas of open water, generally with less than 25% cover of vegetation or soil. Areas characterized by a high percentage (30 percent or greater) of constructed materials

Urban Categories Residential

Includes both heavily built up urban centers (high intensity residential) and areas with a mixture of constructed materials and vegetation (low intensity residential). Includes infrastructure (e.g. roads, railroads, etc.) and all highways and all developed areas not classified as Residential. Vegetation (primarily grasses) planted in developed settings for recreation, erosion control, or aesthetic purposes. Examples include parks, lawns, golf courses, airport grasses, and industrial site grasses.

Commercial/Industrial/Transportation

Urban/Recreational Grasses

8

2.4 Traditional Classifiers Various factors are involved in determining the accuracy of final classification output even when the same training areas are used. The Landcover classification analysis needs to incorporate several factors such as accuracy, land-cover type, software, data availability, ease of use, and level of training etc. (Pal and Mather, 2003). The need to incorporate landscape structure in the selection of classification method has been stressed ( Moy 2002).

According to the United States Geological Survey (USGS), a land-cover classification system should have an accuracy for land cover categories of at least 85% (Smits et al. 1999). Data users surveyed by the USGS in cooperation with The United States Environmental Protection Agency found that 35% of the respondents regarded it as critical and 55% regarded it as desirable to have a classification accuracy of 90% for their application (Lins 1994). On a regional scale for land-cover classifications, these accuracies are rarely achieved especially for the Anderson Level II (Anderson et al. 1976) categories.

Prevalent methods for land-cover analysis are the use of pixel based spectral classifiers. These classifiers include the supervised maximum likelihood classifier (Bolstad and Lillesand, 1991), the spectral angle mapper (SAM), as well as unsupervised Isodata classifier. Among these, the maximum likelihood classifier is widely considered as the most accurate but it requires

9

data to be normally distributed. In the absence of such normally distributed data, classifiers such as the spectral angle mapper, which rely on the spectral shape pattern rather than statistical distribution can give better results (Sohn et al. 2002).There are however, several shortcomings with spectral per-pixel based classifiers with respect to land-cover analysis: (1) It is often difficult to distinguish between landcover classes solely based on their spectral signatures. Residential categories are comprised of mixed pixels of houses, roads, grass, and trees. Use of per-pixel based classification results in residential area being confused with other classes such as transportation or forest. (2) Spectral per-pixel based classifiers ignore the spatial patterns that are obvious in land-cover categories such as the net-like appearance of roads interspersed with rooftops in residential categories. (3) Spectral per-pixel classes are sensitive to scale, i.e. spatial resolution of the satellite image. This dependence can result in multiple features contributing to the spectral response within the pixel. When one pixel covers multiple land-covers, it gives rise to a mixed pixels. Efforts are made to address the problem of mixed pixel often by classifying the mixed pixel based on the majority of the land-cover type (Basin 1997; Zhang and Foody 2001; Aplin and Atkinson 2001). Perfecting a sub-pixel method to model all features contributing to the responses within a single pixel will not be sufficient because the method still would not consider the spatial patterns that are apparent beyond a single pixel.

10

(4) The spectral per pixel classifiers are still limited in differentiating land-cover categories that have the same spectral signature such as urban recreational grasses vs. grasses on natural pastures. In order to address the spatial and contextual patterns that are important in distinguishing certain land-cover types, methods have been developed that incorporate textural or contextual information into the classification scheme.

Contextual or textural classifiers incorporate information from adjacent pixels into the classification process resulting in more realistic classification (Flygare 1997). This information can be texture-based, which describes the spatial variability of tones, or contextual-based, which describes the spatial relationship of a pixel to the remainder of the scene, i.e. proximity of relevant objects in the image which is inherent during visual interpretation (Gurney, 1983; Debeir et al. 2002). Texture measures can be determined from first and second order gray-level statistics (Hsu, 1978; Irons and Petersen, 1981; Haralick, 1986; Jahne, 1991; Gong et al., 1992); texture spectra (Wang and Hee, 1990); low/high pass filters (Cushnie and Atkinson, 1985); structural information (Gong and Howorth, 1990); frequency-based information (Xu et al., 2003); Fourier transforms (Weska et al., 1976; Gong et al., 1992); fractals (Mandelbrot, 1977; Mandelbrot, 1982; Lam, 1990); and stochastic Markov random fields (Dubes and Jain 1989). Contextual information can be determined in reference to ancillary data (e.g. elevation) or in reference to locations on the remotely sensed image itself (Debeir et al. 2002).

11

Incorporating contextual or textural information into the classification process often found successful in increasing the accuracy of land-cover analysis (Swain et al., 1981; Jhung and Swain, 1996, Sharma and Sarkar 1998, Stuckens et al. 2000, Debeir et al. 2002, Xu et al. 2003). With the introduction of textural features and contextual data accuracy of classification analysis can be increased between 0.60 to 0.82 for the Kappa coefficient (Debeir et al. 2002). There are several classification methods available that allows incorporation of spectral, contextual, textural, shape, ancillary, and other features into the classification framework. Among these types of classifiers are artificial neural networks and image segmentation object-oriented classifiers.

Artificial neural networks (ANN) have been studied extensively for landcover classification (Bischoff et al., 1992; Civco, 1993; Ji, 2000; Michelson, 2000; Liu et al. 2002). Artificial neural networks (ANNs) are now widely used, but their operational applications are hindered by the need for the user to specify the configuration of the network architecture and to provide values for a number of parameters, both of which affect performance (Pal and Mather 2003).

ANN involves a large number of simple processing units linked by weighted connections between the processing elements and element

12

parameters. The knowledge of the network is based on the strength of the weighted connections between units. Several researchers (Benediktsson et al., 1990; Hepner et al., 1990; Paola and Schowengerdt, 1995; Skidmore et al., 1997; Openshaw and Openshaw 1997, Liu et al., 2002; Erbeck et al. 2004) found ANN classifiers to be advantageous due to the following; i. possess non-parametric nature, ii. have adaptability, iii. capable of identifying subtle patterns in the training data, iv. capable of filtering noise, v. ability to learn.

The main disadvantage of ANNs is that they can be difficult to train, so the training process has to be carefully controlled thus putting more emphasis on the ability of the analyst.(Michleson et al. (2000). ANNs can be developed that rely more on the spatial information in an image as opposed to conventional classifiers that rely primarily on spectral characteristics (Skidmore et al., 1997; Bruzzone et al. 1997; Erbeck et al. 2004). ANN has been known to increase classification accuracy by 5-6% by adding textural features to the ANN network for classification of seven land-cover categories Ji (2000). Many researchers have found that ANN classification produces similar or superior results to traditional statistical classifiers (Bischof and Pinz, 1992;

13

Kanellopoulos et al., 1992; Li et al., 1993; Paola and Schowengerdt, 1995; Foody and Boyd, 1999; Ji, 2000; Erbeck et al. 2004).

2.5 Object Oriented Approach

Image segmentation and object-oriented processing is a classification approach that can readily incorporate spectral, contextual, textural, and spatial information into the classification process. The basic processing units of object-oriented image analysis are segments or objects, not individual pixels (Benz et al. 2004). The image is divided into connected regions by grouping neighboring pixels of similar properties by low-level information (pixel-based features). Adjacent regions are then merged under selected criterion involving homogeneity or sharpness of region boundaries resulting in creation of segments or objects. Depending upon the stringency of the criterion, segments can be formed at different levels. The higher level objects have spectral, textural, contextual, and shape characteristics that can be used for classification (Benz et al. 2004).

Segmentation algorithms can be region growing/merging, boundary detection, or a combination of both (Stuckens et al. 2000). In region growing algorithms, neighboring pixels that have similar properties are merged to form a larger segment. In boundary detection algorithms, two neighboring pixels that have different pixel properties are assumed to be a boundary of two

14

different segments. Objects are formed by aggregating contiguous non-edge pixels (Stuckens et al., 2000).

Advantages of object-oriented classification are numerous (Benz et al., 2004). i. Meaningful statistic and texture calculations; ii. An uncorrelated feature space with shape attributes; iii. Topological and contextual features of objects; and iv. Close relation between real-world boundaries and object boundaries.

These advantages, however, have failed to generate greater interest in research that compares the classification accuracy of image segmentation and object-oriented classification. Jain and Binford (1991) maintain that the field of image segmentation has been poorly accepted because a thorough comparative evaluation and critical assessment of its major issues have not been achieved on a global scale (Smits et al., 1999). Stuckens et al. (2000) found a 5-6% increase in classification accuracy of a segmentation classification procedure over that of the maximum likelihood classifier for 10 level II landcover categories.This demonstrates that image segmentation may have great potential for improving the classification accuracy of land-cover products that needs further examination.

15

Therefore comparative studies of the performance and accuracy of various land-cover classification methods including these that are newly developed are important for improving the level of accuracy of derived landcover products. There have been numerous studies that have compared the accuracy and performance of various classification methods (Serpico et. Al, 1996; Bastin, 1997; Flygare, 1997; Green et al., 1998; Sharma and Sarkar, 1998; Thomson, 1998; Ji, 2000; Micehlesen et. Al, 2000; Hunter and Power, 2002; Liu et al., 2002; Melesse and Jordan, 2002; Hubert-Moy et al., 2002; Sohn and Rebello, 2002; Emrahoglu et al., 2003; Wardlow and Egbert, 2003; Erbeck et al., 2004; South et al., 2004). These studies compare the performance of a new or introduced classification algorithm to that of the maximum likelihood classier as it is assumed to be the benchmark for parametric classifiers. But as reported by Sohn et al. in 2002 that assumption is not always valid. Therefore, this research compares the performance and accuracy of six methods for landcover classification of eight level II categories in Ohio. These methods are three spectral based classifiers, (1) National Land Cover Data (Vogelman et al. 1998); (2) the spectral angle mapper; (3) the maximum likelihood classifier; (4) the maximum likelihood classifier with texture analysis which adds of texture to the standard maximum-likelihood classifier; (5) a recently introduced hybrid artificial neural network (Frohn and Arellano, 2004) with addition of texture, decision trees, and a customized architecture for each land-cover category to an artificial neural network.; and (6) a recently introduced modified image segmentation and object-oriented processing classifier (Frohn and Chaudhary,

16

2004) which utilizes data transforms, texture, contextual information, and object analysis of segments created at several different scales.

The need for the use of an object oriented approach in remote sensing image segmentation has been in the works since the 1980s, especially in forestry applications. Image segmentation since then has been gaining in popularity in the field of remote sensing. This commonly applied technique in the fields of machine vision and pattern recognition has since proliferated in the field of remote sensing for environmental applications. The availability of inexpensive hi-tech computing has opened up new possibilities for use in landcover analysis. (Pekkarinen 2002, Schiewe 2003). The basic processing units of object-oriented image analysis are objects, not individual pixels (Benz et al. 2004). Initial image segmentation uses low-level information (pixel-based features) and through an iterative process creates higher-level contiguous regions or image objects. These higher level objects have spectral, textural, contextual, and shape characteristics that are useful for land-cover classification analysis. (Benz et al. 2004).

Image segmentation has a number of advantages over per-pixel spectral classifiers: i. segmentation allows the incorporation of spectral, textural, contextual, and shape information (Shackelford and Davis 2003);

17

ii provides classification results with higher accuracy (Stucken et al. 2000, Geneletti and Gorte 2003); iii. reduces within class variance (Hill 1999); iv. classification results provide geo-information in a form that is convenient to use in a GIS database (Geneletti and Gorte 2003); v. ability to handle the data is considerably improved especially with respect to handling future updates. vi. elimination of smaller mapping units results in a more visually attractive classification map (Stuckens et al. 2000) vii. ecologically speaking, image objects are more similar to eco-regions than pixels (Laliberte et al. 2004);. viii. Reduced the sensitivity of the object towards the modifiable area unit problem in remote sensing ( Hay et al. 2003)

A variety of segmentation algorithms is available that can be applied to remotely sensed imagery. These include measurement-space guided spectral clustering; hybrid linkage region growing; centroid linkage region growing; split and merge methods; and area and edge-based methods among many others (Laliberte et al., 2004). Image segmentation can be performed in two stages In first stage the “ global stage”is based on an anlaysis of data in feature space. The objective is to identify clusters in the histogram of the data and form segments from these clusters. The second stage, the “Local Stage” is more

18

common and focuses on the variation of tone or color in a small neighborhood (Kartikeyan et al. 1998). There are two types of segmentation methods are based on local behaviour: (1) edge-detection in which the method finds boundaries between pixels by detecting edges, and (2) region growing methods in which image regions completely surrounded by edge pixels become segments. Thus, pixels either belong to an edge to form a boundary or belong to a segment (Geneletti and Gorte 2003). The primary disadvantage of edge-based methods is that small terrain objects are completely obscured by boundary pixels (Geneletti and Gorte 2003). In region growing segmentation a small neighborhood of pixels is tested for homogeneity criteria. Neighboring pixels that have similar properties are merged to form a larger segment. The resulting segments can be split and merged to create regions of constant tone. Regions can also be grown from seed pixels (Karikeyan et al. 1998, Makela and Pekkarinen 2001, Geneletti and Gorte 2003). One disadvantage of region-growing methods is that results can be affected depending on the order the image is processed (Geneletti and Gorte 2003). The choice of a segmentation method depends on the application. The hybrids of these methods have also been used (Karikeyan et al. 1998).

A number of researchers have presented image segmentation as a tool in the analysis of remotely sensed data (McCauley and Engel 1995, Lobo et al. 1998, Kartikeyan et al. 1998, Hill 1999, Rodriquez et al. 2000, Stuckens et al.

19

2000, Makela and Pekkarinen 2001, Pekkarinen 2002, Geneletti and Gorte 2003, Hess et al. 2003, Mason et al. 2003, Shackelford and Davis 2003, Gitas et al. 2004, Hay et al. 2003, Laliberte et al. 2004, Wulder et al. 2004). These research efforts have utilized a wide variety of data types including AVHRR (e.g. Gitas et al. 2004), Landsat TM (e.g. Makela and Pekkarinen 2001), IRS-IA LISS-I (Karikeyan et al. 1998), SAR (e.g. Hess et al. 2003), LIDAR (Mason et al. 2003), Quickbird, (e.g. Laliberte et al. 2004), IKONOS (e.g. Shackelford and Davis 2003), airborne scanner (e.g. McCauley and Engel 1995), orthophotos (e.g., Geneletti and Gorte 2003), color aerial photographs (e.g. Lobo et al. 1998) and spectrometer data (e.g. Pekkarinen 2002). Applications of image segmentation and object-oriented processing include land-cover classification (Karikeyan et al. 1998, Stuckens et al. 2000, Geneletti and Gorte 2003, Shackelford and Davis 2003), forest inventorying, estimating forest regeneration (Wulder et al. 2004), forest degradation (Filho and Shimabukuro 2002), tropical forest and vegetation type classification (Hill 1999, Rodriquez et al. 2000), mapping burned areas (Gitas et al. 2004), estimating timber volume (Makela and Pekkarinen 2001), wetlands mapping (Hess et al. 2003), and measuring habitat predictability (Mason et al, 2003) among others. Stuckens et al. (2000) found a 5-6% increase in classification accuracy of a segmentation classification procedure over that of the maximum likelihood classifier for ten level II land-cover categories. My research explores the use of image segmentation and object-oriented processing for the classification of eight level-II categories in Ohio.

20

2.6 Accuracy Assessment

The accuracy of land-cover classification products is one of the most common issues in the field of remote sensing. Many different methods have evolved over the years that have improved the accuracy of land-cover classification from remotely sensed imagery. These methods include the use of contextual and textural information (Ryherd and Woodcock, 1996; Stuckens et. al., 2000), GIS and ancillary data (Cetin et al., 1993), elevation (Strahler, 1980), and multi-temporal datasets (Coppin and Bauer, 1996) Although these advances have paid dividends, the accuracy of thematic maps is often too low for practical purposes (Townshend,1992; Wilkinson, 1996). Errors need to be minimized in land-cover maps because error can propagate into further analyses and models that use these maps. A wide variety of classification methods have been developed to improve on the accuracy of conventional spectral classifiers. These include artificial neural networks (Serpico, 1996; Ji, 2000; Michelson et al., 2000, Erbek et al., 2004; Frohn and Arellano-Neri, 2004); decision trees (Pal and Mather, 2003); subpixel analysis (Basin 1997; Zhang and Foody, 2001; Aplin and Atkinson, 2001), contextual and texture-based classifiers (Flygare, 1997; Sharma and Sarkar, 1998; Stuckens et al., 2000; Moy et al. 2001; Debeir et al., 2002; Xu et al. 2003); guided clustering (Reese et al., 2004); fuzzy sets (Melesse and Jordan, 2002); hybrid methods (Wayman et al., 2001; Liu et al., 2002; Kelly et al., 2004; Lo and Choi, 2004), hierarchical classifiers (Ediriwickrema, 1997);

21

multiple classifiers (Liu et al., 2002); image segmentation (Kartikeyan et al., 1998; Stuckens et al., 2000; Benz et al. 2003; Laliberte et al. 2004) and objectbased classification (Walter 2004). The purpose of my research evaluates six classification methods with respect to the accuracy of level II land-cover categories in Ohio.

22

Chapter 3 Methodology

3.1 Data Collection

A 12-band dataset was constructed consisting of bands 1-5, 7 leaf-on (August), and bands 1-5,7 leaf-off (January) from Landsat-7 for an area covering southwestern Ohio (Path 20 Row 33). The size of the study area is 12,564 km2. The area was selected due the familiarity of land-cover types in the region from a combination of fieldwork, analysis of high spatial resolution imagery (IKONOS and Quickbird), maps, and personal experience. The location of the image and study area is shown in Figure 3.1. An eight category, modified Anderson Level II classification scheme was developed consisting of the following categories: evergreen forest, deciduous forest, water, pasture, row crops, commercial/ industrial/ transportation, residential, and urban/recreational grasses. A description of each land-cover category is presented in Table 3.1. Training data for running the classification algorithms were collected by visiting sites with a GPS receiver and digitizing image data. The same training data were used for each classification method except for that of the National Land Cover Data, which has already been classified by the USGS.

23

Table 3.1 Band Descriptions: A summary of the band information is contained in the table below. (Infoterra Ltd., 2001-2006)

Band 1 Band 2 Band 3 Band 4 Band 5 Band 7

Band Width 0.45 - 0.52m (blue) 0.52 - 0.60m (green) 0.63 - 0.69m (red) 0.75 - 0.90m (near infra-red) 1.55 - 1.75m (infra-red) 2.08 - 2.35m (near infra-red)

Spatial Resolution 30 meters 30 meters 30 meters 30 meters 30 meters 30 meters

Band 1- Visible Blue: Designed for water penetration, making it useful for coastal water and lake bathymetry and sediment load mapping. Also useful for differentiation of soil from vegetation, and deciduous from coniferous flora; it is lower for vegetation and coniferous forest. Well fragmented and granular rocks (e.g. some shales, phosphates, evaporites) scatter blue light and result in a high band 1 (and sometimes 2). Band 2 - Visible Green: Designed to measure visible green reflectance peak of vegetation for vigor assessment. Also used to map sediment concentration in turbid waters, and is higher for ferrous iron rich rock compared to ferric iron. Band 3 - Visible Red: A chlorophyll absorption band important for vegetation discrimination. It is higher for rocks and soils rich in iron, especially ferric iron.

24

Band 4 - Near Infrared: Useful for determining healthy vegetation and for delineation of water bodies. It peaks strongly for chlorophyll in healthy vegetation, resulting in a characteristic 'red-edge' between bands 3 and 4. Band 5 - Short wave Infrared: Indicative of vegetation moisture content and soil moisture. Contained water absorbs, resulting in lower values. Dry material results in relatively higher values. It is also useful for discriminating snow and clouds (low for snow, high for clouds). In vegetation free areas band 5 varies according to the type of iron oxide present in rocks and soils, and is generally high for all alteration minerals. Band 7 - Short wave Infrared: This band was selected for its potential for discriminating rocks and for hydrothermal altered zones for mineral exploration. Hydroxyl (OH) molecular bonds in minerals stretch and the resultant electronic vibration causes absorption of energy around 2.2um, resulting in marked low values in band 7 for clay-rich minerals. Carbonate rich materials can also cause the same effect. Silica rich materials, dust in the air and bare soil are often relatively high in band 7.

25

3.2 Study Area The study area covers southwestern Ohio (Path 20 Row 33). The size of the study area was 12,564 km2.

Fig 3.1 Map Showing location of study area.

26

3.3 Classifiers Evaluated A description of the methodology for each classification evaluated in this study is presented below.

3.3.1 USGS National Land Cover Data (NLCD)

The National Land Cover Data (NLCD) set was produced by the USGS in cooperation with The Unites States Environmental Protection Agency in response to growing demand for a consistent and reliable land cover dataset for the conterminous United States. The dataset was produced using 30m Landsat TM data acquired through the Multi-Resolution Land Characterization project (MRLC). The data set utilizes multi-band, bi-seasonal data from Landsat TM; the USGS 3-arc second Digital Terrain Elevation Data (DTED) and derived slope, aspect and shaded relief; the Census Bureau population and housing density data; the USGS Land Use and Land Cover (LUDA), and National Wetlands Inventory (NWI) where available. Additionally State Soil Geographic Database (STATSGO) from USDA, 1994 and data from USGS Biological Division GAP analysis Program is incorporated.(Vogelmann et al. 1998). The NLCD procedure involves generation of leaf-on and leaf-off mosaics for 10 EPA Federal Administrative regions covering the conterminous United States (Stehman et al., 2003). Each region is classified using an unsupervised clustering algorithm ISODATA. The clusters are assigned to a modified Anderson level II classification scheme of 23 categories using National High Altitude Photography program (NHAP) and National Aerial Photography program (NAPP) 27

aerial photographs as a reference. The resulting classes are refined using ancillary data sets and on screen digitization. The classification results from adjacent regions are edge-matched, generating a consistent Land Cover Dataset. The National Land Cover Data (NLCD) has 15 land-cover classes for the state of Ohio. Only 11 of these classes were present in the clipped study area and included high intensity residential, low intensity residential, commercial/industrial/transportation, urban/recreational grasses, quarries/strip mines/gravel pits, deciduous forest, evergreen forest,

Fig. 3.2 NLCD classification

28

Mixed forest, pasture/hay, row crops, and open water. A few adjustments were made to the 11 classes so as to avoid an arbitrary way of defining classes : high intensity residential and low intensity residential were merged into one class, “residential”; quarries/strip mines/gravel pits were merged with “commercial/industrial/transportation,” and the mixed forest class was dismissed.

3.3.2 Spectral Angle Mapper (SAM) The spectral angle mapper (SAM) is a physically-based, per-pixel spectral classifier. SAM compares the angle between the reference spectrum and each pixel measurement vector in n-dimensions, determined by the number of bands. The reference spectrum class with the smallest angle and thus closest match is assigned to the output pixel (Hunter and Power 2002; Jensen, 2005). The spectral angle mappper does not require data to be normally distributed and therefore only perform better in classifying different ecoregions. SAM’s lack of sensitivity to the size of the training data or to the gain factor related to topographic illumination and atmospheric effects gives it a distinct advantage over conventional classifiers (Kruse et al. 1993; Sohn and Rebello , 2002). The main difference between SAM and conventional classifiers is that SAM depends on the spectral shape pattern while conventional classifiers depend on the statistical distribution pattern of the land-cover categories classified. This reliance on shape is what makes SAM perform better (Sohn and Rebello, 2002). However, the SAM algorithm has had

29

mixed results with respect to accuracy in land-cover classification. Sohn and Rebello (2002) found SAM to outperform the benchmark Maximum Likelihood Classifier (MLC), with an increase in overall accuracy of 25% for vegetation types. South et al. (2004) found SAM to outperform four other spectral

Fig 3.3 Spectral Angle Mapper

classifiers, including MLC for classification of crop tillage types. Hunter and Power (2002), on the other hand, found SAM to consistently under-perform MLC for the classification of wetland classes. SAM was selected because of the high number of bands (12) used in the multi-spectral dataset. SAM was applied to

30

the 12-band leaf-on/leaf-off Landsat ETM+ dataset for the classification of the eight land-cover types described above.

3.3.3 Maximum Likelihood Classifier (MLC) The maximum likelihood classifier is one of the most popular methods of classification in remote sensing, The maximum likelihood method has an advantage as it employs probability theory but it is sensitive to normal distribution of data.

It has been used by numerous researchers as a benchmark from which

to compare the performance of other classifiers (Bastin, 1997; Ediriwickrema, 1997; Michaelson et. Al. 2000; Stuckens, et. al., 2000; Hunter and Power, 2002; Liu et al. 2002; Moy et al., 2002; Emrahoglu et al., 2003; Erbeck et al., 2004; Lo and Choi, 2004; South et al., 2004). Factors such as the number, size, and location of training sites; the nature of discriminant variables; and the meaningful evaluation of the classification method have contributed to the perceived high standard of MLC (Ediriwickrema, 1997 and Foody et al, 1992). MLC has been recognized as a stable, robust, and accurate method in standard digital image processing software systems (Ediriwickrema, 1997). MLC incorporates both the variance and covariance matrix of the dataset into the classification decision rule. Maximum likelihood classification assumes that the statistics for each class in each band are normally distributed and calculates the probability that a given pixel belongs to a specific class. With this assumption, the statistical probability of a pixel being a member of a given training class can be computed from the mean vector and the co-variance

31

matrix using a probability density function (Lillesand and Kiefer, 2004). The assumption of normality is often violated in multispectral datasets (South et. al., 2004). Even minor deviations from normality can severely disrupt the classification (Foody, 1992). Land-cover categories with multi-modal histograms should have multiple, individual training samples for each mode to fulfill the normal distribution requirement (Jensen, 2005).

fig. 3.4 Maximum Likelihood Classifier

Four different image band combinations were experimented with to determine the best possible classification result for MLC:

32

i. 12-band dataset (6 leaf-on, 6 leaf-off); ii. ETM+ bands 3,4,5 for both dates; iii. principal components of the 12-band dataset; iv. and a minimum noise fraction transformation (Green et al., 1988) of the 12-band dataset. The 12-band dataset gave the best classification result for MLC and was used for comparison to the other methods.

3.3.4 Maximum Likelihood with Texture Classifier (MLTC) The maximum likelihood classifier is a “spectral classifier” because it is generally applied to spectral values of individual bands in classification analysis. However, it can easily be used as a textural classifier if the bands input into the classification scheme contains textural information as opposed to spectral values (Kartikeyan et al., 1994; Jhung and Swain, 1996). Further, it can be used as a hybrid spectral and textural classifier if both spectral and textural bands are input into the classification scheme. Maximum likelihood with texture classifier (MLTC) was used in this research so that there would be a standard classifier that combines both spectral and textural information to compare to the artificial neural network and the image segmentation object-oriented classifiers, which also incorporate spectral and

33

Fig 3.5 Maximum Likelihood Classifier with Texture

textural information. The MLTC combines the 12-band leaf-on/leaf-off data with a mean co-occurrence texture band to create a 13-band dataset. GrayLevel Co-occurrence Measures (GLCM) use a gray-scale spatial frequency matrix to calculate a variety of texture values. The matrix is determined by comparing pixel values between two neighborhood processing windows with a specified distance and direction. The number of occurrences of the relationship between a pixel and its neighbor within the moving windows are recorded (Haralick 1986). A variety of bands and window sizes were experimented to determine the best possible classification result for MLTC.

34

The study found the mean co-occurrence value of a 3x3 window for leaf-on band 3 (red) produced superior results. This band was combined with the 12band spectral data and the maximum likelihood algorithm was applied to the dataset.

3.3.5 Artificial Neural Network (ANN)

The feedforward neural networks are the first and simplest type of artificial neural networks developed. In this network, the information moves in only one direction, forward, from the input nodes, through the hidden nodes (if any) and to the output nodes. There are no cycles or loops in the network. A feedforward artificial neural network (ANN) was applied to the 12band dataset. This type of ANN is beneficial for supervised classification because of its ability to learn through backpropagation and to generalize a final output (Schalkoff, 1992). There were three layers in the network; one input layer, one hidden layer, and one output layer. The initial ANN used twelve input units (each spectral band), and eight output units (each land-cover category).

The initial ANN produced

unsatisfactory classification results. An attempt was made to improve the classification by adjusting the set of parameters, but in every case a poor generalization was obtained. This type of failure arises when the neural

35

mapping function tries to fit all of the fine details in the training data set rather than capturing the underlying trends in the data (Bishop, 1991).

Fig 3.6 Artificial Neural Network After analyzing the properties of the training sites for each class, the training site statistics, and the results obtained from the initial ANN, a series of changes were made to the ANN classification approach. These changes included: i. Instead of using a general ANN for all classes, a unique ANN was developed for each land-cover class.

36

ii. For urban classes, such as residential and urban/recreational grasses textural information units are incorporated in the input layer of the ANN, and iii. Complementing the ANN classification with a multistage Decision Tree classifier. The modified ANN yielded superior classification results and its classification output was used for comparison to the results of the other classification methods (Frohn and Arellano-Neri 2004).

3.4 Object Oriented Segmentation 3.4.1 Segmentation and Object Oriented Processing (SOOP) The segmentation and object-oriented processing (SOOP) classification method can be divided into two components: i. segmentation to create image objects and ii. classification of the created objects into a land-cover category. SOOP was performed using eCognition software by Definiens Imaging (www.definiens-imaging.com). Bands 3,4,5 of the leaf-off data were used for low-level segmentation of image primitives into higher level image objects.

37

Fig 3.7 eCognition: Multiresolution segmentation

The segmentation was a bottom-up region-merging approach starting with single pixel objects. In an optimization pair-wise clustering process, smaller objects were merged into larger objects based on heterogeneity criteria of color and shape. With each iteration, the pair of adjacent objects with the smallest growth from the defined heterogeneity criteria was merged. The process stopped when the smallest growth for merging of adjacent objects exceeded a pre-defined scale parameter. This procedure simulates the simultaneous growth of segments during each step so that output objects are of comparable size and scale (Benz et al., 2004). The heterogeneity criterion considers both spectral and shape properties. Spectral heterogeneity is

38

determined by weighting factors applied to bands used in the segmentation process. Shape heterogeneity is a value that describes the change of the objects shape with respect to smoothness and compactness (Benz et al., 2004). A color factor of 0.9, a shape factor of 0.1, a compactness value of 0.5, and smoothness value of 0.5 was used in the segmentation process.

Fig 3.8 eCognition showing scale parameter and composition of Homogeneity criterion. A scale parameter is defined in the segmentation process to set a threshold for the maximum increase in heterogeneity of two merging segments. When this parameter is reached, the segmentation process ends. The larger the scale parameter, the larger the segmented objects grow (Baatz and Schape, 1999; Benz et. al, 2004). The dataset was segmented at five different

39

scales depending on the size of object features for different land-cover categories. A scale parameter of 2 was used for classification of small roads and streams; a scale parameter of 5 for larger transportation networks, rivers, evergreen stands, and lakes; a scale parameter of 10 for row crops, pastures, residential areas, small deciduous fragments; a scale parameter of 20 for larger deciduous forest stands and large lakes; and a scale parameter of 30 for the largest deciduous forest patches. A total of 2,868,150 objects were created with scale parameter 2; 436,468 objects with scale parameter 5; 126,656 objects with scale parameter 10; 36,402 objects with scale parameter 20; and 16,780 objects with scale parameter 30.

Fig 3 9. eCognition : Level 3 Objects

40

Fig. 3.10 eCongition: Level 1 Objects.

Fig 3.11 eCognition : Single object for evergreen category.

41

The resulting image objects were classified at various scales depending on the land-cover type into one of the eight level II land-cover categories. Eight bands were used in the classification process. They consisted of bands 3, 4, 5 of leaf-on data; bands 3, 4, 5 of leaf-off data; a mean co-occurrence 3x3 texture band created from band 3 leaf-on; and band 1 of a minimum-noise fraction transformation (MNF) of the original 12-band dataset. The MNF segregates noise in the data, and reduces the computational requirements for further processing (Boardman and Kruse, 1994). The MNF transform consists of two cascaded Principal Components transforms. The first transformation, based on a noise covariance matrix, decorrelates and rescales the noise in the data. The second step is a standard Principal Components transformation of the noise-whitened data (Green et al., 1988). Each land-cover class decision rule was determined from a fuzzy set consisting of membership functions of the object features. A membership function ranges from 0 to 1 for each object feature values with respect to its membership to an assigned land-cover category. Customized membership function decision rules were designed for each land-cover class based on spectral, shape, textural, contextual, and statistical features. For example, for evergreen forest the spectral means of bands 3, 4, 5 of leaf-off were most important to the classification while for small roads shape features were more important in the classification. With respect to distinguishing pasture from urban/recreational grasses, contextual information was important to the classification. And for classification of residential areas, texture was an

42

important object feature. The output classification was determined by assigning each object to the class with the highest degree of membership it had based on all membership features used. Finally a classification-based segmentation was performed to fuse all adjacent objects that were assigned the same land-cover category.

3.4.2 Image Segmentation The segmentation was a bottom-up region-merging approach starting with single pixel objects. In an optimization pair-wise clustering process, smaller segments were merged into larger segments based on heterogeneity criteria of color and shape (Benz et al. 2004):

(1)

Where f is the threshold fusion value for merging segments, hcolor is the heterogeneity criterion for color and hshape is the heterogeneity criterion for shape. The user defined weight parameter w was set to 0.9.

The heterogeneity criterion for color (hcolor) is calculated before and after a potential merging of objects as:

43

(2)

Where, nmerge is the number of pixels within a merged object, nobj1 is the number of pixels in object 1, nobj2 is the number of pixels in object 2, σc is the standard deviation within object of band c. Subscripts merge refer to merge object and obj1 and obj2 refer to the objects prior to a merge,

The heterogeneity criteria for shape describe the improvement of shape with respect to two components smoothness and compactness:

(3)

The user defined weight parameter wcmpct was set to 0.5. The change in smoothness (hsmooth) and compactness (hcmpct) are calculated before and after a potential merging of objects:

44

(4)

(5)

Where, n is the object size, l is the object perimeter, and b is the perimeter of a bounding rectangle.

With each iteration, the pair of adjacent objects with the smallest growth from the defined heterogeneity criteria was merged. The process stopped when the smallest growth for merging of adjacent objects exceeded a pre-defined scale parameter. This procedure simulates the simultaneous growth of segments during each step so that output objects are of comparable size and scale (Benz et al., 2004). A scale parameter is defined in the segmentation process in order to set a threshold for the maximum increase in heterogeneity of two merging segments. When this threshold is reached the segmentation process ends. The larger the scale parameter, the larger the segmented objects grow (Baatz and Schape, 1999; Benz et. al, 2004). Thus one can control the process and its extent by careful setting of this threshold. This gives the user a unique ability to choose different scales for different land-cover categories. In this case the

45

dataset was segmented at five different scales depending on the size of object features for different land-cover categories. A scale parameter of 2 was used for classification of small roads and streams; a scale parameter of 5 for larger transportation networks, rivers, evergreen stands, and lakes; a scale parameter of 10 for row crops, pastures, residential areas, small deciduous fragments; a scale parameter of 20 for larger deciduous forest stands and large lakes; and a scale parameter of 30 for the largest deciduous forest patches. A total of 2,868,150 objects were created with scale parameter 2; 436,468 objects with scale parameter 5; 126,656 objects with scale parameter 10; 36,402 objects with scale parameter 20; and 16,780 objects with scale parameter 30.

Fig 3.12 a eCognition showing Scale parameter and corresponding no. of object

46

Fig.3.12 b eCognition showing Scale parameter and no. of objects

A variety of band combinations from the 12-band dataset were used in the segmentation process. The band combination 3,4,5 of leaf-off produced superior results and was used for object-based classification.

3.5 Object-based Classification The segmented image objects were classified at various scales depending on the land-cover type into one of the eight level II land-cover categories. Eight bands were used in the classification process. They consisted of bands 3,4,5 of leaf-on data; bands 3,4,5 of leaf-off data; a mean co-

47

occurrence 3x3 texture band created from band 3 leaf-on; and band 1 of a minimum-noise fraction transformation (MNF) of the original 12-band dataset. The MNF segregates noise in the data, and reduces the computational requirements for further processing (Boardman and Kruse, 1994). The MNF transform consists of two cascaded Principal Components transforms. The first transformation, based on a noise covariance matrix, decorrelates and rescales the noise in the data.

Fig. 3.13 eCognition showing image band combination.

The second step is a standard Principal Components transformation of the noise-whitened data (Green et al., 1988). Each land-cover class decision

48

rule was determined from a fuzzy set consisting of membership functions of the object features. A membership function ranges from 0 to 1 for each objects feature values with respect to its membership to an assigned land-cover category. Customized membership function decision rules were designed for each land-cover class based on spectral, shape, textural, contextual, and statistical features. For example, for evergreen forest the spectral means of bands 3,4,5 of leaf-off were most important to the classification while for small roads shape features were more important in the classification. For differentiating between crops and pastures bi-seasonal data with bands 3,4 and 5 were used.

Fig 3.14 Segmentation and Object Oriented Programming.

49

With respect to distinguishing pasture from urban/recreational grasses, contextual information was important to the classification. And for classification of residential areas, texture was a significant object feature. The output classification was determined by assigning each object to the class with the highest degree of membership it had based on all membership features used. Finally a classification-based segmentation was performed to fuse all adjacent objects that were assigned the same land-cover category.

3.6 Anderson level II Classification The eight Anderson Level II categories were classified using segmentation and object oriented programming separately based on individual class properties.

3.6.1 Evergreen Forest: Areas characterized by trees where 75 percent or more of the tree species maintain their leaves all year. Canopy is never without green foliage. Around southwestern Ohio it is mostly characterized by Pine and Spruce. Isolated stands of evergreen trees are found scattered throughout the study area.

50

Fig 3.15 Evergreen trees: Pine & Spruce

Fig. 3.16. eCognition: Distribution of Evergreen category

51

Leaf off

Leaf on

Fig. 3.17 eCognition: Object created for evergreen category. For isolating the evergreen forest stand leaf Off Winter Landsat 7 image was most useful. The band combination used was 5, 4, 3. The Evergreen tree stands can be identified easily as green or deep green objects on the image. The scale parameters of 10 and 20 were sufficient depending upon the size of the stand or grouping.

3.6.2 Deciduous Forest: Areas dominated by trees where 75 percent or more of the tree species shed foliage simultaneously in response to seasonal change. The common tree species include pin oaks, maple, birch etc.

52

Fig. 3.18 Deciduous trees: Pin oak, Sugar maple

The deciduous forest type is the most dominant flora in southwest Ohio. It is abundantly found in the study area covering large portions of homogenous land cover type. The scale parameters of 30 and 40 were found to be adequate for classifying deciduous forest category. Band combination used was 5, 4, 3 Leaf off as well as 5,4,3, leaf on was employed. The deciduous forest can be seen as bright green in the leaf-on image but seen as deep brown or reddish brown in leaf-off image.

53

Fig 3.19 eCognition: Object created for deciduous category

3.6.3 Commercial Industrial and Transportation: This category includes infrastructure (e.g. roads, railroads, etc.) and all highways and all developed areas not classified as Residential. Commercial Industrial and Transportation or CIT is a diverse collection of infrastructure that is found in urban areas. It includes roads and highways, which can be seen as linear narrow structures as well as industrial facilities, which are seen as large geometrically shaped objects. The scale parameters of 5 and 10 were used for narrow roads and highways. For classifying larger commercial/ industrial objects, scale parameter of 30 and 40 were used. 54

Fig 3.20 Transportation (a) Highway (b) IKONOS image showing highway

Fig 3.21 eCognition: Level 1 object for CIT category

55

Fig 3.22 eCognition: Level 2 object for CIT category

Fig 3.23 eCognition : Showing Level 5 object for CIT catagory

56

3.6.4 Residential:

Fig 3.24 Residential: (a) Housing development (b) View from IKONOS image

Residential category includes both heavily built up urban centers (high intensity residential) and areas with a mixture of construction materials and vegetation (low intensity residential). Examples include apartment complexes, row houses and singlefamily housing units. The residential area is characterized by a mesh like appearance and its proximity to urban centers. The scale parameter used varied from 20 to 40, depending upon the size and shape of the object. The band combination of 5, 4, 3 leaf off were most useful.

57

Fig. 3.25 eCognition: Level 4 object for Residential category

Fig 3.26 eCognition: Level 2 object for Residential category.

58

3.6.5 Row Crops: Row crops are areas used for the production of crops, such as corn, soybeans, tobacco, and cotton. Both leaf-off and leaf-on images were used with a band combination of 5,4,3. Scale parameter varied from 30-40.

Fig 3. 27 Row Crops

Fig 3.28 eCognition: Level 3 object for Row Crop category 59

3.6.6 Pasture: Pastures are areas of grasses, legumes, or grass-legume mixtures planted for livestock grazing or the production of seed or hay crops. Both leaf-off and leaf-on images were used with band combination of 5,4,3. Scale parameter varied from 30-40.

Fig 3.29 Pastures

Fig 3.30 eCognition: Level 5 object for Pasture category

60

3.6.7 Water:

This category includes areas of open water, generally with less than 25 percent or greater cover of water (per pixel). This category includes small streams as well as big water bodies, such as lakes. The scale parameter of 5-10 was used to classify small streams and rivers. A scale parameter of 20 was used to classify large rivers and small lakes. A scale parameter of 30 -40 was sufficient in classifying large lakes.

Fig. 3.31 eCognition: Level 1 object for Water category.

61

Fig. 3.32 eCognition: Lavel 4 object for Water category.

3.6.8 Urban Recreational Grasses:

Urban recreational gasses are vegetation (primarily grasses) planted in developed settings for recreation, erosion control, or aesthetic purposes. Examples include parks, lawns, golf courses, airport grasses, and industrial site grasses. Leaf-off scene with band combination 5, 4, 3 was employed. Scale parameter from 20-40 was used.

62

Fig 3.33

URG: Golf course

Fig 3.34 eCognition: Object showing Urban Recreational Grasses category.

63

3.7 Accuracy Assessment

Accuracy assessment determines the quality of the land-cover map produced. In this study, a quantitative accuracy assessment was performed by comparing the land-cover map with the reference data. The reference data was collected and labeled using the same classification scheme that was used to create the land-cover map. The classification scheme was mutually exclusive and totally exhaustive (Congalton, 1999), which means each map area fell into one and only one category and every area on the map received a label. A balance between statistical integrity and practical possibility needs to be found. A general “rule of thumb” is to collect more than 50 samples for each category in the error matrix Congalton (1991). This guideline was empirically derived over many years of project experience. For this study, more than 100 samples for each land-cover class were collected following Congalton (1991). A total of 2,942 stratified random points were identified using planimetric and topographic maps, high spatial resolution data (IKONOS 1 m, Quickbird, and digital orthophotoquads), Landsat ETM leaf-on data, Landsat ETM leaf-off data, and site visits. The examples of these images are shown in figure fig. 3.35 3.37.

64

Fig 3.35 Landsat ETM leaf-on data

Fig 3.36

Landsat ETM leaf-off data 65

Fig 3.37 High resolution IKONOS image: A contingency matrix was constructed to compare the reference data to each of the six land-cover classifications. Descriptive statistics were used to evaluate information in each error matrix. Overall accuracy was calculated by dividing the total correct pixels by the total number of pixels in the error matrix. Individual class user and producer accuracies and errors of omission and commission were calculated following Story and Congalton (1986).

66

Fig. 3.38

Design of contingency matrix (Source: Congalton and Green, 1999)

The Kappa coefficient was also calculated to compare the accuracy of the six classifications, which has become a standard component of accuracy assessment (Congalton et al. 1983). Kappa allows the researcher to determine how much the classification produced is significantly better than a random classification.

67

Chapter 4 Results and Discussion

4.1 Overall Classification Accuracy

The segmentation object-oriented processing (SOOP) classification had the highest overall accuracy of 93.8% and Kappa Coefficient of 0.93 followed by the artificial neural network (ANN) classification with an overall accuracy of 87.6% and Kappa of 0.85. The remaining classifications were all below 85% and including the maximum likelihood with texture classifier (MLTC) with an overall accuracy of 82.7% and Kappa 0.80, followed by the maximum likelihood classifier with an overall accuracy of 80.1% and Kappa 0.76. The USGS National Land Cover Data (NLCD) had an accuracy of only 75.1% and Kappa 0.69. The Spectral Angle Mapper (SAM) had the lowest classification accuracy of 66.9% and Kappa 0.60. All three classification methods that incorporated spectral and texture information (SOOP, ANN, MLTC) had higher overall classification accuracies than spectral-based classifications (MLC, NLCD, SAM).

4.2 Producer Accuracies

The producer accuracy is defined as the probability of a reference pixel being correctly classified in a particular category and is a measure of omission error (Jensen 2005). By-class comparisons of producer accuracies for the six

68

classifications are shown in Table 4.2. The SOOP classification was the only method to have all by-class producer accuracies of 90% or greater. SOOP producer accuracies ranged from 89.6% for Urban/Recreational Grasses to 98.0% for water. Producer accuracies for the ANN classification ranged from 61.9% for pasture to 97.8% for deciduous forest. Producer accuracies for the MLC ranged from 66.5% for commercial/industrial/transportation (CIT) to 91% for water and increased slightly with the addition of texture for MLTC, which had a range of 68.3% for urban/recreational grasses to 91% for water. The USGS NLCD had lower producer accuracies ranging from 39.1% for pasture to 86.1% for water. SAM had the lowest producer accuracies, ranging from 41.2 for row crops to 84.6% for water.

4.3 User Accuracies

User accuracy is defined as the probability that a classified pixel actually represents that category on the ground and is a measure of commission error (Jensen 2005). The SOOP classification was the only method to have all byclass user accuracies of 90% or greater. User accuracies for the SOOP classification ranged from 90.8% for deciduous forest to 99.5% for water. Producer accuracies for ANN ranged from 64.4% (row crops) to 99.5% (water). The MLC and MLTC had low user accuracies for urban/recreational grasses and pastures (~33%) and highest user accuracies for deciduous forest (98.7%). User accuracies for NLCD and SAM were lowest for pastures and highest for water.

69

The confusion matrices used to determine the overall, producer, and user accuracies for each of the six classifications are shown in Tables 5-10. The tables are arranged in order of overall accuracy (highest to lowest). The diagonals in the matrices represent correctly classified pixels.

Producer

accuracy was determined by dividing the number of correctly classified pixels in a category by the sum of the class column. User accuracy was determined by dividing the number of correctly classified pixels in a category by the sum of the rows. The error matrices can also indicate which individual categories are being misclassified relative to one another. The results of these categories are examined individually in the next few subsections.

4.4 Class Accuracies

4.4.1 Water

Water had the highest producer accuracies in most of the classification results. This pattern can be attributed to the unique spectral nature and shape of the streams and lakes.

SOOP had the highest producer accuracy for water

of 98% followed by MLC and MLTC (both 91%) and ANN (90.5%). NLCD and SAM had lower producer accuracies for water of 86% and 85% respectively. Omission errors are mainly due to water that was missed in small streams and shallow areas. User accuracies for water were very high for all six classification methods and range from 97.2% (NLCD) to 99.5% (SOOP).

70

4.4.2 Forest Categories

Forest categories include evergreen forest and deciduous forest categories. Forest categories have relatively distinct spectral properties. In addition the use of bi-seasonal data resulted in higher accuracy for these categories by most of the classifiers.

Deciduous Forest

Deciduous forest had producer accuracies above 85% for all classification methods except SAM, which had a producer accuracy of 77%. The ANN and SOOP methods had the highest producer accuracies for deciduous forest (~98%). SAM low producer accuracy for forest is mainly due to the confusion with the crop category. Other spectral classifiers, such as MLC, however, did not have confusion between deciduous forest and crops. SOOP, MLC, and MLTC had user accuracies above 90% for deciduous forest. ANN (83%), NLCD (74%) and SAM (72%) had user accuracies for deciduous forest below 85%. There was some confusion between deciduous forest and residential areas for ANN and NLCD. This confusion is most likely due to classification of stands of trees within residential areas as deciduous forest stands. NLCD also had some confusion between evergreen and deciduous forest.

71

Evergreen Forest

With respect to evergreen forest, SOOP (93%) was the only method to have producer accuracy above 90%. ANN also had relatively high producer accuracy for evergreen forest at 87%. The remaining four classifiers had producer accuracies below 85% for the evergreen category. MLC and MLTC omission errors were mainly due to confusion with the urban/recreational grass (URG) category. Both evergreen and URG have similar signatures because they are green in both leaf-on and leaf-off data. These two classifiers had some difficulty in distinguishing the two categories. The NLCD, on the other hand, had omission errors mostly due to confusion with the deciduous category. The spectral clustering algorithm and the date chosen for the leaf-off Landsat data are potential sources of error for this confusion. User accuracies for all classification methods were much higher than producer accuracies. SOOP (99%), ANN (97%), MLC (95%), and MLTC (94%) had user accuracies above 90% for the evergreen category. Both NLCD (82%) and SAM (83%) had user accuracies below 85% for evergreen. Commission errors for evergreen were caused by confusion with deciduous stands in the NLCD and residential areas in the SAM classification.

72

4.4.3 Agricultural Categories

Figure 4.1 shows an agricultural area within the study area and the resulting classification of that area by the 6 methods. The original image at the top of the figure is the Landsat leaf-off data for the area. Pastures in the image are green and crops are in magenta tones. The SOOP classification is very clean with no noise and appears to classify the pastures and crops well. The ANN classification is also good although it slightly overestimates crops and underestimates pasture and is a bit noisier. The MLC, MLTC, and SAM classifications are very noisy. They tend to overestimate pastures and urban/recreational grasses and underestimate crops. Also, the confusion between pasture and urban/recreational grasses is apparent in these classifications. Finally, the NLCD confuses crop and pasture areas and overestimates crops while underestimating pastures.

Row Crop

SOOP was the only method to have producer accuracies of row crops (96%) above 90%. The next best classifier of row crops was MLTC, which had a producer accuracy of 87% followed closely by MLC, which had producer accuracies of 87%. The next highest classifier with respect to producer accuracies of crops was ANN with 74%. The NLCD performance with respect to row crops had a

73

reasonable producer accuracy of 79%, SAM had the lowest producer accuracies for agricultural categories with 41% for crops. Most of the omission errors were due to confusion with the deciduous forest category. SOOP was also the only classifier to have user accuracies above 90% for both categories. The MLC and MLTC had high user accuracies for crops (93% and 94% respectively. ANN user accuracies for crops were 64%. The majority of commission errors for these two categories were between each other. NLCD and SAM had very low user accuracies for crops (46% and 35% respectively).

Pasture

The main errors of omission for pasture with respect to the MLC and MLTC classification were caused by confusion with the urban/recreational grass (URG) category. This error is explained by the fact that both pastures and URG have grasses with the same or very similar spectral signature. Since the MLTC and MLC classifiers do not incorporate contextual information, such as proximity to urban areas into the classification scheme, these two classes were confused. SOOP on the other hand does incorporate this contextual information into the classification decision rule for URG and therefore has an omission error for the pasture category with respect to URG.

However, the ANN producer accuracy for pasture (62%) was lower than expected. In a previous study by Frohn and Arellano (2004), ANN had much

74

higher producer accuracy for pasture of 90%. The study identified a potential source of error for the decrease in accuracy of pasture in this study. In the previous study, the pasture network architecture incorporated only leaf-off data, while in this study both, leaf-on and leaf-off data were used for pasture. The leaf-on data can cause confusion between very healthy pastures and productive crops. The confusion matrix of ANN shows that the majority of omission errors for the pasture category were in fact due to the crop category. In the future, it is recommend to use only leaf-off data for ANN classification of pastures.

Pastures have a very poor producer accuracy of only 39% using NLCD. Most of the omission errors in the pasture category were caused by confusion with the crop category. In an accuracy assessment of NLCD and GAP land-cover products, researchers have also found that the NLCD had significant confusion between grassland and crop categories (Wardlow and Egbert 2003). They maintained that confusion between these two categories was caused by the suboptimal early spring dates of the NLCD imagery, which is a time when unplanted crops and grassland have similar signatures. SAM had the lowest producer accuracies for agricultural categories with 45% for pasture. As mentioned earlier most of the omission errors for these two categories were due to confusion with the deciduous forest category.

75

SOOP was also the only classifier to have user accuracies above 90% for both categories. The MLC and MLTC had low user accuracies for pasture (63% and 60%, respectively). The majority of commission errors for pasture were with the URG category. ANN user accuracies for pasture were 83%. The majority of commission errors for pasture was due to confusion with row crops. NLCD and SAM had very low user accuracies for pastures (36% and 26% respectively)

4.4.4 Urban Categories

The urban categories consist of residential, commercial/ industrial/ transportation (CIT), and urban/recreational grasses (URG). The classification of urban categories is often found difficult due the diversity of land-cover types that make up these categories (Lo and Choi 2004). SOOP was the only classifier to have producer and user accuracies above 90% for all three urban categories. SOOP producer accuracies were 91% for residential, 91% for CIT, and 90% for URG. SOOP user accuracies were 97% for residential, 92% for CIT, and 92% for URG. ANN producer and user accuracies were the second highest and were at least 85% for the urban categories. These two classifiers are the only classifiers to have high accuracies for all three urban classes because they are the only ones that incorporate spectral, textural, and contextual information into the classification scheme. In residential areas, textural information is necessary for high classification

76

accuracy. Spectral and textural information is necessary for high classification accuracy of CIT. And, spectral and contextual information is necessary for high classification accuracy of URG. Some researchers have used sub-pixel methods to deal with problems of mixed pixels in the urban environment (Zhang and Foody, 2001; Aplin and Atkinson, 2001). Some have recommended that high spatial resolution imagery (