An automated satellite image classification design using object-oriented segmentation algorithms: A move towards standardization

Expert Systems with Applications Expert Systems with Applications 32 (2007) 616–624 www.elsevier.com/locate/eswa

An automated satellite image classification design using object-oriented segmentation algorithms: A move towards standardization Ruvimbo Gamanya *, Philippe De Maeyer, Morgan De Dapper Department of Geography, Gent University, Krijgslaan 281, S8, B-9000 Gent, Belgium

Abstract Numerous segmentation algorithms have been developed, many of them highly specific and only applicable to a reduced class of problems and image data. Without an additional source of knowledge, automatic image segmentation based on low level image features seemed unlikely to succeed in extracting semantic objects in generic images. A new region-merging segmentation technique has recently been developed which incorporates the spectral and textural properties of the objects to be detected and also their different size and behaviour at different stages of scale, respectively. Linking this technique with the FAO Land Cover Land Use classification system resulted in the development of an automated, standardized classification methodology. Testing on Landsat and Aster images resulted in mutually exclusive classes with clear and unambiguous class definitions. The error matrix based on field samples showed overall accuracy values of 92% for Aster image and 89% for Landsat. The KIA values were 88% for Aster images and 84% for the Landsat image.  2006 Elsevier Ltd. All rights reserved. Keywords: Object-orientation; Land use and land cover; Automation; Standardization

1. Introduction There is a demand for tangible landscape objects at several scales which are internally relatively homogeneous on which one can apply spatial statistics and, to assess changes. Various approaches in computer vision, pattern recognition, image analysis, landscape ecology and environmental monitoring are in search of such objects. The classification quality is directly dependent on the quality of extracted objects. Object-oriented image classification is based on the fact that important semantic information necessary to interpret an image is not represented in single pixels, but in meaningful image objects and their mutual relations. A strong and experienced evaluator of segmentation techniques is the human eye/brain combination. By

*

Corresponding author. Tel.: +32 9 264 4694; fax: +32 9 264 4985. E-mail address: [email protected] (R. Gamanya).

0957-4174/$ - see front matter  2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2006.01.055

applying segmentation procedures to the automation of image analysis, the activity of visual digitizing is replaced. Recent applications of image segmentation and image understanding techniques require increased robustness, better reliability and high automation of the algorithms. While traditional methods of region based (region growing, texture based segmentation) (Aguado, Montiel, & Nixon, 1998; Bezdek, 1981; Li, Wang, & Wiederhold, 2000; Liu & Yang, 1994; Ronfard, 1994) and edge based (snakes, dynamic programming) (Canny, 1986; Pal & Pal, 1994) image segmentation algorithms continue to be explored with attempts to improve their performance by adding expert knowledge to their detection criteria, new model driven techniques have been developed—active shape models, active appearance models (Cootes, Edwards, & Taylor, 1998; Cootes, Taylor, Cooper, & Graham, 1995). Performance of traditional approaches to edge-based image segmentation is being improved by utilizing advanced segmentation criteria that reflect higher level of knowledge about the segmented object (Gorte, 1998).

R. Gamanya et al. / Expert Systems with Applications 32 (2007) 616–624

In most applications, however, this expert knowledge is not derived automatically, but introduced by an observer and then translated to mathematical form. This approach requires expert knowledge of the given problem and is usually time consuming when new application has to be designed. Use of information about object shape and form is usually very limited or is not considered at all in the pixel-based approaches. Baatz and Scha¨pe (1999a, 1999b, 2000) developed a new region-merging segmentation technique which incorporates the spectral and textural properties of the objects to be detected and also their different size and different behaviour on different stages of scale, respectively. The concept of scale is important in image analysis as most environmental dimensions possess one or more domains of scale (Wiens, 1989) at which the individual spatial or temporal patches can be treated as functionally homogeneous. The concept of heterogeneity (and homogeneity) is perhaps the key characteristic of every landscape and underlies the scale factor in images. It may be defined as the uneven, non-random distribution of ecological units. There are three types of heterogeneity: temporal, functional and spatial. In the image multi-scale segmentation, the quantitative criterion for the evaluation of the segmentation results is that the average heterogeneity of pixels is minimized. Each pixel is weighted with the heterogeneity of the image object to which it belongs. The qualitative criteria are the fact that any segmentation results have to satisfy the human eye and the information, which can be extracted from image objects for further successful processing. With satellite remote sensing techniques now becoming the single most effective method for land cover/land use data acquisition, it is imperative that standardized, object-oriented approaches to image analysis be developed. Anderson, Hardy, Roach, and Witmer (1976) developed a hierarchical land use and land cover classification system for use with remote sensor data. The Anderson derived land use and land cover classifications have been adopted in most contemporary systems including the FAO Land Cover Classification System (FAO, 2000). This system is created specifically for mapping purposes; it uses a set of independent diagnostic criteria, the classifiers, rather than being nomenclature based. This allows each class to be clearly and systematically defined, thus producing internal consistency. The LCCS specific design allows incorporation into GIS and databases and it was combined with the multi-resolution image segmentation procedure. The primary objective of this research was to develop a standardized, object-oriented classification method, with automation capabilities. This article reports on the development of a method based on linking object-oriented image analysis techniques in eCognition and the FAO land cover classification system. ASTER datasets were used and the automation capabilities of this technique were tested on other ASTER images taken at the same period. To assess application of the method on another satellite images, classification of a Landsat TM image was also undertaken.

617

2. Background 2.1. A ‘new’ paradigm—multi-scale image segmentation Multi-resolution segmentation is a bottom up regionmerging technique starting with one-pixel objects. In numerous subsequent steps, smaller image objects are merged into bigger ones. Throughout this pairwise clustering process, the underlying optimization procedure minimizes the weighted heterogeneity nh of resulting image objects, where n is the size of a segment and h an arbitrary definition of heterogeneity. In each step, that pair of adjacent image objects is merged which stands for the smallest growth of the defined heterogeneity. If the smallest growth exceeds the threshold defined by the scale parameter, the process stops. Doing so, multi-resolution segmentation is a local optimization procedure. Baatz and Scha¨pe (2000) developed decision heuristics to determine the image objects that will merge at each step and definition of homogeneity of image objects to compute the degree of fitting for a pair of objects. Given a feature space, two objects are considered similar when they are near to each other in this feature space. For a d-dimensional feature space the heterogeneity h is described by ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi rX 2 h¼ ðf1d  f2d Þ ð1Þ d

Examples for appropriate object features are for instance mean spectral values or texture features, such as the variance of spectral values. The distances can be furthermore standardized by the standard deviation over all segments of the feature in each dimension vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u uX f1d  f2d 2 h¼t ð2Þ rfd d The degree of fitting of two adjacent image objects can be defined by describing the change of heterogeneity hdiff in a virtual merge. Given an appropriate definition of heterogeneity for a single image object the increase of heterogeneity in a merge should be minimized. There are different possibilities to describe the change of heterogeneity hdiff before and after a virtual merge h1 þ h2 ð3Þ 2 This definition corresponds to the requirement of a quantitative criterion, which aims to minimize the average heterogeneity of image objects when evaluating a segmentation. It can be improved by taking into account the object’s size n

hdiff ¼ hm 

hdiff ¼ hm 

h1 n1 þ h2 n2 n1 þ n2

ð4Þ

An alternative is to weight image object heterogeneity with the object size, thus fulfilling the quantitative criterion which states that average heterogeneity of image objects weighted by their size in pixel should be minimized

618

R. Gamanya et al. / Expert Systems with Applications 32 (2007) 616–624

hdiff ¼ ðn1 þ n2 Þhm  ðn1 h1 þ n2 h2 Þ ¼ n1 ðhm  h1 Þ þ n2 ðhm  h2 Þ

2.2. LCCS technical concepts ð5Þ

This definition can be generalized to an arbitrary number of channels c, each having a weight wc: X wc ðn1 ðhmc  h1c Þ þ n2 ðhmc  h2c ÞÞ ð6Þ hdiff ¼ c

Appropriate definitions for spectral heterogeneity of image objects are for instance—the variance of spectral mean values or the standard deviation of spectral mean values. 2.1.1. Form heterogeneity There are two definitions for form heterogeneity; the first is the deviation from the ideal compact form given by the relation between factual edge length l and the root of the object size n in pixels, that is, the edge length of a square with n pixels 1 h ¼ pffiffiffi n

ð7Þ

Another definition is the deviation from the shortest possible edge length given by the bounding box b of the segment. It is the relation between factual edge length 1 and the edge of the bounding box. In a raster the edge length of the bounding box is also the shortest possible edge length for an arbitrary segment, that is, l P b holds for any image object (Baatz & Scha¨pe, 2000) h¼

l b

ð8Þ

Global mutual fitting is the strongest constraint for the optimization problem and it reduces heterogeneity most over the scene following a pure quantitative criterion. Its main disadvantage is that it does not use the treatment order and builds first segments in regions with a low spectral variance leading to an uneven growth of the image objects over a scene. It also causes an unbalance between regions of high and regions of low spectral variance. Comparison of global mutual fitting to local mutual fitting results show negligible quantitative differences, the former (local) always performs the most homogeneous merge in the local vicinity following the gradient of the degree of fitting. The growth of image objects happens simultaneously as well in regions of low spectral variance as in regions of high spectral variance. A distributed treatment order together with the local mutual best fitting guarantees a systematic handling of regions of different texture and highly minimizes heterogeneity of resulting image objects at the same time. Convincing results are obtained by defining spectral heterogeneity according spectral variance or spectral deviation of the spectral mean values (Eq. (6)). The resulting image objects are highly homogeneous, image objects represent contrasts consistently and segmentation results are highly reproducible with only slight differences of the edges between objects of low contrast.

The LCCS is based on a standardized a priori classification system and not the traditional a posteriori method. An a priori approach is based upon definition of classes before any data collection actually takes place. The advantage is that classes are standardized independent of the area or means used. The a posteriori system, though flexible and adaptable, is unable to define standardized classes (Di Gregorio & Jansen, 1998). The LCCS classification is comprehensive in the sense that any land cover identified anywhere in the world can be readily accommodated. 2.3. Similarities between LCCS and eCognition eCognition enables the development of knowledge bases for elaborate classification of local context and land use using a class hierarchy. This simple hierarchical grouping offers a wide range for the formulation of image semantics and for different analysis strategies. This hierarchical grouping capability is similar to the LCCS. Both systems have an underlying Boolean logic at each decision point. Whereas the LCCS is limited to a few descriptors to define class cover, in eCognition a broader information base is used in distinguishing classes. In the final classes only structural vegetation types that are recognizable on satellite imagery and in the field are used. In vegetation classification, the final classification is not intended to be floristic or ecological vegetation classification. All land covers can be accommodated in this highly flexible system; the classification could therefore serve as a universally applicable reference base for land cover and land use derived from satellite imagery, thus contributing towards data harmonization and standardization. 3. Methods A synopsis of the method is shown in Fig. 1. The LCCS software available at http://www.africover.org/LCCS.htm and eCognition software (Definiens-Imaging GmB) are used for this technique. Image registration and preprocessing need to be undertaken in other image processing software as eCognition does not support these capabilities. 3.1. Data acquisition and preprocessing An ASTER image (09-10-01) was used to develop the protocol. The ASTER and Landsat images were geo-referenced to the 1:250 000 topographic maps corresponding to the UTM 36 K projection, E001. The error was less than 0.5 m. ASTER images have a resolution of 15 m, whilst Landsat has a resolution of 30 m. Cloud covered areas were masked on the images. Automation was performed on another ASTER image taken on the 16th of October of the same year. Color matching was performed were necessary on ASTER images to allow the application of the protocol in the automation process.

R. Gamanya et al. / Expert Systems with Applications 32 (2007) 616–624

619

LCCS

eCognition

Determine class at Dichotomous Level

Perform feature selection and threshold values for splitting classes Class names are recorded in legend

If field data is present, use known objects to select correct feature threshold values

Perform class-related classification for the first two classes Record Protocol steps with the best classification result

Analyze result, if misclassified objects are present, then go back to Step 3

Repeat Steps 1 - 4 for the next class splitting at lower levels

Standardized end-class names with LCCS Map Code, Basic Classifier & Modifier

Other Image Processing Software

Final classified image & legend produced with same LCCS class names.

Save final protocol and class hierarchy for use in automation

Automation

Perform color matching if necessary for images. Import georectififed Image

Load class Hierarchy & execute

Perform radiometric normalization of multi-temporal objects

x – images classified

Accuracy Assessment

Fig. 1. Synopsis of the method.

To validate the classification method a cloud-free Landsat TM image from August 1989 was used. This was necessary to test if application on slightly lower resolution image was possible. A separate protocol was developed for this image. A tasseled cap transformation for enhancing spectral information content was computed and used in the classification of Landsat TM. Tasseled Cap transformation especially optimizes data viewing for vegetation studies (Kauth & Thomas, 1976). This was in addition to the use of other vegetation indices (NDVI, Ratio bands).

eCognition. Field verification of the objects conclusively set the threshold limits at each decision node thereby amending the final protocol adopted. Tasseled Cap vegetation index was calculated from data of the related six TM bands. Three of the six tasseled cap transform bands are often used: band 1 (brightness, measure of soil); band 2 (greenness, measure of vegetation); band 3 (wetness, interrelationship of soil; and canopy moisture). The Tasseled Cap Transformation for Landsat satellite imagery is calculated with the following coefficients:

3.2. Feature selection Brightness ¼ 0:3037ðTM1Þ þ 0:2793ðTM2Þ þ .4743ðTM3Þ

Initial distinction was between primarily vegetated and non-vegetated areas based on the use of the Normalized Difference Vegetation Index (NDVI) (Rouse, Haas, Schell, & Deering, 1973) and other ratio indices. The NDVI is calR R RED culated as follows: NDVI ¼ RNIR . NIRþR NIR Threshold values are initially established based on visual analysis, and the use of the feature space definition in

þ 0:5585ðTM4Þ þ 0:5082ðTM5Þ þ 0:1863ðTM7Þ Greenness ¼ 0:2848ðTM1Þ  0:2435ðTM2Þ  0:5436ðTM3Þ þ 0:7243ðTM4Þ þ 0:0840ðTM5Þ  0:1800ðTM7Þ Wetness ¼ 0:1509ðTM1Þ þ 0:1973ðTM2Þ þ 0:3279ðTM3Þ þ 0:3406ðTM4Þ  0:7112ðTM5Þ  0:4572ðTM7Þ

620

R. Gamanya et al. / Expert Systems with Applications 32 (2007) 616–624

Table 1 The classification features used at decision nodes Spectral information

• • • • •

Computed features (based on spectral info)

• NDVI • Tasseled Cap Vegetation Index (Landsat) • Ratio indices (Band 4/2/Band 3/2 etc)

Generic shape features

• • • • •

Texture (after Haralick)

• GCLM homogeneity • GCLM dissimilarity • GCLM standard deviation

Class Related Features

• • • • •

LCCS

• Life form (Woody/Herbaceous/Lichens or mosses) • Cover/Macro-pattern (Closed/Open/ Sparse) • Height • Spatial distribution • Leaf type (broadleaved/Needleaved/Aphyllous) • Leaf phenology (Evergreen/Deciduous)

Standard deviation Ratio Minimum pixel value Maximum pixel value Relationship to neighbor objects

Area Compactness Rectangular Fit Shape index Density

Relations to neighbour objects Relations to sub objects Relations to super objects Membership to class x Classified as x

Table 1 shows the various features used to split classes at decision nodes. It is important to note that the features listed do not represent the full sets available but only a limited selection was used in this research.

and TM imagery. For each of 130 objects identified, extensive field information including species composition, level and type of management, soil characteristics, topography, forest layers and canopy characteristics were collected. Two methods of object verification were carried out. Objects were selected systematically based on the use of road transects. At every 20 km point along major roads (which are easily recognizable on the images). After a point was selected along the transect line (on either side of the road at a set distance). The point was then located on the satellite image as belonging to a particular object and the object characteristics were assessed. The relational and context features used in its classification were double checked for accuracy and crosschecked with other objects. The second method of targeted object verification was also undertaken. Recognizable objects were first identified on the image and were located in the field using GPS. Identifying the boundary of an object was not easy and was therefore not attempted as that information was not required in the classification process, as each sample point could be easily related to a particular object. An a priori classification of the images had been performed since the parameters used to reach a decision at each node generalized classes to identified, that is, at the upper level of the classification tree. Finer end-class detailing would then require field verification. Field verification of objects was used to conclusively determine threshold values. For example, each object would be simultaneously identified on the image and in the field. Because classification was initiated from a broad to a fine level, if an object was correctly classified at a lower level it automatically implied that all the other threshold values used to split classes at higher levels (nodes) were correct. If the object was incorrectly classified, necessary adjustments to the threshold values at problem nodes were made as outlined in Fig. 1 allowing correction of all objects with similar characteristics.

3.3. Study area 3.5. Automation of methodology The study was based on satellite imagery taken over Central Zimbabwe. The whole area covered was UL 30d34 0 49.0200 E, 17d43 0 08.6700 S and LR 31d12 0 53.2600 E 18d05 0 56.2100 S. For the purposes of this paper, a smaller area was focused on. The subset includes Lake Chivero and Harare, the capital city located to the south east. 3.4. Field reconnaissance High quality training data are essential for accurate land cover classification of any sort. Extensive field reconnaissance was conducted throughout the entire study area in the summer of 2005. As many training sites as possible were visited in each cover type as defined by a modified LCCS classification system (Di Gregorio & Jansen, 1998). To ensure high spatial and thematic precision of collected data, field sampling was conducted in a GIS framework using laptop computers and GPS information and orthorectified, geo-referenced, and resolution-enhanced ASTER

The automation capabilities of this classification system were tested using overlapping ASTER images. There were no significant differences in the spectral quality between the primary image (09-10-01) and an ASTER image (16-10-01) used for testing the protocol although color matching was performed. The protocol was recorded and adjusted as the method was being developed. Sun elevation adjustments and color matching are necessary steps in the application of the automation protocol if the images were taken at slightly different times. Same day adjacent images do not need this correction. 4. Results and discussion 4.1. Protocol development The intention of the research was to facilitate rapid automated classification of satellite same-date images of

R. Gamanya et al. / Expert Systems with Applications 32 (2007) 616–624

the same area. It is imperative that an image representing all the potential classes is used in the protocol building process and knowledge-based rules are developed. Failure to adequately represent all the land cover and land use classes results in misclassified objects during the automation of subsequent multi-date images. The protocol was developed through step-wise splitting at each decision node into two classes or more classes using distinguishing feature. The derived classes’ names at each node are dependent on the LCCS class guide lines thus enabling repeatability of methodology across different environs and users. The standardization capabilities of this methodology linked with the automation procedure contribute significantly to the development of satellite image classification approaches. The classifiers or features used in this classification based on the image spectral, shape and texture information, vegetation indices and LCCS classifiers were adequate in performing classification in both the ASTER and Landsat images. The challenge arose in identifying representative threshold levels for the classifiers used. For example, visual analysis and later field verification showed that areas with an NDVI value of 0.09 were characterized by bare areas; rocky outcrops, built-up areas or water in the ASTER image. This threshold value therefore adequately represented the cut-off between primarily vegetated and

621

non-vegetated areas. In the Landsat image the threshold value for this classes’ split was an NDVI value of 0.013. These values are image specific and can be easily adjusted in an existing protocol for use with other multi-date images. These threshold cut-off values were selected based on field verification of a number of objects. This technique of defining and adjusting threshold levels at every node was successfully used before field verification and post classification stage. Each threshold level was verified by cross-referencing with field data. Fig. 2a–c shows classification of an image at the LCCS dichotomous level into four classes. The next stages in the classification are at the modular-hierarchical level as shown in Fig. 2d. The process is continued until 14 classes were derived as shown in Fig. 3. A subset of the image was selected to show in greater detail the classification of an individual object at the initial and final stage of the classification (Fig. 4). This object was finally classified as Open (40  (20–10)% Forest (Woodland). A way of ensuring the use of the same threshold values is by performing radiometric normalization of the same-date datasets to ensure that only real change is detected. Application of methodology to multi-date images is facilitated by adoption of the same protocol, with adjustments to the threshold values at the nodes.

Fig. 2. Dichotomous and modular-hierarchical level classification of the ASTER image.

622

R. Gamanya et al. / Expert Systems with Applications 32 (2007) 616–624

Fig. 3. Final ASTER classification and legend.

Fig. 4. Example of an individual object classification: Final Class Name—Open (40  (20–10)% Forest (Woodland)). (a) Original image with object highlighted in red and (b) final class with classified object. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

4.2. LCCS classifiers Selecting classes at the hierarchical and modular-hierarchical levels highly depends on the purpose of the classification. Skipping a particular level does not adversely affect the classification. There were certain limitations with fully adopting the LCCS class definitions. For example, according to the LCCS definition, the primarily vegetated areas class applies to areas that have a vegetative cover of at least 4% for at least two months of the year. This definition cannot be adopted in the classification of remotely sensed data as this relies on the status of the LULC at a particular time. Objects are therefore either vegetated or non-vegetated at the time the satellite image is taken. Bare fields for example, which would have been classified under primarily vegetated areas in the LCCS are classified under

the primarily non-vegetated group, if there was no vegetation at the time of image capture. The LCCS can result in a tremendous amount of end classes; this flexibility allows the users to determine which classes they can clearly identify from the image and therefore select an appropriate classification route. Depending on the type of satellite image, certain classifiers cannot be used, for example it might not be possible to use the height classifier with most of the medium resolution satellites, unless a digital elevation model (DEM) of the area is available as an additional information layer, the spatial distribution classifier can be used instead. Another example of the flexibility of LCCS system is seen in the classification of herbaceous vegetation, which can be divided into forbs and graminoids, distinction between these two classes is only possible through application of spectral signature

R. Gamanya et al. / Expert Systems with Applications 32 (2007) 616–624

623

Fig. 5. Landsat image subset final classification and legend.

statistics as a feature during the classification process. Because this information was not available, herbaceous vegetation was adequately classified using only the cover aspect, that is, Closed, Open and Sparse classifiers. 4.3. Validation using Landsat image To validate the methodology, a Landsat image subset was classified as shown in Fig. 5. The subset clearly shows three distinct forest types and herbaceous cover. The water hyacinth, a colonizing specie around the lake is distinctly classified into Aquatic or regularly flooded vegetation using a neighborhood classifier and a high band ratio (band 4/band 2) value. Fewer classes were identified in the Landsat compared to the ASTER image as spectral variation depends on sensor resolution. The classification stability of the Landsat image was very high at 91%. Based on field data, the error matrix for the Landsat TM classification showed an overall accuracy of 89% and a KIA value of 84%. In ASTER, the overall accuracy was much higher at 92% and the KIA value was 88%. 4.4. Standardization of methodology It is most likely that subjectivity can remain a weakness unless clear class rules are developed and documented.

Classification rules need to be developed in relation to field data, for example, life form and cover, spatial distribution/ macro-pattern classifiers thus standardizing the methodology. An example is the use of the density feature to differentiate cover, that is, a density value or range of x represents Closed Forest (>(60–70%), y is Open Forest (between (60–70) and is (10–20%)) and z is Sparse (below (10–20) percent but >1%). Field verification of the feature values or ranges can then be used to validate a particular land use and land cover. 4.5. Automation of method Fig. 6a–c shows the results of the automatically classified ASTER image subset at different scales. High accuracy was achieved with the error matrix showing an overall accuracy of 88.7% and a KIA value of 86%. The classification assessment showed over 90% classification stability. A few misclassifications were noted but these were corrected by adjusting the threshold values in the applied protocol. This shows that this methodology can be successfully automated by careful initial protocol development. In the automated ASTER image classification, the issue of scale was assessed. In Fig. 5, a scale parameter of 8 extracted small objects and a scale parameter of 20 resulted in larger segmented objects being classified. Comparison of

Fig. 6. (a) Original ASTER image for automation test. (b) Classified result at segmentation level 1 (scale 8). (c) Classified result at segmentation level 2 (scale 20).

624

R. Gamanya et al. / Expert Systems with Applications 32 (2007) 616–624

the classification accuracies showed that a smaller scale parameter had higher overall accuracy of 91.4% compared to 87.6% for the larger scale parameter. This is attributed the fact that a large scale results in less homogeneous objects or in objects with mixed spectral characteristics. Use of the scale parameter is maximized in eCognition because the normal classification process enables object derivation at different scales derived at different levels and the merging of all objects into a single classification. 5. Summary and future directions This technique produced highly distinct classes with clear and unambiguous class definitions. This approach offers more capabilities due to the large number of classifiers and features available and in both the FAO LCCS and eCognition, which can be successfully combined by users. Standardization of the thresholding process at the nodes needs to be related to field data. Upon identification of verified objects with clear class definitions in the field, it is possible to apply this methodology to future datasets of the same area without the need for additional field verification. It is imperative that local people and users of the technology be involved in the process of field verification and class selection to avoid confusion and promote data standardization. The field verification process therefore plays a crucial role in contributing to the knowledge base especially in multi-spectral and multi-temporal images. For, example, an object classified as Closed Forest (>(60–70%)) can be inventoried based on its documented spectral properties, vegetation index values, density, and other features for future use. The a priori nature of the technique ensures repeatability and avoids class confusion. This approach therefore contributes significantly to object-oriented image analysis and in promoting harmonization of land use and land cover classification methodologies. The automation process is necessary as satellite data acquisition becomes easier and more images are made available. References Aguado, A. S., Montiel, E., & Nixon, M. S. (1998). Fuzzy image segmentation via texture density histograms. EU project Nr. ENV4CT96-0305—Fuzzy Land Information from Environmental Remote Sensing (FLIERS) Final Report, 1998. Anderson, J. R., Hardy, E., Roach, J., & Witmer, R. (1976). A land use and land cover classification system for use with remote sensor data. US Geological Survey Professional Paper, 964. Washington, DC (p. 24).

Baatz, M., & Scha¨pe, A. (1999a). GIS-Aktualisierung und Image Fusion basierend auf einem neuen Verfahren zur objektorientierten und multiskalaren Bildanalyse. In J. Strobl & T. Blascke (Eds.), Angewandte geographische Informationsverarbeitung XI, beitra¨ge zum AGIT Symposium Salzburg 1999 (pp. 12–24). Karlsruhe: Herbert Wichmann. Baatz, M., & Scha¨pe, A. (1999b). Object-oriented and multi-scale image analysis in semantic networks. In Proceedings of the 2nd international symposium on operationalization of remote sensing (pp. 16–20). Enschede: ITC. Baatz, M., & Scha¨pe, A. (2000). Multiresolution segmentation—an optimization approach for high quality multi-scale image segmentation. In J. Strobl et al. (Eds.), Angewandte Geographische Informationsverarbeitung XII. Beita¨ge zum AGIT-symposium Salzburg 2000 (pp. 12–23). Karlsruhe: Herbert Wichmann Verlag. Bezdek, J. (1981). Pattern recognition with fuzzy objective function algorithms. New York: Plenum Press. Canny, J. (1986). A computational approach to edge detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 8, 679–698. Cootes, T., Edwards, G. J., & Taylor, C. J. (1998). Active appearance models. In H. Burkhardt & B. Neumann (Eds.). 5th European conference on computer vision (vol. 2, pp. 484–498). Springer. Cootes, T., Taylor, C. J., Cooper, D. H., & Graham, J. (1995). Active shape models—their training and application. Computer Vision and Image Understanding, 61(1), 38–59. Di Gregorio, A., & Jansen, L. J. M. (1998). Environment and natural resources service, GCP/RAF/287/ITA Africover—East Africa Project and Soil Resources. Management and Conservation Service. FAO, Rome. FAO (2000). Land Cover Classification System (LCCS): Classification concepts and user manual for software version 1. Rome. Gorte, B. (1998). Probabilistic segmentation of remotely sensed images. In ITC Publication series, No. 63. Kauth, R. J., & Thomas, G. S. (1976). The tasseled cap—a graphic description of the spectral–temporal development of agricultural crops as seen in Landsat. In Proceedings of the symposium on machine processing of remotely sensed data, LARS, Purdue University, West Lafayette, IN (pp. 441–451). Li, J., Wang, J. Z., & Wiederhold, G. (2000). IRM: Integrated region matching for image retrieval. In Proceedings of the eighth ACM international conference on multimedia (pp. 147–156). Liu, J., & Yang, Y. H. (1994). Multiresolution color image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 16(7), 689–700. Pal, N. R., & Pal, S. K. (1994). A review on image segmentation techniques. Pattern Recognition, 26(9), 1277–1294. Ronfard, R. (1994). Region-based strategies for active contour models. International Journal of Computer Vision, 13(2), 229–251. Rouse, J. W., Jr., Haas, R. H., Schell, J. A., & Deering, D. W. (1973). Monitoring vegetation systems in the Great Plains with ERTS. In S. C. Freden, E. P. Mercanti, & M. Becker (Eds.), Third earth resources technology satellite-1 symposium technical presentations, section A (vol. I, pp. 309–317). Washington, DC: National Aeronautics and Space Administration, NASA SP-351. Wiens, J. (1989). Spatial scaling in ecology. Functional ecology, 3(8), 120–129.