• No results found

AUTOMATIC CLASSIFICATION OF LAND COVER FEATURES WITH HIGH RESOLUTION IMAGERY AND LIDAR DATA: AN OBJECT-ORIENTED APPROACH

N/A
N/A
Protected

Academic year: 2021

Share "AUTOMATIC CLASSIFICATION OF LAND COVER FEATURES WITH HIGH RESOLUTION IMAGERY AND LIDAR DATA: AN OBJECT-ORIENTED APPROACH"

Copied!
11
0
0

Loading.... (view fulltext now)

Full text

(1)

Sohel Syed, Paul Dare and Simon Jones. Automatic classification of land cover features with high resolution imagery and lidar data: an object-oriented approach.

Proceedings of SSC2005 Spatial Intelligence, Innovation and Praxis: The national biennial Conference of the Spatial Sciences Institute, September, 2005. Melbourne: Spatial Sciences Institute. ISBN 0-9581366-2-9

AUTOMATIC CLASSIFICATION OF LAND COVER FEATURES

WITH HIGH RESOLUTION IMAGERY AND LIDAR DATA: AN

OBJECT-ORIENTED APPROACH

Sohel Syed

1

, Paul Dare

2

, Simon Jones

3

1,2

School of Geography, Population & Environmental Management, Flinders University, GPO Box

2100, Adelaide, SA 5001, Tel: + 61 8 8182 4000, Fax: +61 8 8285 6710

3

Mathematical & Geospatial Sciences, RMIT University, GPO Box 2476V, Melbourne, VIC 3001,

Tel: +61 3 9925 2419,

Email: [email protected]

KEYWORDS

: Image segmentation, object-oriented classification, lidar, high resolution imagery

ABSTRACT

By using high resolution imagery it is possible to detect individual buildings and tree crowns more easily than with conventional lower resolution satellite image data. However, due to the high spatial resolution, automatic classification of such imagery based only on the spectral characteristics of the features can become difficult, especially, in spectrally homogeneous areas. Object-based image processing techniques overcome this problem by incorporating both spectral and spatial characteristics of objects.

In this research, an object-oriented classification scheme was developed which used a lidar derived DSM with multi-spectral image data for the initial segmentation and subsequent object classification. In the data pre-processing stage, the lidar and multi-spectral layers were co-registered into the same coordinate system with uniform spatial resolution. Then a normalised digital surface model (nDSM) was created from the first and last returns lidar data by applying height thresholds. In the multi-resolution segmentation process, the influence of the nDSM and the multi-spectral bands on object generation was controlled by layer weight, scale parameters, the amount of colour and shape factors. As a result, a hierarchical tree of image objects was generated wherein each object knew its neighbouring objects in both horizontal and vertical directions. Within this class-hierarchy each assigned class was described either by one or more fuzzy-membership functions and a nearest neighbour classifier was applied for the final classification. The accuracy of the results using this approach is promising compared to pixel-based maximum likelihood classification. Results indicate that object-oriented approaches has great potential for integration of lidar and high-resolution imagery for multi-sensor classification.

BIOGRAPHY OF PRESENTER

Sohel Syed received his master’s degree in GIS from Curtin University of Technology, Western Australia. He began his PhD with an ARC funded scholarship to the School of Geography, Population and Environmental Management, Flinders University, South Australia. He also began his career as a research assistant in Airborne Research Australia (ARA), where he investigates automated feature extraction algorithms from high resolution imagery. His PhD research topic is “the integration of multi-spectral imagery and laser scanner data for land cover mapping”. Sohel's expertise lies at the interface of GIS and remote sensing for land cover mapping and monitoring, 3D visualisation and automated feature extraction.

(2)

INTRODUCTION

Remote sensing from airborne and spaceborne platforms provides valuable data in various forms and scales for mapping and detecting individual land cover features. The trend has increased dramatically in recent years due to availability of high-density laser scanner and high-resolution multi-spectral data. However, a vast majority of applications rely on pixel-based image processing concepts, which were developed in the 1970s. The main drawback of per-pixel classification in a multi-dimensional feature space is that it doesn’t make use of any spatial concept (Blaschke and Strobl, 2001). Especially in high-resolution images it is very likely that neighbouring pixels belong to the same land cover class due to spatial patterns of differing complexity or texture. Thus, the output from conventional classification algorithms may possess some uncertainty. They also have difficulty adequately or conveniently exploiting expert knowledge or contextual information. Object-oriented image processing overcomes these difficulties by first segmenting the image into meaningful multi-pixel objects of various sizes, based on both spectral and spatial characteristics of groups of pixels. Then the segmented image objects are classified using expert knowledge within fuzzy logic and a hierarchical decision tree. In object-oriented image analysis, multi source data fusion is an automatic side-product of object building (de-Kok et al., 1999). A set of geo-referenced data from different sources defines the topology of image objects, and allows these different types of data to be brought together in a concrete local relation. An advantage of this process is that image objects can be extracted from one data layer, and subsequently in the image analysis step those image objects are able to take into account the attributes of the other data layers (Baatz and Schape, 2000).

The application of traditional automatic classification systems, based on the pixel’s spectral value, have been showing very unsatisfactory results when applied to high resolution images (Rego and Koch, 2003; Blaschke et al., 2001). In high resolution imagery each pixel is related not to the character of object or area as a whole, but to the components of the image. As a result, a lot more classes are detected when it is classified (Smith et al., 2000).

This paper highlights some results of a study where a new automated classification scheme is applied to data from a combination of modern sensors. A new object-oriented classification approach was applied to high resolution multi-spectral image and lidar derived DSM data. An approach for automated land cover feature detection is discussed, using multi-resolution segmentation and a knowledge-based classification technique.

BACKGROUND

Pixel vs. object-based classification

Digital image processing techniques are either pixel- or object-based (Soille, 2003, p. 294). Pixel-based classification techniques utilize spectral pattern value combinations associated with different feature types, each assigned a unique Digital Number (DN), evaluating spectral reflectance values present within each pixel to find meaningful patterns. A class of pixels is determined from the overall DN values and statistics are derived from the spectral characteristics of all pixels in the image. Maximum likelihood classification (MLC) was one of the common in pixel-based spectral classification schemes. Object-based classification starts with the grouping of neighbouring pixels into meaningful areas. This means that the segmentation and subsequent object topology generation is controlled by the resolution and the scale of the expected objects (deKok et al., 1999). In an object-based classified image, the elementary picture elements are no longer the pixels, but connected sets of pixels (Fig.1). Once the image has been segmented, measurements are performed on each region and adjacency relations between regions can be investigated (Soille, 2003, p. 287).

Figure 1: An illustration of the hierarchical network of image object.

Level 1

Level 2 Level 3 Pixel level

(3)

Nowadays many new segmentation algorithms as well as applications have been developed, but not all of them lead to convincing results while being robust and operational. One reason is that the segmentation of an image into a given number of regions is a problem with a huge number of possible solutions. The high degrees of freedom must be reduced to a few which satisfy the given requirements (Blaschke and Strobl, 2001). A new algorithm called “the fractal net evolution approach” (Baatz and Schape, 2000) has apparently achieved this requirement by defining homogeneity in combination with local and global optimization techniques. In this algorithm, a scale parameter is used to control the average image object size and homogeneity criteria are defined by the spectral and spatial information.

Spectral information sometimes is not enough to recognize objects within high spatial resolution aerial and satellite images (<5m GSD). In this situation, illumination and sun angle plays a vital role in recognizing particular land cover objects. Normally in pixel based spectral analysis, the same objects are recognized as different objects due to this factor. The higher spatial resolution generally reduces the problem of mixed pixels, but the internal variability and the noise within individual land cover classes are increased (Schiewe et al., 2001). As a consequence, traditional classification approaches are producing too many or not well defined classes, because their clusters are built upon spectral homogeneities only. The major drawback of pixel-based classification techniques is that they do not take spatial information into account. A pixel is classified depending on its spectral values irrespective of the values of the neighbouring pixels. Classifications are therefore very sensitive to noise and they often lack of spatial consistency (Soille, 2003, p.302). As an alternative to pixel based classification, object-based classification methods improve accuracy and interpretability in high resolution images (Aplin et al., 1999). It also s

Another advantage of object-based classification methods is to incorporate fuzzy classification system. Fuzzification describes the transition from a crisp system to a system. Fuzzy logic using graded or qualified statements rather than ones that are strictly true or false. The membership functions of fuzzy sets are chosen either on the basis of expert knowledge, or by using methods of numerical taxonomy (Burrogh and McDonnel, 1998). The first approach is known as Semantic Import (SI) model, where classes based on expert knowledge are usually imposed or imported classes that are set up without direct reference to the local data set. The second approach is known as Similarity Relationship (SR) model, where the value of membership function is a function of the classifier used.

Multi-spectral image and lidar data integration

Lidar is a recent development in remote sensing with great potential for creating high resolution and accurate digital elevation models (DEMs). It has proven a mature state of art technology to acquire three-dimensional point clouds for describing the Earth’s surface. The most dramatic feature of recent airborne lidar systems are their ability to discriminate between first and last pulse reflections. A laser pulse that is fired over an object usually has multiple reflections. Some of the laser pulse may be reflected by the top of the object and therefore represents the first returning pulse. The remainder is likely to be reflected by the ground and hence generates the last returning pulse. In this way, lidar produces a fast and highly accurate three-dimensional surface (Rottensteiner et al., 2003). Lidar data provides accurate measurement of landcover structures in the vertical plane; however, current lidar sensors have limited coverage in the horizontal plane. Multi-spectral data provide extensive coverage of landcover classes in the horizontal plane but are relatively insensitive to variation in their height. Therefore, the integration of lidar and multi-spectral data can greatly improve the measurement and mapping of landcover classes.

In land cover feature classification, the combination of lidar and high-resolution satellite imagery has a promising future. Lidar provides very accurate position and height information, but less direct information on the object’s geometrical shape, while high-resolution imagery offers very detailed information on objects, such as spectral signature, texture, shape etc. Combining these two kinds of complementary datasets is quite promising for building extraction, 3D city modelling etc.(Tao and Yasuoka, 2002). Rottensteiner et al. (2003) integrated lidar and high-resolution multi-spectral data for the automatic detection of buildings with heterogeneous appearances; a hierarchic integration technique was adopted to detect buildings in urban settings. Hofmann (2001) used an object-based classification scheme to detect buildings and roads in IKONOS data using additional elevation information.

In light of these findings, the research presented in this paper describes the use of high resolution multi-spectral imagery combined with a lidar derived DSM for initial segmentation and subsequent object-based classification. This approach has been adopted to make a more accurate classification of land-cover features for landuse mapping.

(4)

STUDY AREA

The study area is situated in the small town of Mathoura in southern New South Wales. Figure 2 shows the geographical location of Mathoura and the study area. The study area is a part of Mathoura’s central town and covers around 9600m2. This area was part of a large data sets combine with aerial photographs and lidar data. The sample area is a good mixer of land cover features, which consists of rail line, storage sheds and silos, office building, vegetated areas and open space.

Figure 2: The study region at Mathoura.

DATASETS AND SOFTWARE

High resolution imagery

The image data, collected by AEROmetrex, was acquired using a Zeiss LMK 152 camera with a calibrated focal length of 152.261mm. It was captured at an approximate flying height of 850m above the ground level equating to an average scale 1:5500, which was later scanned at 15µm to provide a pixel size of 8.25 cm. The photographic images were orthorectified with the help of accompanying exterior orientation parameters (Xo, Yo, Zo, , , ), which were captured using onboard GPS and IMU sensors. Figure 3(a) shows the photographic dataset of the study area.

Lidar data

Lidar data was acquired in July 2001 by AAMGeoScan (now AAMHatch). The lidar system used was the ALTM 1225, which operates with a sampling intensity of 11000 Hz at a wavelength of 1.047 µm. Approximate flying height of this sensor was 1100m and the laser swath width was 800m. Vertical accuracy was 0.15m (1 ), the internal precision was 0.05m, and the original laser footprint was 22cm in diameter. The data was subsequently resampled to a 1m grid. More details of the lidar data and its processing information can be found in Molloy and Bruce (2004). The processed lidar image is shown in figure 3(b).

In this research, eCognition (http://www.definiens.com) Professional version 4.0 was used for the object-oriented classification, which included multi-resolution segmentation and knowledge-based object extraction. ERDAS Imagine 8.7 was used for pixel-based maximum likelihood classification and image processing including image co-registration and resampling.

NSW

VIC

Study Area

(5)

Figure 3: Photographic and lidar data sets of the study area.

METHODOLOGY

The proposed method is based on the integrated use of high-resolution imagery and lidar data for object-oriented land cover classification. Object-oriented classification requires the identification of meaningful objects over the image and classifying them with class attributes. Therefore, the overall procedure consists of a sequential application of

segmentation and classification. A flow chart summarising the proposed object-oriented classification technique is provided in Fig. 4.

Figure 4: Flow chart for the object-oriented classification. (*) Denoted as a process number, which later referred into text. (a)

Photographic image (R1,G2,B3)

(b)

normalized Digital Surface model (nDSM)

Lidar DSM High res. image Multi-resolution segmentation (1) Colour Non-vegetation Non-colour (4) Mean & ratio of B2 (2) Grey surface Open space Grey roof (6) (5)

Tree Colour roof Grey roof

Neighbour(7) Non-grey surface Vegetation Tree Grass Stddev of B2 (3) Brightness & ratio of B3 Brightness & ratio of B3 Mean height diff.

(6)

Multi-resolution segmentation

The basic processing units of object-oriented image analysis are segments or image objects, and not single pixels. In the segmentation process, for each image object a meaningful statistic is calculated in an increased uncorrelated feature space using shape, texture and topological features. This information improves the value of the final classification and cannot be fulfilled by common, pixel-based approaches (Benz et al., 2004).

Figure 5: Segmented image of the study area

Image-object primitives (Fig. 5) are created through multi-resolution segmentation (see fig. 4(1)). These objects are polygons of roughly equal size exhibiting interior homogeneity. In the segmentation process, the user selects the size and homogeneity criteria of the objects. For size, user specifies the scale parameter, which determines the maximum allowed heterogeneity for resulting image objects. By modifying the value of the scale parameter the user can vary the size of the image objects. For homogeneity, the user chooses the relative weight to apply spectral versus shape criteria to reduce heterogeneity. Here shape, smoothness and compactness criterion are applied in a mixed form to define homogeneity for the image objects.

Layers Resolution Std. Dev Weight

MS band 1 0.4m 64.2 1.0

MS band 2 0.4m 64.1 1.0

MS band 3 0.4m 69.8 1.0

nDSM 0.4m 2.2 5.5

Table 1: Image layers description for segmentation.

Equal weight was assigned to each of the multi-spectral bands (Table 1). This emphasis was chosen because of the lack of colour homogeneity visually observed within the same features in the image. On the contrary, lidar derived nDSM layer had more homogeneity in grey level; therefore more weight was given to this layer (Table 1). By visual interpret different image segmentation results, a scale parameters 15 was chosen to create local homogeneity and to keep global heterogeneity. Similarly, a ratio of smoothness to compactness weight is defined. Here, 3:7 was specified (Table 2), emphasizing the discrete, compact nature of building roofs. A higher smoothness emphasis would be used to define objects observed to have greater variability between features (Baatz et al., 2004). The compactness weight made it possible to separate objects that had quite different shapes but not necessarily a great deal of colour contrast, such as building roofs versus roads within the study area.

(7)

Scale Parameter 15 Shape Factor 0.275

Compactness 0.3 Smoothness 0.7

Table 2: Segmentation parameters

Class hierarchy

The class hierarchy is the frame of object-oriented classification to create the knowledge base for a given classification task. It contains all classes and is organized in a hierarchical structure (Baatz et al., 2004). The class hierarchy passes down class descriptions from parent classes to their child classes. It reduces the redundancy and complexity in the class descriptions and creates a meaningful grouping of classes.

Figure 6: The class hierarchy for the object-oriented classification.

In this research, the class hierarchy was developed through inheritance hierarchy, which refers to the physical relations between the classes. Initially, vegetation and non-vegetation were the parent classes and they further divided into grass,

tree and colour, non-colour classes as child classes (Fig. 6). Later, vegetation class became a parent class and grass and

tree were the child classes. Within this class-hierarchy each class is described either by one or more fuzzy-membership

functions, a nearest neighbour classifier or by a combination of both. Membership functions are determined by the Semantic Import (SI) model, which is based on the expert knowledge of the features. A stepwise refinement of the class-hierarchy was achieved using the inheritance mechanism.

Classification based on spectral properties

Since the generated image objects hold more spectral information compared to pixels’ digital numbers, the object-oriented classifier offers a huge variety of derivative spectral features (Hofmann, 2001). Brightness and spectral ratios of the image objects were calculated using all image layers. Textural features were calculated using standard deviations of layer values, spectral mean values of sub-objects, and average spectral differences of sub-objects. Contrast information were generated though spectral differences to neighbouring objects and super-objects. Context related features included mean spectral differences to a given class.

In this research, spectral nearest neighbour classification was used for its efficiency in simple classes. Initially, vegetation and non-vegetation were separated by a fuzzy membership description of mean and ratio of the green spectral band. A image object was represented as vegetation if its mean of the green band was larger or equal to 30 and also the ratio was larger or equal to 0.37 (see fig. 4(2)). Vegetation was subdivided further as grass and tree classes. This was described by the standard deviation of the green band (see fig. 4(3)). In fuzzy membership function a green object represented grass if its standard deviation of the green band was smaller or equal to 18. On the contrary, for a tree the standard deviation was larger or equal to 18 (Fig. 7).

(8)

Grass Tree

Figure 7: Fuzzy membership functions for grass and tree classes.

Colour Surface was the sub-class of non-vegetation class. This class was defined by the object’s brightness and the ratio of the blue band (see fig. 4(4)). For this class, the brightness was larger or equal to 70 and the ratio of the blue band was larger or equal to 0.38. Grey Surface was defined by the object’s brightness and the ratio of the red band (see fig. 4(5)): the brightness is larger or equal to 70 and the ratio of band 3 is larger or equal to 0.28. Shadow was the subclass of non-Grey surface class and is defined by the brightness. An image object was a shadow if its brightness was less or equal to 70.

Classification based on DEM-properties

When taking the DEM information into account, it is obvious that it is not the absolute but the relative height within classes, which characterises them. In eCognition this property can be modelled by describing the difference in elevation to neighbouring objects (Hofmann, 2001). In this research, rooves and open areas were discriminated by the mean difference of the nDSM layer (see fig. 4(6)). A image object was represented as open space if its mean difference of nDSM was smaller or equal to 1.2m (Fig. 8). The definition of roof was the opposite of the open space.

Open-space Roof

Figure 8: Fuzzy membership function for discriminating between open-space and roof area.

Classification enhancements using contextual information

As shadows are typically created by other features, most of them can be detected by describing just their surrounding features. Additionally, shadows may vary their spectral properties. Thus it is useful to classify shadows by describing their contextual criteria and subsequently their different spectral properties. Therefore, depending on the type, shadows may inherit their spectral properties from an appropriate super-class and then be identified by their surroundings. The logic applied for this was “if an object classified as shadow is surrounded sufficiently by objects classified as building, it should be classified as building shadow”. In this research, colour roof shadow was a subclass of shadow. It was defined by the neighbour-object relation (see fig. 4(7)). A shadow object is classified, as a grey roof shadow if its border to neighbour-object relation for grey object is larger or equal to 0.025m. A shadow object is classed as tree shadow if its border to vegetation neighbour-object is larger or equal to 0.02m.

Left border Centre point Right border

Fu zz y m em be rs hi p va lu e

Left border Centre point Right border

Fu zz y m em be rs hi p va lu e

(9)

Pixel-based classification

In this research, training sites were selected after careful analysis of topographic maps and aerial photographs. The training regions were shown to contain an adequate number of pixels and to be spectrally separable. This issue is important in supervised classification to avoid the misinterpretation of land cover features for those with a similar spectral signature. In a similar way, statistics were generated for the high resolution image. After running the maximum likelihood classification scheme with equal prior probability, the classified image was generated with eight types of land cover features, the same as the number for the object-oriented classification.

RESULTS AND DISCUSSION

The results of both the object-oriented classification and maximum likelihood classification are shown in Figure 9. Statistics in Table 3 show the mean, standard deviation, minimum and maximum fuzzy logic membership values. The fuzzy membership values are between 0 (totally ambiguous) and 1 (totally unambiguous). Figure 9(b) shows in map form the object-oriented classification of each of the objects. The grey, colour roof, colour roof shadow, and tree shadow, were all classified with the highest stability, as noted by their high mean, minimum and maximum membership values and low standard deviations. The grass class (0.906) and tree class (0.926) had the lowest mean membership values, suggesting that this class should be refined to attain more stable image object memberships.

Figure 9: Comparison between maximum likelihood classification (a) and the object-oriented classification (b).

From visually inspecting the two classification results (Fig. 9), the main difference is the sharpness of the features, which look clearly better delineated in the object-oriented classification due to the combined use of the lidar derived DSM with multi-spectral image. This is particularly evident when observing the open space and the building roofs in the pixel-based classification. Open space and roofs were mixed like a ‘pepper and salt’ effect on to the ground. The problem of the pixel based classification was also clear in the shadow classes, which were poorly generated. Conversely, when using contextual information in the object-oriented classification, these classes were perfectly identified.

(b)

Object-oriented classified image

Open space Grey roof

Grey roof shadow Colour roof

Colour roof shadow Grass

Tree

Tree shadow

(a)

Maximum Likelihood classified image

(10)

Feature Type No. of objects Mean SD Min Max.

Open space 128 0.999 0.0740 0.196 1

Grey roof 33 1 0 1 1

Grey roof shadow 28 0.999 0.0008 0.995 1

Colour roof 5 1 0 1 1

Colour roof shadow 3 1 0 1 1

Grass 14 0.906 0.2060 0.388 1

Tree 47 0.926 0.2310 0.107 1

Tree shadow 10 1 0 1 1

Table 3: Statistics for all objects in object-oriented classification.

In this study, object-oriented scheme was found more flexible than the MLC in automatic classification. The improvement can be achieved with a relatively simple and unrefined application of the object-based classification approach. It allows highly automated classification in association with highly refined and specialized membership functions, which in a more complex case should increase the accuracy even further. There is a great potential for further improving object-oriented classification quality through refining the decision rule structure, whereas MLC offers very little potential for improvement in this case other than through a procedure that tries to imitate the object-oriented approach.

CONCLUSION AND FURTHER RESEARCH

The goal of this study was to automatically classify land cover features using high-resolution imagery and lidar derived DSM data, and the object-oriented approach did this effectively. A semi-rural image subset was analysed using two classification schemes, resulting in much more accurate land cover maps than are attainable using the pixel based MLC. Shadow objects are particularly susceptible to misclassification when using pixel spectra. Even so, object-oriented classification has proven effective in correctly identifying the shadows where this is difficult by other methods. Although a very simple approach was used to create the classification hierarchy in this study, significant improvement over MLC was obtained. Object-oriented classification allows automated classification in association with highly refined and specialized membership functions, which in a more complex case, should increase the margin of accuracy even further.

Further field survey is being conducted to validate the classification results generated from object-oriented and MLC. From this object-oriented approach, it is envisaged that incorporating lidar data into the high-resolution imagery can make the automatic feature extraction more effective. It is anticipated that this method will facilitate a flexible and cost-effective production of detail maps demanded by the emerging users of spatial industries.

ACKNOWLEDGEMENT

This work was supported by an Australian Research Council (ARC) Discovery Project (DP0450889).

REFERENCES

Aplin, P., Atkinson, P.M. and Curran, P.J. (1999) "Per-field classification of land use using the forthcoming very fine spatial resolution satellite sensors: problems and potential solutions", in: Advances in remote sensing and GIS Analysis, Atkinson, P. M. and Tate, N. J. (eds.), John Wiley & Sons, Chichester, p. 219–239.

Baatz, M., Benz, U., Dehghani, S. and Heynen, M. (2004) eCognition User Guide 4, Definiens Imagine GmbH, Munchen, Germany,pp.

Baatz, M. and Schape, A. (2000) Multiresolution Segmentation: an optimization approach for high quality multi-scale image segmentation, Proceedings of Angewandte Geogr. Informationsverarbeitung XII, Strobl, J. and Blaschke, T. eds., Wichmann, Heidelberg, pp. 12-23.

Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I. and Heynen, M. (2004) Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information, ISPRS Journal of Photogrammetry and Remote Sensing,Vol. 58, p. 23.

(11)

Blaschke, T., Conradi, M. and Lang, S. (2001) Multi-scale image analysis for ecological monitoring of heterogeneous, small structured landscapes, Proceedings of SPIE, Toulouse, pp. 35-44.

Blaschke, T. and Strobl, J. (2001) What’s wrong with pixels? Some recent developments interfacing remote sensing and GIS, Proceedings of GIS – Zeitschrift für Geoinformationssysteme 6/2001, Hüthig GmbH & Co. KG, Heidelberg, pp. 12-17.

Burrogh, P.A. and McDonnel, R.A. (1998) Principles of Geographical Information Systems, Oxford University Press, New York, USA. 333 pp.

deKok, R., Schneider, T. and Ammer, U. (1999) Object based classification, applications in the alpine forest environment, International Archives of Photogrammetry and Remote Sensing,Vol. 32, No. 7-4-3 W6, 3-4 June, 1999.

Hofmann, P. (2001) Detecting buildings and roads from IKONOS data using additional elevation information, Proceedings of GIS – Zeitschrift für Geoinformationssysteme 6/2001, Hüthig GmbH & Co. KG, Heidelberg, pp. 28-33.

Molloy, N. and Bruce, D. (2004) Application and analysis of three-dimensional object reconstruction using two dimensional imagery in a tactical environment, Proceedings of The 12th Australasian Remote Sensing and Photogrammetry Conference, Smith, R. and Dawbin, K. eds., Fremantle, Western Australia, 18-22 October 2004.

Rego, F. and Koch, B. (2003) Automatic classification of land cover with high resolution data of the Rio de Janeiro city Brazil comparison between pixel and object classification. In:, Proceedings of The International archives of the photogrammetry, remote senisng and spatial infoamrion sciences, Carstens, J. ed., Regensburg, Germany, 27-29 June.

Rottensteiner, F., Trinder, J., Clode, S. and Kubik, K. (2003) Building detection using LiDAR data and multi-spectral images, Proceedings of VIIth Digital Image Computing: Techniques and Applications, Sun, C., Talbot, H., Ourselin, S. and Adriaansen, T. eds., Sydney, 10-12 Dec., pp. 673-682.

Schiewe, J., Tufte, L. and Ehlers, M. (2001) Potential and problems of multi-scale segmentation methods in remote sensing, Proceedings of GIS – Zeitschrift für Geoinformationssysteme 6/2001, pp. 34-39.

Smith, G.M., Fuller, R.M., Hoffmann, A. and Wicks, T. (2000) Parcel-based approaches to teh analysis of remotely sensed data, Proceedings of The Remote Sensing Society Conference, Nottingham, UK.

Soille, P. (2003) Morphological image analysis: principles and applications (2nd Edition), Springer, Berlin ; New York. 316 pp.

Tao, G. and Yasuoka, Y. (2002) Combining High Resolution Satellite Imagery and Airborne Laser Scanning Data for Generating bareland DEM in Urban Areas, Proceedings of International Workshop on Visualization and Animation of Landscape, International Archives of Photogrammetry, Remote Sensing and Spatial Information Science, Kunming, China, 26 - 28 February 2002.

References

Related documents