Land cover classification of QuickBird multispectral data ... · PDF fileLand cover classification of QuickBird multispectral data with an object-oriented approach E. Tarantino Polytechnic

Land cover classification of QuickBird multispectral data with an object-oriented approach

E. Tarantino Polytechnic University of Bari, Italy

Abstract

Recent satellite technologies have produced new data as an alternative to aerial photographs. These Very High Resolution (VHR) data offer widely acceptable metric characteristics for cartographic updating and increase our ability to map land use in geometric detail and improve accuracy of local scale investigations. Even if it is possible to distinguish spatial entities obscured in the past by low resolution sensors, high spatial resolution involves a general increase of internal variability within land covers. Moreover, accuracy of results may decrease on a per-pixel basis. If the full potentiality of the new image data for urban land use mapping is to be realised, new inferential remote-sensing analysis tools need to be investigated. This study analyses a procedure of classification which aims to reduce negative impacts related to spectral ambiguity and spatial complexity of land cover classes of VHR imagery. In order to avoid such disadvantages, an object-oriented classification methodology on QuickBird multispectral data has been implemented on a peripheral area of Bari, in the region of Apulia (Italy). The analysis consists of two phases: multiresolution segmentation and nearest neighbour classification of the resulting image-objects. With this procedure the generated segments act as image objects whose physical and contextual characteristics can be described by means of fuzzy logic. Validity and limitations of such methodology for land cover classification is finally evaluated

1 Introduction

The applications of data obtained from satellite platforms at 300 to 600 miles above the earth have remained limited for urban areas until now for a variety of

© 2004 WIT Press, www.witpress.com, ISBN 1-85312-728-0Management Information Systems, C. A. Brebbia (Editor)

reasons, including low resolution of images complexity of ground features, and technological differences with conventional photography. With satellite imagery technology the different bands are recorded synchronously so that their pixels may be precisely matched up and compared with their counterpart pixels in other bands. This means that spectral ("colour") differences can be used to identify urban features to the extent that colours are diagnostic - sort of a coarse spectroscopy from space. Spectral sensor technology, however, coupled with the complexity of ground features in urban areas, can make visual interpretations of satellite imagery both labour intensive and uncertain. To implement image interpretation, commercial or free software can be used to "train" computers to recognise spectral signatures from sample locations. The software program will then read through an entire satellite dataset to tag other pixels that have similar characteristics [11]. This procedure works well with images of agricultural crops or forests in rural areas, but computers have had trouble distinguishing between different urban features. For example, roofs of building are small compared to satellite pixels and have many colours and shadings due to different materials used and to different orientation towards the sunlight. As a result, this approach has had limited success in dealing with data for urban environments [10, 13]. During the past few years improvement in the resolution of satellite images has broadened applications for satellite images to areas such as urban planning, data fusion with aerial photos and digital terrain models (DTMs), and the integration of cartographic features with GIS data. However, previous high-resolution satellites, such as 1-m resolution IKONOS, still could not replace the use of aerial photos, which have resolution as high as 0.2 to 0.3 m. The successful launch of QuickBird and its high resolution sensors has narrowed the gap between satellite images and aerial photos, permitting easy updating of urban information. The result is that the panchromatic resolution is increased from 1-metre to 70-centimetre and multi-spectral from 4-metre to 2.80-metre resolution. Panchromatic imagery is collected in 11-bit format (2048 grey levels) and delivered in 16 bit format for superior image interpretation (shadow detail, etc.), or 8-bit format (256 grey levels) supported on desktop GIS and mapping application packages. Four-band multi-spectral imagery consists of blue, green, red and near-infrared bands, delivered in 16-bit and 8-bit formats. Such data are likely to stimulate the development of urban remote sensing [9], but optimal results still usually need human visual interpretation of the computer enhanced images. For these reasons, much of the research in satellite data analysis has focused on agricultural and forested areas, where the spectral responses from large fields with homogeneous types of vegetation provide relatively uncomplicated landscapes. If the full potential of the new image data for urban land use mapping is to be realised, new inferential remote-sensing analysis tools need to be investigated [5]. At the same time, it has been found that the land cover classification of VHR images results in a complex structural composition, which might inhibit recognition of distinct urban land use categories. High resolution of the data does not automatically lead to higher classification accuracy. This is due

126 Management Information Systems


to the heterogeneity of objects within an urban area which leads to either misclassified pixels or unwanted details [8, 17]. In this paper a classification procedure aimed at reduction of negative impacts from land cover classification of VHR imagery was investigated. In order to avoid such disadvantages, an object-oriented classification methodology on QuickBird multispectral data has been implemented on the test area of Bari, in the region of Apulia (Italy). The analysis is conducted with eCognition software and consists of two phases: multiresolution segmentation and nearest neighbour classification of the resulting image-objects. With this procedure the generated segments act as image objects whose physical and contextual characteristics can be described by means of fuzzy logic. Validity and limitations of such methodology for land cover classification are finally drawn.

2 From pixel -based to object-based classification

In remote sensing techniques the informational classes of a thematic mapping are not directly registered, but must be derived indirectly by using evidence contained in the spectral data of an image. When standard procedures of per-pixel multispectral classification are applied to VHR data, the increase of spatial resolution leads to augmentation in ambiguity in the statistical definition of land cover classes and a decrease of accuracy in automatic identification. These sources of imagery are likely to generate other problems. Even if the radiometric resolution is enhanced (11 bits for QuickBird imagery), spectral capabilities are generally limited compared to those of the previous generation sensors (seven bands for the Landsat TM). Moreover, associated with an increase in spatial resolution there is, usually, an increase in variability within land parcels (‘noise’ in the image), generating a decrease in accuracy of land use classification on a per-pixel basis [16]. In order to solve such problems, some post-classification procedures were investigated on the basis of intrinsic contextual information of data [15]. Although a reduction of noise in the classified image was obtained, substantial improvements in overall accuracy of the results was not see. Moreover, a loss of meaningful information in classified data was shown because of geometric and dimensional non-correspondence of real elements with moving window implementation matrix (for example with majority logical filter). Such methods need extensive editing operations on classified images in order to be stored in GIS databases. An alternative technique to per-pixel classification is the per-field classification (so called because fields, as opposed pixels, are classified as independent units), and this takes into account the spectral and spatial properties of the imagery, the size and shape of the fields and the land cover classes chosen. In fact, this approach requires a priori information on the boundaries of objects in the image. ‘Field” or ‘parcel’ refers to homogenous patches of land (agricultural fields, gardens, urban structures or roads) which already exist and are superimposed on the image. Some studies [2, 7] indicate that this

Management Information Systems 127


methodology is contributing positively to the classification of remote sensing imagery of high to medium geometric resolution. Problems arise in cases where no boundaries are readily available or when exactly those boundaries should be updated [6]. One solution is image segmentation. In many cases, image analysis leads to meaningful objects only when the image is segmented in ‘homogenous’ areas [3, 12]. Because an ‘ideal’ object scale does not exist, objects from different levels of segmentation (spatially) and of different meanings have to be combined for many applications. The human eye recognises large and small objects simultaneously but not across totally different dimensions. From a balloon for instance, the impression of a landscape is dominated by land use patterns such as the composition of fields, roads, ponds and built up areas. Closer to the ground, one starts to recognise small patterns such as single plants while simultaneously small scale patterns loose importance or cannot be perceived anymore. In remote sensing, a single sensor highly correlates with a specific range of scales. The delectability of an object can be treated relative to the sensor’s resolution. A coarse rule of thumb is that the scale of image objects to be detected must be significantly bigger than the scale of image noise relative to texture. This ensures that subsequent object oriented image processing is based on meaningful image objects. Therefore, among the most important characteristics of a segmentation procedure is the homogeneity of the objects [6]. The resulting segmentation should be reproducible and universal, thus allowing application to a large variety of data [1, 3]. Moreover, multiresolution image processing based on texture and utilising fractal algorithms can alone fulfil all main requirements at once. Their ‘fractal net evolution approach’ uses a local mutual best-fit heuristics to find the least heterogeneous merge in a local vicinity following the gradient of the best-fit. Furthermore, their algorithm can be applied with pure spectral heterogeneity or with a mix of spectral and form heterogeneity [6].

3 Data and methods

The study area of this research is in Italy, i.e. part of the peripheral district of Bari, in the region of Apulia Italy). The flat geomorphology and settling features justify the choice of this area, as t permits testing of the validity of advanced classification methodologies on a over-regional scale for the new generation satellite sensor imagery. The data sources acquired for the analysis consist of:

a) QuickBird multispectral images (Acquisition Date/Time: 2002-08-01 / 17:11:18). To reduce the size of image to be manipulated and, to maintain geographic consistency during analysis, equivalent sub-scenes were defined, fig. 1.

b) training set and the ground reference data obtained by land use survey (September 2002) and additional maps.



Figure 1: QuickBird multispectral image processed in RGB mode- Bari (Italy).

3.1 Classification strategy

Object-oriented image analysis consists of two steps, segmentation and (fuzzy) classification of the resulting image-objects. The multiresolution segmentation algorithm implemented in eCognition software extracts homogenous image-objects at different resolution hierarchy-levels respectively. For each segment (object) object-features are calculated and stored in a database. Compared to conventional pixel-based classification approaches, utilizing only the spectral response, image objects contain additional information, like, among others, object texture, shape, relationship to neighbours and sub-objects Fuzzy classification is based on the different object-features available. In contrast to pixel-based statistic classifiers, fuzzy classification replaces the strict class definitions of "yes" and "no" by a continuous range where all numbers between 0 and 1 describe a certain state of class membership [4]. In this paper class description is performed using the fuzzy approach of nearest neighbour that supports a classification approach based on marking typical objects as representative samples. The nearest neighbour (NN) classifier is able to evaluate the correlation between object features and handles the presence of overlaps in the feature space much more easily. Moreover, it allows very fast and easy handling of the class hierarchy for classification (without class - related features). With the aim of exploring the potentiality of multiresolution segmentation of QuickBird multispectral data, different scale parameters were implemented, by



producing three levels of spatial representation, as shown in figures 2, 3, 4. Such algorithms join neighbour regions that show a degree of fitting – computed with respect to their spectral variance and/or their shape proprieties – which is smaller than a pre-defined threshold (scale parameter). The choice of this parameter determines the number and size of resulting segments [14]. In order to assess the capability to discriminate urban features, level 1 of segmentation for the following classification was selected. Classes separated by object-features can represent parts of a thematic unit and can be combined into semantic groups. The analysed territory includes three main semantic groups of land use (urban settlements, agricultural fields and airport), which include low level objects. Following the classification methodology of eCognition software, the definition of classes is based on the possibility of distinguishing one class from another in certain features. This is realized by developing a knowledge base suitable for each classification strategy enabling description of each class. Seven land use classes for the study area were individuated (asphalt roads, country roads, buildings, meadows, uncultivated land, arable land, olive-groves). Generally, remotely sensed reflectance is related to land cover and not to land use, but in the present case each land use class was assumed to correspond to spectrally-separable land covers. Additional rule-set definitions were introduced, such as relations to neighbour objects and interpretation of deduced information (vegetation indices; band ratios; relevant land use, such as the airport complex). After this, the nearest neighbour (NN) classifier was trained on the representative sample of each of the land use classes, fig. 5.

Figure 2: Multiresolution segmentation of QuickBird multispectral data (level 1).







Figure 5: NN classification (level 1) of QuickBird multispectral data.

4 Results

The presented classification results show the ability of the so far developed rule-sets to distinguish between objects and to generate semantic classes. This is demonstrated by the high value of the overall accuracy (0,962) of the classification, as shown in table 1.

Table 1: Results of NN classification – level 1: error matrix and statistics.



There are various causes for obvious "classification" errors, with relevant difficulty in distinguishing between different urban features. For example, roofs of building have many colours and shadings due to different materials used and their different orientation towards the sunlight. These proprieties caused intense overlaps with the asphalt road class. For spectral separable classes, such as natural features, like meadows, olive grove, uncultivated land a good recognition rate was obtained. The arable land class shows a classification accuracy with light overlaps to country roads. Semantic classes that present errors often contain different class-types causing a partly incorrect representation. For example a class for urban settlements can include trees or grass. Object-oriented classification distinguishes between these objects and classifies trees or grass (as meadow) depending on the object-level.

5 Conclusions

The semantic classes derived from satellite data cannot exactly fit the classes for large scale investigation, because of the resulting aggregation of information. Very high resolution satellite data contain significantly more detailed information than previous satellites (e.g. Landsat TM) but problems arise from their high spectral variance. The so called "salt & pepper" effect is a handicap for the recognition of semantic classes. Object oriented classification systems enable us to analyse high resolution data and avoid some of the above mentioned problems. Based on a preceding segmentation, a fuzzy approach of nearest neighbour is used to classify the QuickBird data. The primary task is the creation of object lists that take the abilities of different remote sensing data into account. Major problems are the generation of meaningful objects that fit semantic classes of the object lists. Classification rules for all object types and their representation in VHR data from different sensors and on varying acquisition times have to be developed. The next step will be the assignment of the deduced classification rules on more complex scenes. Additional knowledge from thematic layers has not been used in this study so far but will be included in our future work.

References

[1] Andresen, T., Mott, C., Zimmermann S., Schneider, T., Monitoring of reed population on Bavarian lakes with high-resolution satellite data, Proc. of ISPRS Workshop “High Resolution Mapping from Space 2001”, Hannover, September 19-21, (CD-ROM), 2001.

[2] Aplin, P., Atkinson, P., Curran, P., Per-field classification of land use using the forthcoming very fine resolution satellite sensors: problems and potential solutions, Advances in remote sensing and GIS analysis, eds. Atkinson, P. & Tate, N., Wiley & Son, Chichester, pp. 219-239, 1999.



[3] Baatz, M., Schäpe, A., Multiresolution Segmentation – an optimisation approach for high quality multi-scale image segmentation, Angewandte Geographische Informationsverarbeitung, eds. Strobl, J., Blaschke, Griesebner, XII, Wichmann-Verlag, Heidelberg, pp. 12-23, 2000.

[4] Baatz, M., Heynen, M., Hofmann, P., Lingenfelder, I., Mimler, M., Schäpe, A., Weber, M., Willhauck, G., eCognition User Guide. München, Definiens AG, 2000.

[5] Bauer, T., Steinnocher K., Per-parcel land use classification in urban areas applying a rule-based technique, GeoBIT/GIS, 6, pp. 12-17, 2001.

[6] Blaschke, T., Lang, S., Lorup, E., Strobl, J., Zeil, P., Object-oriented image processing in an integrated GIS/remote sensing environment and perspectives for environmental applications, Environmental Information for Planning, Politics and the Public, eds. Cremers, A., Greve, K., Metropolis Verlag, Marburg, 2, pp. 555-570, 2000.

[7] Caprioli M., Tarantino E., Accuracy assessment of per-field classification integrating very fine spatial resolution satellite sensors imagery with topographic data, Journal of Geospatial Engineering, 3 (2), pp. 127-134, 2001.

[8] Donnay, J.-P., Use of remote sensing in-formation in planning, Geographical In-formation and Planning, eds. Stillwell, J., Geertman, S. & Openshaw, S., Berlin, Springer-Ver-lag, pp. 242–260, 1999.

[9] Fritz, L. W., High resolution commercial remote sensing satellites and spatial information, 1999. www.isprs.org/publications/highlights/ highlights0402/fritz.html.

[10] Gao, J., Skillcorn, D., Capability of SPOT XL data in producing detailed land cover maps at the urban-rural periphery, International Journal of Remote Sensing, 19 (15), pp. 2877-2891, 1998.

[11] Gong, P., & Howarth, P. J., An assessment of some factors influencing multispectral land-cover classification, Photogrammetric Engineering and Remote Sensing, 56 (5), pp. 597-603, 1990.

[12] Gorte, B., Probabilistic Segmentation of Remotely Sensed Images, ITC Publication Series, 63, 1998.

[13] Harris, P. M., Ventura, S. J., 1998. The integration of geographic data with remotely sensed imagery to improve classification on an urban area, Photogrammetric Engineering and Remote Sensing, 61(8), pp. 993-998, 1998.

[14] Schiewe, J., Tufte, L., Ehlers, M., Potential and problems of multi-scale segmentation methods in remote sensing, GeoBIT/GIS, pp. 34-39, 2001.

[15] Townshend, J. R. G., The enhancement of computer classification by logical smoothing. Photogrammetric Engineering and Remote Sensing, 52, pp. 213-221, 1986.

[16] Townshend, J. R. G., Land cover, International Journal of Remote Sensing, 13, pp. 1319-1328, 1992.

[17] Woodcock, C., Strahler, A., The factor of scale in remote sensing, Remote Sensing of Environment, Vol. 21, pp. 311–322, 1987.



Documents

Land cover classification of QuickBird multispectral data ... · PDF fileLand cover classification of QuickBird multispectral data with an object-oriented approach E. Tarantino Polytechnic