Detailed Land Cover Mapping by Introducing Higher-Spatial Resolution Panchromatic Bands in Multispectralclassification: Examples using Landsat ETM+ and Quickbird Imagery

Projo Danoedoro1,2 and Stuart Phinn1
1Centre for Remote Sensing and Spatial Information Science (CRSSIS),
School of Geography, Planning and Architecture,
The University of Queensland, St Lucia QLD 4072 Australia.

2Department of Cartography and Remote Sensing,
Faculty of Geography, Gadjah Mada University Yogyakarta 55281, Indonesia.

Land management and physical planning activities normally require detailed information on land-cover/land-use (LC/LU). The advent of high-spatial resolution imagery is also expected to contribute such information. Although previous works showed that the high-spatial resolution imagery has been successfully used to generate LC/LU information, there were only few categories presented on the derived maps. Regarding this current situation, an image-based detailed LC categorisation was developed in the frame of a versatile LU classification scheme to support local planning tasks in Indonesia. With respect to this detailed categorisation, Semarang area in Indonesia was mapped using multispectral and higher-spatial resolution panchromatic bands of Landsat-7 ETM+ and Quickbird imagery. Multispectral classification procedure was applied to compare the effectiveness of original and the texturally-filtered bands involved. Two approaches in textural transformation were used, i.e. textural filtering of the original multispectral bands and textural aggregation applied to the panchromatic bands. It was found that the multispectral classification using original multispectral bands gave the best overall accuracy levels. In contrast to the previous works' findings, the use of textural filtering and aggregation techniques in integration with original multispectral bands gave lower accuracy levels. Since the detailed LC categorisation was purposively prepared for spectral-related LC labelling, it is not surprising that the textural information -which is related to spatial dimension of the LU- performed less accurately. Nevertheless, this study demonstrates the potential of integrating higher-spatial resolution panchromatic imagery in multispectral classification.


1.1. General Context

Land management and physical planning activities normally require detailed information on land-cover/land-use (LC/LU). Environmental planning at local level, in particular, needs biophysical data that can be extracted from various remote sensing images with relatively large scales. The advent of high-spatial resolution imagery (SPOT HRVIR, IRS-P, Ikonos, and Quickbird) is also expected to contribute such information. In addition, Landsat 7 ETM+ also delivers panchromatic data in 15 m spatial resolution, which is relatively high when its coverage (180 x 185 km) is also taken into account. Those satellite imaging systems are thus very useful to support a spatial database development for local planning.

Although previous works showed that the high-spatial resolution imagery has been successfully used to generate LC/LU information (Aplin et al., 1999; Jenkins and Phinn, 2002; Ehlers et al., 2003), there were only few categories presented on the derived maps. Regarding this current situation, an image-based detailed LC categorisation has been developed in the frame of a versatile LU information system (VLUIS) to support local planning tasks in Indonesia (Danoedoro et al., 2004). In the VLUIS, LC/LU information is broken down into five separate layers, i.e. representing spectral-related cover, spatial, temporal, ecological, and socio-economic function dimensions respectively (Table 1).

The spectral dimension requires automatic spectral classification methods for extracting information. Many studies showed that this task can easily be done using standard multispectral classification (Jensen, 2004; Mather, 2004). Discussion and examples on the use of such methods are normally given in limited number of classes, which is considered to be relevant to particular purposes. It is rarely discussed how such automatic classification methods can be used to generate more detailed categorisation.

The advent of high-spatial resolution imagery such as Ikonos and Quickbird also gives new opportunity for generating more detailed information on LC/LU. Its high spatial resolution enables mapping of features that cannot be recognised using long established satellite systems such as Landsat ETM+ and SPOT HRVIR. However, many applications make use of such high-spatial resolution imagery as a basis for visual interpretation, particularly for urban land assessment. During recent years, automatic classification based on very high-spatial resolution multispectral imagery was also performed by several authors, (Sawaya et al., 20023 Ehlers et al., 2003; Laliberte et al., 2004; Mesev, 2005) to map general LC and (semi)natural features such as vegetation and coral types.

Digital classification of LC/LU sometimes makes use of textural information extracted from the imagery (Zhang, 1999; Campbell, 2002; Mather, 2004). Textural parameters are sometimes used to give additional information, by taking into account the values of neighbouring pixels. Textural filtering techniques are usually applied to bands sensitive to phenomena of interest on LC/LU. The texturally-filtered bands are then added to the original multispectral dataset, to be used as input to multispectral classification process. The works of Danoedoro (2002) and Chen et al. (2004), showed that the use of such bands successfully increases the accuracy of classification results.

On the other hand, some of satellite imaging systems provides higher-spatial imagery in panchromatic spectrum to accompany the multispectral bands. Landsat ETM+ has 15 m spatial resolution in addition to the 30 m multispectral bands, SPOT HRVIR has 5 m panchromatic band besides the 10 m multispectral ones. Ikonos and Quickbird provide 1 m and 0.61 cm panchromatic bands respectively. All these panchromatic bands are usually combined with the multispectral images in pan-sharpened colour composite images, e.g. using Brovey transform (Vrabel, 1996), as a basis for visual interpretation.

Table 1. Versatile Land-use Classification Scheme.
The spectral-related cover categorisation used in this study is given in detail,
while the other dimensions categorisations are only outlined up to level III.
This table is a revised version of the one published in Danoedoro et al. (2004)

On a band-to-band observation basis, the higher spatial-resolution imagery gives textural information in addition to the corresponding lower-spatial resolution bands. Every single pixel of Landsat ETM+ multispectral band may be associated to corresponding 2x2 pixels of its panchromatic band. The same association may be applied to Quickbird imagery, on which every single 2.4 m pixel may be related to 4x4 pixels at 0.6 m panchromatic bands. In spite of this, the potential of using textural information based on the higher-spatial resolution panchromatic band was rarely explored.

1.2. The Study Area
Semarang municipality in Central Java, Indonesia, is an example of a rapidly growing urban area. The city administratively covers an area with contrasting features of landscape, i.e. hilly terrain in the south consists of semi-impervious rocks and flat coastal alluvial plain in the north. The hilly terrain is predominantly occupied by agricultural land, while the coastal alluvial plain is mainly occupied for residential, industrial, and commercial uses. The new residential and industrial areas are expanding fast in the southern part due to rapidly growing population. Accurate and up to date spatial data including LC/LU is needed for monitoring, planning, and managing development in this area. Higher spatial resolution satellite imagery is expected to contribute significantly in solving the problem.

1.3. Problem Formulation
Detailed LC/LU classification and mapping is necessary for various local planning activities. As many local institutions require different types of LC/LU information, the classification scheme should satisfy a range of needs. The currently developed VLUIS offers a detailed categorisation to respond these needs (Danoedoro et al., 2004). The problem is how can the detailed LC/LU mapping be done using moderate and high-spatial resolution imagery? Moreover, what is the contribution of higher-spatial panchromatic imagery to the classification accuracy?

The objective of this study was to evaluate the contribution of Landsat ETM+ and Quickbird panchromatic bands in multispectral-based LC classification. To do so, a spectral-related categorisation under VLUIS was used as a reference. The other LU dimensions are being studied separately, in terms of extraction methods development and their versatility.


3.1. Materials and Data Prepration

This study made use of two image datasets representing moderate- and high-spatial resolution satellite imagery, i.e. Landsat ETM+ with 30 m multispectral and 15 m panchromatic bands, and Quickbird with 2.4 m multispectral and 0.61 m panchromatic bands. The Landsat ETM+ dataset was recorded on 21 August 2002, while the Quickbird dataset was acquired on 31 August 2002. Both image datasets were available at level 1G for Landsat ETM+ (systematic correction in terms of geometric and radiometric preprocessing applied, and the product has been registered to UTM cartographic projection) and standard level for Quickbird. In addition to these datasets, panchromatic B/W aerial photographs of the study area at scale of 1:10,000 and topographic map at scale of 1:25,000 were also available.

A classification scheme for spectral-related cover dimension was prepared (see Table 1), with respect to the work of Danoedoro et al. (2004). This categorisation contains classes with generic names, which are assumed to have direct relationship with their spectral characteristics. The classification scheme is multilevel in character, at which the level I is prepared for imagery with 100 m spatial resolution or coarser, level II is for imagery with 20-100 m spatial resolution, level III for imagery with 5-20 m spatial resolution, and level IV is for imagery with 0.5 - 5 m spatial resolution.

Figure 1. Landsat ETM+ and Quickbird covering the study area.

A geometric correction and coordinate transformation was applied out to the image datasets by using topographic map and GPS receiver-assisted ground control points. After that, principal component analysis (PCA)-based pan-sharpened colour composite datasets were created in order to prepare enhanced colour display with higher-spatial resolution. With this method, two new 'multispectral' datasets were created, i.e. at 15 m for Landsat ETM+ and at 0.60 m for Quickbird.

3.2. Methods
The methods used in this study are summarised in Figure 2. There was initial processing in terms of geometric correction and pan-sharpened colour composite image composition. After that, parallel processing in terms of (a) textural filtering applied to the original multispectral bands of Landsat ETM+ and Quickbird imagery, and (b) aggregation of higher-spatial resolution panchromatic bands into lower-spatial resolution, equal to their corresponding multispectral bands. Multispectral classification followed using the original, texturally filtered, and aggregated panchromatic bands. Based on the classified images, evaluation on the contribution of higher-spatial resolution imagery to the accuracy of classification results was done.

Figure 2. Method used in this study. All geometric pre-processing have been excluded from this figure.

A convolution filtering technique using textural filter was applied to the original bands of both image datasets. A variance filter at a size of 7x7 window kernel was chosen in order to represent the heterogeneity of neighbouring pixels within the window. The kernel size of 7 x 7 was recommended by several authors, e.g. Chen et al. (2004) and Puissant et al. (2005) for optimal texture expression. Based on the obtained results, texturally-filtered red, near infrared and first middle infrared bands of Landsat ETM+ and similarly treated red and near infrared bands of Quickbird were then selected. These three textural bands were then added to the original bands, creating a new dataset with nine and six bands for Landsat ETM+ and Quickbird respectively.

Parallel to the processing with convolution filtering technique, an aggregation technique of higher-spatial resolution panchromatic bands was applied. In contrast to convolution filters, the aggregation method does not work with moving window. Instead, it reduces every block consisting n x n pixels of the higher-spatial resolution panchromatic bands into single pixel at the size of lower-spatial resolution multispectral bands. Thus, every 2 x 2 pixels of 15 m resolution panchromatic band was aggregated into a single 30 m resolution pixel, while every 4 x 4 pixels of 0.60 m resolution panchromatic band was aggregated to a single 2.4 m resolution pixel. A variance parameter of the original neighbouring pixels was used to define new values of the derived lower-spatial resolution images. By doing so, additional 'bands' at 30 m and 2.4 m containing textural information conveyed from the higher-spatial resolution panchromatic band were added to the multispectral datasets of Landsat ETM+ and Quickbird respectively.

A spectral sampling procedure was run using regions of interest (ROIs). Spectral samples were taken from 76 ROIs for Landsat ETM+ and from 134 ROIs for Quickbird. Independent field reference datasets containing 61 ROIs for Landsat ETM+ and 112 ROIs for Quickbird were also prepared for accuracy assessment. The aerial photographs were used as reference, supported by local knowledge of the study area. During this sampling process, an interactive separability index was calculated using Jeffries-Matusita (Jensen, 2004). Thus, with every new ROI taken, a separability index was immediately computed as a basis for decision to include or exclude every ROI as a sample. By doing so, pixel samples can be expected to give a better prediction to the classification result.

Based on the collected samples, a supervised classification was executed using maximum likelihood algorithm (MLH). After that, the immediate result of the MLH was regrouped into smaller number of classes based on the VLUIS classification scheme, i.e. level III for the Landsat ETM+ and level IV for the Quickbird imagery. The regrouped classes were then filtered using global majority and selective majority. By using global majority filtering, all classes were taken into account for determining the predominant class within a given window. In a selective majority filtering, particular classes are excluded from the calculation in order to preserve their existence in the kernel's centre, whatever the predominant class was.

The same procedure and classification process using the same ROIs was applied to the other datasets, i.e. (a) the layer stack containing texturally-filtered multispectral bands, as well as (b) the layer stack with texturally aggregated panchromatic bands. After that, the same class merging and majority filtering processes was applied

Accuracy assessment was then applied using an independent dataset, in terms of separately collected ROIs. A confusion matrix as a basis for overall accuracy computation (Short, 1982) and Kappa index (Congalton, 1991) was also developed to evaluate class error levels.


4.1. Textural Filtering and Aggregation

Textural filtering technique in terms of variance computation of the neighbouring pixels within a 7 x 7 kernel window gave different effects on each spectral band. For Quickbird multispectral imagery in general, the visible bands (blue/1, green/2, and red/3) give a better, enhanced, appearance in urban area as compared to the near infrared band (band 4). Conversely, near infrared band 4 gives more detailed textural information in the agricultural and forested areas. All these phenomena are related to the designated spectral sensitivity of each band. However, red and near infrared bands perform better than others in a region with mixed features of urban and agricultural lands like Semarang. For Landsat ETM+ imagery, bands 3 (red), 4 (near infrared), and 5 (middle infrared I) give better textural results as compared to others. All better performing bands were then added to the original multispectral datasets, to be used as input to the multispectral classification process.

A different approach was used for extracting textural information from the panchromatic band with higher-spatial resolution. Instead of using convolution technique with moving window, an aggregation method was applied, i.e. by converting every 2 x 2 pixel size (Landsat ETM+) and 4 x 4 pixel size (quickbird) blocks of the panchromatic bands to single pixels with new values. The generated new images have identical pixel size with the multispectral bands (30 m for Landsat ETM+ and 2.4 m for Quickbird), while the new values were computed using a textural parameter, i.e. variance.

Figure 3. Effects of textural filtering on middle infrared-I band of Landsat ETM+ (A)
and red band of Quickbird (C). The features look more blocky and noisy as compared
to the results of textural aggregation of
panchromatic bands of Landsat ETM+ (B) Quickbird (D).

Comparison at the same spatial resolution shows that, generally speaking, the texture-based aggregation of the panchromatic band gives better results than the texturally filtered multispectral bands, in the sense that it keeps the original boundary of features, particularly the linear objects such as airport runway, coastal line, and rivers (Figure 3). The textural filtering technique with variance computation of the original bands tends to show very sharp, blocky 'edge-effect' alongside the boundaries between homogeneous features. However, it also tends to create 'noisy' appearance on the heterogeneous phenomena such as mix of residential and industrial, as well as urban and agricultural, areas. Figures 3A and 3C depict the middle infrared-I band of Landsat ETM+ and red band of Quickbird respectively, which have been treated using variance-based textural filter. The use of larger size of filter size, e.g. 13 x 13 instead of 5 x 5 or 7 x 7, may reduce the noisy features. On the other hand, it may also reduce the effective resolution of the image, even though the pixel size remains the same. Figures 3B and #D show the aggregated panchromatic bands of Landsat ETM+ and Quickbird respectively. These bands were aggregated to lower-spatial resolution equal to that of multispectral bands (30 m for Landsat ETM+ and 2.4 m for Quickbird).

4.2. Classification Results
The classification procedures were applied to different datasets, i.e. original bands (OB), original plus texturally filtered bands (OTFB), and original plus texturally aggregated panchromatic bands(OTAPB). These datasets were processed using the same spatial resolution and ROI-based samples. The same further processing in terms of postclassification enhancement was also applied.

The sampling procedure used for ROI selection was mainly based on OB displayed on several colour combinations at once; so there were true colour, standard false colour and other colour composite images on screen. Pan-sharpened colour composite images were also used to visually examine the homogeneity of ROIs, prior to the statistical computation. During the sampling process, it was found that different types of surface could be discriminated based on their separability indices, as far as their size is large enough as compared to the image's spatial resolution. The use of true colour in addition to other colour composites was also found useful in determining the colours and types of object on the Quickbird imagery, e.g. red-painted concrete roof tiles, blue-painted metal roof tile, new and completely rusted metal roof tiles. Using Landsat ETM+ imagery, discrimination was also made possible for more general categorisation, e.g. clay and fibre-cement roof tiles as well as asphalt surface.

Discrimination of dry-wet soil surfaces, broadleaf - needleleaf vegetation types, and vegetation densities could easily be done using Landsat ETM+ imagery. This image was also found useful in discriminating water turbidity level. General soil types with different origin or characeristics (e.g. sandy volcanic soils, light-toned sandy calcareous soils, and loamy-clay soils with high iron oxide content) can be discriminated as well. However, discriminating vegetation densities using Quickbird imagery was found to be more difficult. This may be due to 'density' being spatial resolution-sensitive. At a moderate-spatial resolution, like Landsat ETM+'s, canopy density of several stands may be expressed by a single pixel, but it may not the case in spatial resolutions finer than 3 metres where a single stand may be represented by more than one pixel.

Based on the spectral samples, the maximum likelihood classification algorithm delivered 40 and 85 cover classes for Landsat ETM+ and Quickbird respectively. By using VLUIS's spectral-related cover dimension classification scheme, those classes were regrouped into 27 (Landsat ETM+) and 48 (Quickbird) categories. This regrouping procedure applied to all datasets containing OB, OTFB, and OTAPB.

Detailed classification results show that the use of OB as input provides the best accuracy level as compared to that of OTFB and OTAPB. As presented on Table 2, the immediate results of multispectral classification using OB gives the highest overall accuracy as well as Kappa index for both Landsat ETM+ and Quickbird datasets. When the texturally filtered bands were introduced to accompany the OB, lower accuracy levels were obtained. Moreover, the use of texturally aggregated panchromatic bands in combination with the OB also generates results with lower accuracy levels for both datasets.

However, there were different trends in the effects of class merging. The class merging is necessary to group similar categories into meaningful classes as outlined in the VLUIS' spectral-related cover dimension classification scheme. When the 40 classes of Landsat ETM+ were merged into 27 new classes, consistent increases of accuracy levels in all datasets were performed, with the OB dataset's class merging shows the highest, while the combination of OTAPB shows the lowest accuracy levels. Class merging from 95 to 48 categories of the Quckbird datasets performed inconsistent changes in accuracy level. Despite the highest accuracy level of the OB, the OTFB shows the lowest accuracy level as compared to the other two.

Table 2. Accuracy assessment1 of the classification results based on original bands (OB),
original plus texturally filtered bands (OTFB),
and original plus texturally aggregated panchromatic bands (OTAPB).

Note :
  1. All percentages indicate overall accuracies, while the Kappa indices are presented in brackets.
  2. Immediate results: 40 classes for Landsat ETM+ and 85 classes for Quickbird
  3. After class merging results: 27 classes for Landsat ETM+ and 48 classes for Quickbird
  4. Further class merging was applied toQuickbird datasets referring to the same classification scheme used for Landsat ETM+ dataset. Smaller number of classes was found on Quickbird datasets due to the difference in area coverage as compared to that of Landsat ETM+

The findings suggest that the introduction of texturally aggregated information to the detailed classification process is more effective on image dataset containing much higher spatial resolution panchromatic band like Quickbird. With Landsat ETM+, aggregation of higher-spatial resolution (i.e. panchromatic) band tends to lower the accuracy level as compared to texturally-filtered bands. The difference in spatial-resolution ratio between Landsat ETM+ and Quickbird may be used to explain these findings. Landsat ETM+ imagery has relatively low spatial-resolution panchromatic band (15 m) as compared to its counterpart multispectral bands (30 m). This means that information on every single pixel of any multispectral band gains textural information from the statistics of only 4 pixels of corresponding panchromatic band. On the other hand, 16 pixels of panchromatic band more effectively contribute textural information of every single pixel on Quickbird multispectral band.

The negative contribution of textural information in lowering the classification accuracy (as compared to the use of OB) is probably related to the classification scheme used in this study, although this needs a further research to prove.. Previous works (e.g. Danoedoro, 2002; Chen et al., 2004) showed that the introduction of textural information could improve the classification accuracy. It should be noted, however, that the classification schemes used in those studies do not purposively developed for generic objects related to LC as assumed in the VLUIS classification scheme used here. Danoedoro (2002), for example, used the texturally filtered bands for extracting information on LC containing spatial and socio-economic function aspects, such as (a) urban settlement, (b) mix of urban settlement and commercial areas, (c) typical industrial area, (d)dry fields with low percentage of crop cover, (e) urban settlement with sparse vegetation. Chen et al. (2004) used modified USGS

classification scheme containing information like (a) single family residential area, (b) multifamily residential area, (c) commercial/industrial area, (d) artificial (irrigated grassland), and (e) high density vegetation. In the VLUIS classification schemes, those categories are specified under various dimensions other than spectral-related cover types (see Table 1).

Because the classification scheme of the VLUIS' cover dimension was spectrally-based and purposively developed for multispectral extraction method, it is not surprising that the use of OB generates the highest accuracy level. On the other hand, density is a spatial resolution-sensitive concept, and thus related to spatial dimension. Several classes containing density information on Quickbird imagery were more accurately classified using OTFB and OTAPB, e.g. C2211/low-density woody broadleaves and C2221/medium density woody broadleaves. Texturally, those classes contain high variance values as compared to high density woody broadleaves. On the other hand, categories with low contrast between pixels (which means having smoother texture) show lower accuracy levels. C12 (shallow water) and C321 (mud) are the examples. With landsat ETM+ datasets, there was non clear difference in accuracy lvels between OB and OTFB, even though in general OTAPB performs less. Table 3 summarises the findings.

Another thing important to note is that the use of 7 x 7 kernel size in this study was based on the assumption that it may work properly in Indonesian urban and rural environment. The works of Chen et al. (2004) and Puissant et al. (2005) recommending the use of that textural filter kernel size may need an evaluation, particularly for urban and rural environment in developing countries like Indonesia.

Table 3. Example of classes containing different levels of textural variance, and their contribution to the producer's and user's accuracies. All accuracy levels are presented in percent.

OB: original bands dataset. OTFB: Original plus texturally-filtered bands dataset. OTAPB: original plus texturally-aggregated panchromatic band

The table 3 showed that the introduction of textural information is generally not suitable for automatic classification using VLUIS' spectral-related cover dimension. However, textural information extracted from the higher-spatial resolution such as that of panchromatic bands performs better, particularly when they are texturally aggregated using relatively large window size into lower spatial resolution. The use of textural information extracted from the higher-spatial resolution panchromatic band may be considered as an alternative to replace the texturally-filtered bands.

Figure 4. Detailed land-cover classification result based on original bands datasets

Based on those research findings, it can be concluded that:
  1. Multispectral bands of Landsat ETM+ imagery gives higher accuracy for detailed LC classification as compared to much finer spatial-resolusion imagery like Quickbird. This finding is in accordance with most previous works' findings;
  2. Original datasets in terms of multispectral bands are supportive to detailed LC mapping using classification scheme of spectral-related cover dimension. On the other hand, textural information built using filtering technique and aggregation of higher-spatial resolution panchromatic band performs less accurately. Since texture is closer to spatial than spectral dimensions, it does not work properly in discriminating objects using spectral criteria;
  3. Textural aggregation of higher-spatial resolution panchromatic band performs better than textural filtering, particularly when a contrast combination of lower- and higher-spatial resolution is in use. Quickbird imagery with 1:4 spatial ratio performs better than that of Landsat ETM+ with 1:2.

Since this study was carried out with only a single size of textural filtering window, i.e. 7 x 7, it is important to further study the use of various sizes of window in order to check their effectiveness in improving classification accuracy. This study also recommend a further study on the use of textural information, in terms of both filtering and aggregation, for extracting spatial aspects of LC. The further study should also carried out in relation with the progress of the currently developed object-based image segmentation methods.

  • Aplin, P., Atkinson, P.M., and Curran, P.J. (1999). Fine Spatial Resolution Simulated Satellite Sensor Imagery for Land Cover Mapping in the United Kingdom. Remote Sensing of Environment, 68, 206-216
  • Campbell, J. B. (2002). Introduction to Remote Sensing (3 ed.). New York: Guilford Press.
  • Chen, D., Stow, D.a., & Gong, P. (2004) Examining the Effect of Spatial Resolution and Texture Window Size on Classification Accuracy: An Urban Environment Case. International Journal of Remote Sensing, Vol. 25, No.11, 2177-2192
  • Cihlar, J., & Jansen, L. J. M. (2003). From Land Cover to Land Use: A Methodology for Efficient Land Use Mapping over Large Areas. Professional Geographer, 53(2), 275-289.
  • Danoedoro, P. (2002). Integrating Spectral, Textural and Terrain Information for Land-use Mapping of wet Tropical Region: a Case Study of the Semarang-Ungaran Area, Central Java, Indonesia. Project assignment report on Advance Remote Sensing of Environment. Department of Geographical Sciences and Planning the University of Queensland.
  • Danoedoro, P., McDonald, G.T., & Phinn, S.R. (2004). Developing a Versatile Land-use Information System Based on Satellite Imagery for Local Planning in Indonesia Phase I: Establishment of Classification Scheme. Paper presented at the 7th International Seminar on GIS for Developing Countries (GISDECO-7) "GIS Capacity Building and Infrastructure", Universiti Teknologi Malaysia, Skudai, Johor, 10-12 May 2004
  • Ehlers, M., Gahler, M., & Janowsky, R. (2003). Automated Analysis of Ultrahigh Resolution Remote Sensing Data for Biotope Type Mapping: New Possibilities and Challenges. ISPRS Journal of Photogrammetry and Remote Sensing, 1252, 1-12.
  • Jenkins, R. B., & Phinn, S. R. (2002). Mapping the Internal Structure of Urban Areas by Spatial Reclassification of High Resolution Imagery (Poster). Paper presented at the The 11th Australasian Remote Sensing and Photogrammetry Association Conference -- Image to Information, Brisbane.
  • Jensen, J. R. (2004). Introductory Digital Image Processing -- A Remote Sensing Perspective (3 ed.). Englewood Cliffs, N.J.: Prentice Hall.
  • Laliberte, A.S. Rango, A.Havstad, K.M., Paris, J.F., & Beck, R.F. (2004). Object-oriented Image Analysis for Mapping Shrub Enchroachment from 1937 to 2003 in Southern New Mexico. Remote Sensing of Environment No.93, 198-210
  • Lillesand, T. M., Kiefer, R. W., Chipman, J. (2004). Remote Sensing and Image Interpretation (5 ed.). New York: John Wiley and Sons.
  • Mather, P. M. (2004). Computer Processing of Remotely Sensed Data: An Introduction (2 ed.). Brisbane: John Wiley and Sons.
  • Mesev, V. (2005). Identification and Characterisation of Urban Building Patterns using IKONOS Imagery and Point-based Postal Data.. Computers, Environment and Urban Systems. (article in press)
  • Puissant, A., Hirsch, J., & Weber, C. (2005) The Utility of Texture Analysis to Improve Per-pixel Classification for High to Very-high Spatial Resolution Imagery. International Journal of Remote Sensing, Vol. 26, No.4, 733-745
  • Sawaya, K.E., Olmanson, L.G., Heinert, N.J., Brezonick, P.L. & Bauer, M.E. (2003). Extending satellite Remote Sensing to Local Scales: Land and Water Resource Monitoring using High-resolution Imagery. Remote Sensing of Environment No.88, 144-156
  • Vrabel, Jim, (1996), Multispectral Imagery Band Sharpening Study, Photogrammetric Engineering & Remote Sensing, Vol. 62, No. 9, pp. 1075-1083.
  • Zhang, Y. (1999). Optimisation of Building Detection in Satellite Image by Combining Multispectral Classification and Texture Filtering. ISPRS Journal of Photogrammetry and Remote Sensing, 54, 50-60.
Page 1 of 1