Extracting Land-Use Information Related to Socio-Economic Function From Quickbird Imagery: A Case Study of Semarang Area, Indonesia
Department of Cartography and Remote Sensing,
Faculty of Geography Gadjah Mada University, Yogyakarta Indonesia /
Centre for Remote Sensing and Spatial Information Science (CRSSIS)
The University of Queensland, StLucia Australia
Recently, high-spatial resolution imagery has widely been used for environmental assessment and mapping in Indonesia. However, most of these studies made use of visual interpretation instead of digital classification for generating land-cover/land-use information from such imagery. On the other hand, digital classification was usually applied to deliver land-cover information or mixed information between land-cover and land-use with relatively general categorisation, so that the results were not adequate to support planning. This study tried to extract land-use information related to socio-economic function by combining spectral classification, image segmentation and visual interpretation of Quickbird imagery covering Semarang area, Indonesia. To do so, a multi-spectral classification was run to derive detailed spectral-related land-cover classes. Image segmentation and visual interpretation were also carried out to generate spatial pattern of the land-cover features. A classification scheme under the versatile land-use classification system (VLUIS) was used as a reference. Integration of the spectral-related land-cover and spatial pattern maps was controlled using a knowledge-based approach, by formalising knowledge about spatial relationship between land-cover, socio-economic function, spatial pattern and their ecological context into a set of GIS rules. The result showed that Quickbird imagery could be used for generating socio-economic function of land-use at 83.63% accuracy (Kappa=0.821). In addition, several limitations related to the methods used and inaccurately mapped categories were identified.
1. INTRODUCTION: BACKGROUND AND PROBLEM
Land-cover/land-use information is recognised as an important input to planning (Lindgren, 1985; Lein, 2003). However, Fresco (1994) claimed that accurate data on actual land-use were not easily found at both global/continental and national/regional scales. In many developing countries, problems related to the availability of land-use information were caused by the lack of coordination between technical institution and the rarely available spatial data in terms of compatibility, relevance and newness. These lead to the need for the development of a flexible, multipurpose classification system (Danoedoro, 2004). From remote sensing point of view, the problems were caused by difficulties in generating land-use information from automatic classification procedures, since the automatic classification could only derive land-cover classes.
Land-use and land-cover are different. Clear differentiation between land-cover and land-use concepts has been made by several authors. Campbell (1983), for example, showed the difference in concrete-abstract dichotomy, where land-cover is concrete and land-use is abstract. That is, land-cover can be mapped directly from images, while land-use requires land-cover and additional information on how the land is used. However, both concepts were sometimes mixed in use (Anderson et al., 1976; Malingreau and Christiani, 1981), although land-use information normally contains attributes of land-cover. As a result, many land-use maps derived from remotely-sensed imagery could not consistently present land-use information, and these could not be effectively used in various applications related to planning. Due to its abstract character, the land-use information extraction is still a challenging task in remote sensing studies.
2. PREVIOUS WORKS
Visual image interpretation could normally generate land-use information by combining a set of interpretation elements including colour/tone, texture, shape, shadow, size, pattern, site and association (Lillesand et al., 2004). With digital multspectral classification, only land-cover could usually be extracted, as the land-cover types are related to their spectral responses recorded by the remote sensors (Jensen, 2004; Mather, 2004). The recorded spectral responses are comparable to tones or colours in visual interpretation. Further development in image analyses provided textural approaches based on statistical measure of a pixel group within a particular size of a moving window (Chen et al., 2004; Danoedoro, 2005). Current development of image processing system showed that shape and pattern could be recognised and classified using object-based image segmentation approach (Baatz and Schappe 2000; Ranasinghe, 2006).
However, land-use could not be simply interpreted from the colour, texture and shape of pixel groups. There are other factors that should be incorporated such as association and location of the object. If the land-use is viewed from its socio-economic function of the land cover, there would be more efforts required for deriving such information from a digital image dataset. Image-based LC/LU mapping, according to van Gils et al. (1990) could be carried out using photo-guided, photo-key, and land-ecological approaches. The photo key approach uses aerial photographs or printed imagery as a guide during field observation and measurement, while the photo key approach relies heavily on the photomorphic features as basis for objects recognition. Land-ecological approach, on the other hand, views land-cover, land-use and land characteristics as a unity. Therefore, interpretation of more detailed information related to LC and LU could be done by taking land characteristics into consideration.
Classification schemes also played an important role in the success of image classification. Danoedoro (2006) showed that the use of a classification scheme with classes that could not be directly related to the spectral classification might lead to a lower accuracy result. Similarly, classification scheme with classes relevant to spectral differentiation could lead to less accurate result when it was used for classifying imagery using non-spectral approach. In addition, there was another problem of relating classification schemes and remote sensing methods. Most digital classification methods were used for land-cover/land-use with limited number of classes, i.e. equal or less than 10 (Aplin and Atkinson, 2000; Sawaya et al., 2003; Wang et al., 2004; Puissant et al., 2005), while many applications related to planning require more detailed categorisation. Detailed categorisations used in remote sensing projects were performed by some authors, e.g. Loveland and Belward (1997), which dealt with global land cover mapping. Detailed FAO land-cover classification was prepared for visual interpretation (Jansen and Gregorio, 2003). Detailed categorisation was also found in the USGS land-cover/land-use classification system (Anderson et al., 1976). But in general detailed classification of land-use based on digital processing of remotely sensed imagery was rarely available. The previous works showed that developing method for land-use information extraction using digital remotely sensed data is important. Land-use information should contain information related to ‘uses’, or socio-economic function, rather than land-cover types. In addition, the use of appropriate classification scheme is also critical to the success of such efforts.
3. OBJECTIVE AND GOALS
This study tried to generate information related to socio-economic aspect of land-use based on a Quickbird image dataset covering Semarang area, Central Java, Indonesia. The derived information related to the socio-economic function referred to a newly developed classification system called Versatile Land-use Information System (VLUIS, Danoedoro et al., 2004). In the VLUIS, land-use information is broken down into five dimensions mappable from remotely sensed imagery, i.e. spectral-related land-cover, spatial, temporal, ecological, and socio-econonomic function. The goals of this study were (a) to produce a map showing the socio-economic dimension of land-use of the study area, generated from the Quickbird imagery; and (b) to carry out accuracy assessment of the derived map.
4. STUDY AREA
Semarang Municipality is the capital of Central Java Province, Indonesia. In terms of spatial data availability, the Semarang local authority for planning (Bappeda) does not have adequate data to support spatial planning in its area. The Bappeda only has a Key Dataset for Local Development (KDLD) with relatively old maps, including LU map published in 1995. Moreover, all spatial data are stored in paper format and there is no GIS system for storing and modelling the spatial data for planning purposes. Semarang is inhabited by approximately 1.33 million (2001) with most people concentrated on the north coastal alluvial plain of Java Sea. Stretching from 6º 50’ - 7º 10’ S and 109? 35’ – 110?50’E, Semarang municipality is unique among other provincial capital cities in Indonesia. In terms of area, it ranks fourth (373.7 km2) among other Indonesian cities, and it administratively covers both urban and rural areas at roughly the same proportion. The rural land-uses situated in the rolling and hilly terrain in the southern part are currently decreasing very rapidly due to the rapid urban area expansion. Young fluvial deposits in the coastal area currently experience land subsidence, causing rob – a local name for flooding process related to sea tidal activities. The rob has periodically inundated the urban area intermittently during the past 15 years, provoking inhabitants to move to the higher place with rough terrain in the southern part of the city. Figure 1 shows the study area covering the western part of the Semarang Municipality.
Figure 1. The study area as shown by a colour composite of Quickbird multispectral imagery
5. MATERIALS AND METHODS
This study used Quickbird multispectral and panchromatic datasets at 2.4 m and 0.6 m spatial resolutions respectively, covering western part of Semarang city, Central Java, Indonesia. The datasets were acquired on 31 August 2002 and delivered at standard correction mode, i.e. only the systematic geometric and radiometric errors were corrected. Research Systems Inc ENVI 4.0 software was used for creating a pan-sharpened multi-spectral dataset and multi-spectral classification. ITC’s ILWIS version 3.3 raster-based GIS and image processing software was utilised for on-screen digitisation of terrain units and spatial-related land-cover features. In addition, a trial version of eCognition version 4.0 software was also used for object-based image segmentation.
As summarised in Figure 2, a multispectral classification based on the Quickbird’s multispectral bands was carried out using maximum likelihood algorithm. The VLUIS spectral-related cover classification scheme (Table 1) was used as a reference, although the regions of interest (ROIs) of sample were taken and labelled using generic names. At this stage, 85 spectral-related land-cover classes were generated. After that, a class merging procedure followed, generating 48 land-cover classes. A selective majority filtering was applied to the merged classes in order to remove isolated classified single pixels and to aggregate the speckled classified pixels. The majority-filtered classification result was then used as a land-cover map, ready for further processing, i.e. deriving socio-economic aspect of land-use.
Figure 2. Method used in this study
Figure 3 Summarised criteria for the unsupervised object-based image segmentation
After obtaining all required input maps, a knowledge-based image classification for generating land-use information related to socio-economic function was run. Classes for the socio-economic function referred to the VLUIS socio-economic dimension’s classification scheme (Table 4). In this process, relationship between land characteristics, spatial pattern, and the land-cover types were formalised into a set of rules, applicable in a raster-based GIS environment. The accuracy of the derived land-use map was then assessed using overall accuracy and Kappa.
6. RESULTS AND DISCUSSION
Multi-spectral classification using maximum likelihood algorithm showed that the Quickbird dataset could generate 85 spectral-related land-cover classes at 67.75% overall accuracy (Kappa = 0.6813%). This accuracy result was considered low, compared to the accuracy threshold for land-cover/land-use map (Campbell, 1983). After the classes were merged to 48, the overall accuracy increased to 79.02% (Kappa=0.7829). This resultant map actually showed a land-cover map presented using VLUIS spectral-related cover dimension. When a global majority filtering was applied to the 48 classess, the overall accuracy became 87.05% (Kappa = 0.8656). Furthermore, when a selective majority filtering process applied to the same number of classes, an overall accuracy of 85.90% (Kappa = 0.8539) was achieved. Although the selective majority filtered produced a slightly lower accuracy level, the appearance of the resultant map was better, in the sense that the linear features such as road network and water courses could be preserved. This map was then used a final land-cover map, utilised in the subsequent processing for generating a final land-use map containing socio-economic function information. Figures 4 and 5 show the accuracy assessment result and the land-cover map obtained from this stage.
Figure 4. Accuracy assessment result obtained from the multi-spectral classification using 85 and 48 classes, and 48 classes that were processed using global majority and selective majority filtering.
Figure 5. The spectral-related cover dimension map obtained from the multi-spectral classification of Quickbird dataset covering the study area of north-western Semarang.
Terrain classification and delineation was carried out using on-screen digitisation. The backdrop image used for this purposes was a pan-sharpened colour composite imagery. Based on the terrain classification using visual interpretation, a general terrain unit map relevant to the variability of land-use in the study area was produced. The land characteristics representing each terrain mapping unit is presented in Table 2. They were interpreted from the imagery, then were checked and corrected during the fieldwork. The land characteristics, spatial dimension map, and the land-cover maps were then correlated to the socio-economic function-related land-use classes, in order to develop a set of computer-based rules for deriving the land-use map in a raster-based GIS environment.
A spatial dimension map showing land-cover-related spatial units characterised by shape, location, regularity, pattern, and density of the land-cover was generated using on-screen digitisation. This map was interpreted from the pan-sharpened Quickbird magery, so that detailed features such as road network with 4-5 m wide could be delivered. Mapping units of such spatial dimension map served as objects that were used to transform spectral-related land cover classes to other categories, e.g. socio-economic function-related land-use classes. Figure 6 shows the spatial dimension map generated from the on-screen digitisation of the pan-sharpened Quickbird imagery.
Figure 6. The spatial dimension map obtained from the on-screen digitisation.
The immediate result of the object-based image segmentation was relabelled using the VLUIS spatial dimension classification scheme (Figure 7b). In order to assess its accuracy, this spatial dimension map was superimposed with the other spatial dimension map obtained from the visual interpretation (Figure 7d). By overlaying both maps, an error matrix could be delivered, resulting an overall accuracy of 62.27 % (Kappa = 0.5978). This accuracy level showed that the object-based image segmentation could not be used for generating spatial dimension map according to the VLUIS classification scheme. This result also lead to the use of the spatial dimension map derived using on-screen digitisation in the subsequent process, i.e. deriving socio-economic dimension of land-use information.
Figure 7 (a) small portion of the north-western part of the study area used for testing the segmentation result; (b) segmentation result; (c) skeleton of the segmentation result superimposed on the colour composite image; (d) skeleton of the segmentation result superimposed on the visual interpretation-based spatial dimension map
By integrating the spectral-related cover dimension, terrain unit, and spatial dimension maps, the socio-economic dimension map was derived. In this process, the matrix showing relationship between terrain units (representing land characteristics) and the socio-economic classes could be used as a basis for making decision, whether any spectral-related land-cover class exist on the particular terrain unit could be judged as a particular socio-economic class, or it should be consulted to the corresponding spatial dimension category. For example, typical shallow water (C12) in the tidal flat could (Mtf2) be judged as coastal fishpond (F1222). Old clay roof tile (C4212) in the strongly dissected piedmont (Dpmx) was labelled as lower class residential area because (F4111), according to the spatial dimension map, it was built up area with irregularly spaced small sized units interleaved by vegetation (S4111). The result was a socio-economic dimension map presented in Figure 8.
Figure 8. The land-use map containing socio-economic function information derived from the spectral-related land-cover, spatial dimension, and terrain unit maps.
Based on the matrix, a strategy could be set up, involving the use of two-dimensional tables for controlling overlay between the terrain unit, spatial dimension, and spectral-related land-cover map. The terrain unit and the spatial dimension maps were superimposed first to generate a map representing a combination of terrain and spatial aspect of land-cover. After that, the result was superimposed with the spectral-related land-cover map to derive the socio-economic dimension map. This means that the socio-economic labelling was determined by the land characteristics, land-cover types, and the spatial pattern, regularity, density, size, and pattern of the land-cover.
The socio-economic function was specified in four main categories: (a) water-based utilisation, (b) forest-based utilisation, (c) agricultural uses, and (d) settlement and infrastructures. Further categorisation details were derived with respect to this major grouping. For example, under agricultural uses, the following categories were mapped: (a) F3221/rubber estate, (b) F3412/rice field with continuous (three times) rice, and (F3523) dry field with mixed cash crops. In addition, under the settlement and infrastructures, the following categories were mapped: (a) F4111/lower class residential area, (b) F4112/ medium-higher class residential complex, (c) F4312 parking area, (d) F4341/waste recycle area, and (e) F4342/waste disposal land.
Accuracy assessment of the northern part of the Quickbird imagery showed that overall accuracy of 83.63% was achieved, with Kappa index of 0.8209. Confusions were constituted by continuous rice field with three times rice (F3411) and continuous rice field with twice rice (F3412); homestead garden (F3610) and forest garden (F3640); as well as forest garden (F3640) and mix garden (F3620). In comparison with the spaial dimension map, the socio-economic dimension map was less pixelated, because the objects, not individual pixels, played a more important role during rule-based classification. On the other hand, the features representing the socio-economic function classes looked more blocky then those of the spectral-related land-cover map, because the grainy features have been aggregated with the aid of the spatial dimension map.
Most of the mappable classes derived from the Quickbird datasets were presented at level 4. At this level, maps for supporting local planning activities are normally prepared at scale of 1:5,000 - 1:10,000. However, it was also found that several categories (particularly annual crops), which have been defined at level 4, were not possible to extract using Quickbird imagery. Crop types (spectral-related cover dimension) were the example. Due to its wide spectral range of each band and the abundance of crop types grown in the study area, the multi-spectral bands utilised by the Quickbird sensor was not adequate to differentiate various crop types grown at once in one place. However, by correlating the general categories of crop (e.g. cash crops, rice), more detailed categories at level 4 under other dimensions (particularly ecological and socio-economic dimensions) could be derived. Crop rotational pattern, for example, is strongly controlled by the terrain characteristics. By correlating the crop calendar, date/season of recording, general group of the crop types that could be recognised, and the terrain characteristics, the socio-economic dimension of land-use in terms of rotational pattern and possible planted crops could be generated.
Socio-economic function reflects how people use the land for satisfying some of their needs. The classification scheme under the dimension of socio-economic function showed a consistent categorisation in terms of “uses”. In addition, this study also showed that integration of digital image processing and raster GIS could generate relatively detailed land-use information. Some categories also required additional information with administrative boundaries, which could only be supplied by secondary data, e.g. land status. Some other categories could easily be defined using spectral approach supported by limited ancillary data, for examples teak (Tectona grandis) production forest areas, since virtually all areas with teak trees were intended for production forest.
It should also be realised that land-use is a multifaceted concept. It may not be modelled and mapped merely using five dimensions based on remotely sensed imagery. In more developed areas, land-uses are normally less correlated to the external forms of the features, e.g. building shape and pattern. The use of remote sensing technology to model the land-use into multidimensional fashion should be put in the context of complementary or alternative approach, where field surveys often fail to generate comprehensive, efficient and rapidly provided information required in a planning process. Furthermore, all these dimensions should be used in combination with field-based land-use survey and measurement, particularly for urban land-use planning purposes.
Remote sensing normally produces accurate results under certain circumstance in relation to images and processing. Indonesia is situated in a wet tropical region so that cloud cover and hazy atmosphere impede regular acquisition of good quality imagery at any day throughout a year. Increasing number of satellite systems that are capable to make recording with various viewing angles is expected to overcome this problem.
From the spatial analytical perspective, the use of simple map overlays with Boolean logic operators (If-then-else) for deriving other dimensions’ categories needs to be improved in further studies. The ecological and socio-economic function dimensions, for example, may be modelled using more sophisticated spatial analyses tools, e.g. neighbourhood analysis. The presence of particular socio-economic functions is occasionally related to their neighbours, so that analysis of adjacency or contiguity in vector data model needs to be explored.
This study has shown that the land-use information related to socio-economic function could be generated from high-spatial resolution imagery like Quickbird, at relatively high overall accuracy level. Such information could be delivered when the spectral and spatial dimensions of land-cover/land-use were integrated with terrain unit map containing land characteristics relevant to the variation of socio-economic aspect of land-use. To achieve that result, a multi-spectral classification should be combined with visual interpretation or image segmentation generating spatial dimension of the land-cover/land-use as well as terrain unit classes.
This study also showed that the use of object-oriented image segmentation was not successful yet, in the sense that the accuracy level achieved was relatively low. However, in the near future, improvement in the methods and algorithms may be carried out, so that a better result may be expected.
|© GISdevelopment.net. All rights reserved.|