COMBINING IMAGE SEGMENTATION AND MULTISPECTRAL CLASSIFICATION FOR GENERATING LAND-USE INFORMATION

This research developed methods for generating land-use information which is relevant to a broader study, i.e. land capability assessment and classification for sustainable development of brackishwater aquaculture systems in Indonesia. The broader study requires land-use information covering coastal areas with various utilisations, e.g. coastal fishponds, rice fields, rural settlement, mangrove-based conservation, and urban uses. In order to meet that requirement, the land-use information needs to be delivered in terms of spatial, ecological, and socio-economic dimensions. To do so, a versatile land-use information system (VLUIS) which has been developed for local planning in Indonesia was used as a reference. In the VLUIS, the land-use information is broken down into five layers representing spectral, spatial, temporal, ecological, and socio-economic dimensions. As the study area, two small portions of Landsat ETM+ image covering Maros and Jeneponto, South Sulawesi, Indonesia were chosen. In this study, a combination of multispectral classification and object-based image segmentation was applied. The multispectral classification was carried out to generate spectral-related land-cover types, while the object-oriented image segmentation was run to derive spatial dimension of land-use. A terrain unit map obtained from visual interpretation was used to support the integration of the spectral-related land-cover and the spatial dimension maps. A knowledge-based classification incorporating spectral, spatial and terrain characteristics of the study area was carried out. By this method, new spatial information in terms of maps representing ecological and socio-economic dimension of land-use were generated using different rules. This study showed that a single source of imagery could be processed in various ways to derive different types of spatial information, and all information could then be integrated to generate versatile land-use information relevant to particular planning tasks. The multispectral classification was found to be accurate enough to provide spectral-related land-cover types. It was also found, however, that the object-based image segmentation was still less accurate to classify objects with respect to their shape, size, and pattern simultaneously, particularly in comparison with the visual interpretation. Nevertheless, in the near future, improved methods of this approach may be expected to provide more useful and accurate information.