Stanisław Lewiński - Monographic Series No 12
Price: 45 PLN
Object-oriented classification of satellite images as a method for deriving information object on land cover and land use
Increase of application of satellite images for environmental monitoring can be observed recently. This is particularly evident while analyzing tasks undertaken within European GMES initiative, which assumes collection of information on land cover and land use on the basis of satellite images more frequently and with higher level of spatial detail. In order to fulfill these tasks new methods of creating spatial databases should be commonly implemented, with the aim to replace conventional method of visual interpretation.
The presented work includes proposal of methodology for creating land cover/land use database on the basis of object-oriented classification of medium-resolution satellite images. The proposed methods enable to create vector database automatically, with defined minimum mapping unit and with class boundaries similar to those obtained with the use of conventional methods of visual interpretation.
Object-oriented classification of Landsat ETM+ satellite image was done for the study area of 423 km2, comprising city of Legionowo. The results of classification were next processed with the use of specially designed algorithm. The whole approach consists of 8 stages: preparation of satellite data for classification, segmentation, object-oriented classification, generalization of classification results, preparation of vector database, smoothing of class boundaries, verification of classification, preparation of final database.
In the phase of preparation of Landsat ETM+ satellite image for classification multispectral and panchromatic data were fused, applying PANSHARP algorithm. Segmentation of image content was done on the basis of panchromatic channel and data obtained after fusing. In the course of object-oriented classification various classification methods were applied, including self-prepared ZABUD1 criterion, which was used for discriminating built-up land classes. Generalization of classification was done, using minimum mapping unit of 1 ha for built-up land and 4 ha for the remaining classes. It caused changes reaching only 1.1 % of the study area. Next raster clasified image was converted to vector one, forming vektor data base, and smoothing of class boundaries was performed to make them resembling map obtained as a result of visual interpretation. Vector database was next verified by visual interpretation; in the course of this process interpreter could change class codes and location of class boundaries. In the last stage of the work final database was prepared, which included changes entered in verification process. As a result of the applied object-oriented classification and further processing of classification image the database, comprising 18 land cover/land use classes was created; its accuracy was assessed as reaching 94 %.
In order to make comparison of the methods, land cover/land use classification was done, using conventional classification approach. So-called hybrid method was used, which combines supervised classification and next unsupervised approach for non-classified pixels. Pixel-based classification allowed to recognize only 8 classes. The resultant image was generalized using the same rules as in case of object-oriented classification; this process caused changes in classification image reaching over 26 % of the study area. Next results of classification were compared with database obtained in the course of object-oriented classification, which served as a reference material. Low consistency of both classifications was achieved – 72 % and 61 %, depending on the applied method of comparison.
In order to confirm the applied assumptions and approach for objectoriented classification, the action was done to prepare land cover/land use database for a new region. Study area covering sheet of topographic map at a scale of 1:50000, called “Pulawy” (M-34-20-D sheet designation), was selected for this purpose. Landsat ETM+ image, used previously at IGiK for preparing CORINE CLC2000 database, was utilized for this work. The sate llite data were not original, but already pre-processed within CORINE project.
This pre-processing included image rectification, in which cubic convolution resampling method was applied, as well as change of pixel size was done (for multispectral data from 30 m to 25 m, for panchromatic data from 15 m to 12.5 m). Such an approach was aimed at obtaining colour composites with higher level of detail, to be used later for visual interpretation. Application of object-oriented classification enabled to recognize 15 land cover/land use classes. Automatically prepared vector database was compared with CORINE CLC2000 database. Despite big difference in spatial resolution of two compared databases (4 ha and 25 ha, respectively) high consistency of areas for particular classes was achieved.
The proposed approach, aimed at automatic preparation of land cover/land use databases, can significantly support, or even replace visual interpretation. It allows to prepare database with similar information content as in case of visual interpretation in a shorter time, giving much more detail.
The presented method of preparing results of object-oriented classific ation is applied according to the strictly defined rules; it assures repeatability of the results.