Wyniki wyszukiwania - BazTech

1

Hybrid approach for mapping wetland habitats based on application of VHR satellite images

Bochenek Z., Shrestha S., Malek I.

Geoinformation Issues

|

2013

|

Vol. 5, no. 1(5)

21--28

EN

The article presents results of work aimed at developing methods for classifying wetland habitats, based on different types of satellite data and on various classification approaches. Very high resolution WorldView-2 satellite images were used for the research work as basic input data. An object-oriented, rule-based approach was applied to achieve high accuracy of classification of wetland vegetation classes. As a result of the research a semi-automatic classification method has been prepared within the eCognition environment, which enables high accuracy of the resultant map (ca. 90%) to be reached. At the final stage of the research, applicability of radar Terra SAR-X images for vegetation classification was studied.

PL

W artykule zostały przedstawione wyniki prac ukierunkowanych na opracowanie metod klasyfikacji obszarów podmokłych, bazujących na różnych typach danych satelitarnych oraz na różnych podejściach klasyfikacyjnych. Do prac badawczych jako podstawowe materiały wykorzystano wysokorozdzielcze obrazy satelitarne WorldView-2. Zastosowano metodę klasyfikacji obiektowej do osiągnięcia wysokiej dokładności klasyfikacji zbiorowisk roślinnych na obszarach podmokłych. W wyniku prac utworzono półautomatyczną metodę klasyfikacji w środowisku eCognition, która umożliwia osiągnięcie wysokiej dokładności (rzędu 90%). W końcowej części prac badawczych przeprowadzono analizę stosowalności obrazów radarowych Terra SAR-X dla celów klasyfikacji typów roślinności.

2

Application of object-oriented method for classification of VHR satellite images using rule-based approach and texture measures

Lewinski S., Bochenek Z., Turlej K.

Geoinformation Issues

|

2010

|

Vol. 2, no. 1(2)

19-26

EN

New approach for classification of high-resolution satellite images is presented in the article. That approach has been developed at the Institute of Geodesy and Cartography, Warsaw, within the Geoland 2 project - SATChMo Core Mapping Service. Classification algorithm, aimed at recognition of generic land cover categories, has been elaborated using the object-oriented approach. Its functionality was tested on the basis of KOMPSAT-2 satellite images, recorded in four multispectral bands (4 m ground resolution) and in panchromatic mode (1 m ground resolution). The structure of the algorithm resembles decision tree and consists of a sequence of processes. The main assumption of the presented approach is to divide image contents into objects characterized by high and low texture measures. The texture measures are generated on the basis of a panchromatic image transformed by Sigma filters. Objects belonging to the so-called high texture are classified at first steps. In the following step the classification of the remaining objects takes place. Applying parametric criteria of recognition at the first group of objects four generic land cover classes are classified: forests, sparse woody vegetation, urban / artificial areas and bare ground. Non-classified areas are automatically assigned to the second group of objects, which contains water and agricultural land. In the course of classification process a few segmentations are performed, which are dedicated to particular land cover categories. Classified objects, smaller than 0.25 ha are removed in the process of generalization.

PL

W artykule przedstawiona jest metoda klasyfikacji wysokorozdzielczych zdjęć satelitarnych. Została ona opracowana w Instytucie Geodezji i Kartografi i w ramach europejskiego projektu Geoland 2 - serwisu SATChMo. Algorytm klasyfikacyjny, którego celem jest rozpoznanie podstawowych klas pokrycia terenu, został opracowany z zastosowaniem podejścia obiektowego. Jego działanie zostało sprawdzone na podstawie zdjęć KOMPSAT-2 rejestrujących obrazy w czterech kanałach wielospektralnych (4 m) oraz w kanale panchromatycznym (1 m). Struktura algorytmu zbliżona jest do drzewa decyzyjnego i składa się z szeregu kolejno wykonywanych procesów. Podstawowe założenie przyjętego sposobu postępowania stanowi podział treści zdjęcia na obiekty charakteryzujące się niskimi i wysokimi wartościami tekstury. Jest on wykonywany na podstawie przetworzonego filtrami Sigma kanału panchromatycznego. Najpierw klasyfikowane są obiekty z grupy tzw. wysokiej tekstury a następnie pozostałe. Stosując parametryczne kryteria rozpoznania, w pierwszej grupie obiektów klasyfikowane są lasy, roślinność rozproszona, zabudowa oraz tereny pozbawione pokrywy roślinnej. Obiekty niesklasyfikowane są automatycznie dołączane do drugiej grupy obiektów, w ramach której rozpoznawane są wody oraz tereny rolnicze. W toku procesu klasyfikacji jest wykonywany szereg segmentacji dedykowanych poszczególnym klasom. Obiekty mniejsze od 0.25 ha są poddawane generalizacji.

3

Porównanie klasyfikacji obiektowej z tradycyjną klasyfikacją pikselową z punktu widzenia automatyzacji procesu tworzenia bazy danych o pokryciu i użytkowaniu terenu

Lewiński S.

Roczniki Geomatyki

|

2007

|

T. 5, z. 1

63-70

EN

The capabilities of land cover and land use classes identification using object-oriented classification and traditional, so-called pixel-based classification are compared in the paper. The comparison is based on the Landsat satellite image showing a study area of over 423 km2, located within the borders of the Commune of Legionowo (near Warsaw). The results of both classifications were generalised, using a working unit of 1 ha for built-up areas and water and 4 ha for the remaining classes. Object-oriented classification was performed within eCognition software environment. The applied tools of object-oriented classification enabled identification of 18 classes. Subsequent generalisation caused changes only to the area constituting 1.1% of the entire study area. Classification accuracy assessment using the method of visual interpretation and creation of the final land cover and land use database was the final stage of works. The accuracy for the entire study area reached over 94%. Traditional pixel-based classification was performed using so-called hybrid classification, which involves performing supervised classification and then unsupervised classification for unclassified pixels. The pixel-based approach enabled identification of only 8 classes. In the process of generalisation, based on the same principles as in the case of object-oriented classification, 26% of the area of the analysed image was changed. The accuracy of pixel-based classification, assessed by comparing the post-generalisation image to the database obtained after the visual verification of object-oriented classification, reached 72% and 61%, according to the comparison method applied. The results of comparing these two methods of classification prove a significant advantage of objectoriented classification over traditional pixel-based classification. The tools of object-oriented classification enabled identification of twice as many number of classes and a high level of accuracy of the classification process. Moreover, object-oriented classification enables proper generalisation, necessary for creating a land use and land cover database with a defined level of spatial resolution of class recognition.

4

Rozpoznanie form pokrycia i użytkowania ziemi na zdjęciu satelitarnym Landsat ETM+ metodą klasyfikacji obiektowej

Lewiński S.

Roczniki Geomatyki

|

2006

|

T. 4, z. 3

139-150

EN

The results of object-oriented classification based on multispectral and panchromatic Landsat ETM+ data, conducted with the use of eCognition software, are presented in the paper. The classification image was prepared using an algorithm aimed at obtaining a database similar to the one resulting from traditional visual interpretation. After the classification, generalisation of data was performed using a working unit of 1 ha for built-up areas and 4 ha for the remaining classes. Next, raster to vector conversion was performed and the edges of objects delineations were smoothed. Verification using a method of visual interpretation was the last stage of works. After combining the verification results with the classification, the final database was obtained. The applied methods of classification enabled identification of 18 land cover and land use classes, at least four of which cannot be identified using traditional methods. The obtained total accuracy of classification reached 94%. The principles of segmentation of the Landsat ETM+ image based on the panchromatic channel and fused multispectral and panchromatic data are specified in the paper. Fusion was based on PanSharp algorithm within PCI Geomatica software, which preserves spectral characteristics of the original data. The adopted principles of land use and land cover classes were also described. What is particularly worth attention is the method of identification of four built-up land classes, which were extracted from the general class of built-up areas classified using the nearest neighbour method. This task involved use of a parameter defined as a square root of the sum of squares of differences between spectral values of particular channels, while the classification of shadows of buildings was used for identification of built-up areas with apartment blocks. The presented method of classification and processing of the obtained results can support or, in certain cases, entirely replace traditional visual interpretation of satellite images, aimed at creating a land cover and land use database.

5

Klasyfikacja obiektowa narzędziem wspomagającym proces interpretacji zdjęć satelitarnych

Lewiński S.

Roczniki Geomatyki

|

2005

|

T. 3, z. 2

97-106

EN

This paper presents the works performed during realisation of a project, which enabled designing an automatic support method of satellite image interpretation of land cover forms. The method is based on object-oriented classification and it comprises five basic stages: image segmentation, classification, generalisation, conversion of the classification images into vector format, verification of the classification using the method of visual interpretation. Based on a study area of nearly 800 sq. km, rules of object-oriented classification of ASTER satellite images were defined. Object-oriented classification was carried out with a division into 19 classes of land cover and land use. The result of classification was generalised using a working unit of 4 hectares and 1 hectare for water and build-up areas (working units are connected with the scale of 1:50000). Next, raster classification images were converted into vector format. The polygon edges of the vector layer were smoothed in order to make them more similar to the borders identified during visual interpretation. Assessment of classification was performed in order to verify correctness of classification codes and borders identified during visual interpretation. To this end, the procedure used in the CORINE 2000 programme was applied. Interpretation resulted in obtaining information about the differences between classification and interpretation. On the basis of these results, it was possible to precisely specify the accuracy of classification of all classes (within the entire study area) and to create an accurate database of land cover and land use. In the process of object-oriented classification, diverse classification criteria were applied. The metod of classification of mixed forest and apartment blocks is particularly interesting: mixed forests were classified as deciduous or coniferous forests characterised with high non-uniformity, while apartment blocks were identified according to shadows of high buildings. During generalisation of the images, only 1.4% of the study area was changed, which indicates that satellite image segmentation was performed properly. Total accuracy of classification was over 86% and half of the classification mistakes occurred as a result of the fact that an image was taken in spring. The suggested method may accelerate interpretation of land cover and land use even by 50% and in some cases it may even replace visual interpretation. The condition for the method to be effective is defining the rules of object-oriented classification for all types of satellite images, as it was done for the ASTER image. The rules of classification do not necessarily have to cover all classes of land cover (sometimes it may even be impossible). Correct automatic identification of even a few classes will accelerate the process of land cover database creation.