Classification scheme

In Sub-section 5.1.2 it has been focused only on certain vegetation habitats (based on a subjective visual recognition of characteristic vegetation patches using ancillary data), not aiming at the identification of each vegetation pattern in the test site. Therefore, a part of the area remained unclassified (in the case of DR, 2008 it was more than 30%). Defining class

labels for each vegetation pattern present in the image, and hereby, analysing the complete vegetation cover gives a better understanding about the target site and it is required for the testing of classification transferability (Chapter6).

In the ecological community mapping study of Rapp et al. (2005) applied in an open and forested wetland site in the United States it has been described that the addition of new classes (i.e., using an extended classification scheme) describes the complexity of local veg-etation patterns appropriately and improves the vegveg-etation classification result.

5.2.1 Extending the classification scheme

The aerial image from 2008 for the test site of Dunaremete was applied for the investigations of classification scheme extension. Before the step of sample selection for the complementary classes, existing classes in the classification scheme had to be reconsidered. Nevertheless, due to the findings in Sub-section 5.1.7, it was reasonable to apply the originally chosen classes and complement them with additional ones.

Extending the classification scheme with classes mostly smaller in size than the earlier iden-tified ones is not straightforward, if the time of image acquisition (2008) and the collection of ground reference information (botanical field survey, 2004) is significantly different and in this case uncertainties are expected. In order to complement the classification scheme, botanical and silvicultural maps and visual interpretation of the image were applied. A hi-erarchical classification scheme was built with two levels, however, the most significant one is the “detailed” level (Level II), where classes were directly distinguished based on physiog-nomical appearance and contrary to that the upper level (Level I) groups these classes by semantics only. In the study ofJohansen et al.(2007) structural vegetation classes followed the classification scheme of Terrestrial Ecosystem Mapping derived from aerial (analogue and digital) imagery, which also helped here to find the appropriate classes for Level I.

Complementary classes to the detailed level (Level II) were partly defined using habitat maps (O10: Vegetation on edges and dams, T1: Arable land from AppendixB, FigureB.2), using silvicultural inventory (FATI1, 046: Domestic poplar-robinia; 058: Other hardwood, from Appendix B, Figure B.3) and applying visual interpretation (Road, Bare soil mixed with grass, Young stand, Shadow). Although the class of Shadow does not build a real vegetation class, its application was important, since those areas significantly vary from

Chapter 5. Spectral-textural classification 58 other classes and have a common occurrence in the current image. The new hierarchical classification scheme is presented in Figure5.10.

Class of water bodies was separated, as it has been described in Sub-section5.1.1followed by the classification of road, both based on irregularly shaped segments after the combination of quadtree and multi-resolution segmentations with the use of vegetation index (BlueNDVI) and Brightness (average DN of the bands) values in the classification.

Focusing on the classification of vegetation, class descriptions for the CDBF classification algorithm were derived based on sample objects (chessboard segments as applied in Sub-section 5.1.2) using the earlier chosen descriptors (one spectral and 4 textural parameters).

In addition, a small modification was applied to the use of vegetation index. Besides the formerly chosen BlueNDVI (called as modified NDVI in Table 5.2), GreenNDVI=(NIR-G)/(NIR+G) was tested, since it is vital from the aspect of further analysis (temporal classification transferability) to consider transferable indices in case of images with different band combinations in different years (Table 3.1). That’s why the use of GreenNDVI (with NIR and G bands) was preferable in comparison to BlueNDVI (with NIR and B bands).

Figure 5.10: Hierarchical classification scheme (Level I, II) applied to the test site of Dunaremete, 2008

Figure 5.11: Classification result with the complemented classification scheme by the CDBF algorithm using GreenNDVI as vegetation index and the earlier chosen textural

parameters, applied to the test site of Dunaremete, 2008

5.2.2 Result & discussion

Classification result is presented in Figure5.11by the class description based fuzzy algorithm with the use of GreenNDVI besides the same textural (GLCM/GLDV) features as applied previously. Around 17% of the site still remained unclassified, although this was much less than in case of the originally applied classification scheme (around 30%). The best overall accuracy was 88% and Kappa index was 0.87 (not including the class of road in the accuracy calculations because of the 20 m×20 m chessboard-based accuracy assessment method). These accuracy values are very close to the accuracy measures with the originally applied simple classification scheme (Table 5.4). The difference between the application of different vegetation indices (BlueNDVI or GreenNDVI in the combined feature set with textural measures) is summarized in Table 5.8.

An important issue regarding the classified vegetation habitats is related to classes Domestic poplar-Robinia and Hybrid poplar. Although, the mixed type of Domestic poplar-Robinia (DP-R) is only present in one compact site (Appendix B, Figure B.3), in the classification result other HP stands were labelled as DP-R. It could be established from the silvicultural

Chapter 5. Spectral-textural classification 60 data concerning species related tree ages, that those forest stands have the same age, which is different from any other stands in the study area. Nevertheless, a statistically significant analysis for the assessment of age structure could not be carried out on that area because of the small size of the test site.

Table 5.8: Comparison of accuracy measures based on different vegetation indices in the combined (spectral & textural) feature set applied in the CDBF algorithm, DR, 2008.

VI in the feature set OA Kappa BlueNDVI 86% 0.84 GreenNDVI 88% 0.87