Grassland Vegetation Classification in The Krkonose Mountains National Park using WorldView-2 Imagery
Kupkova, L.1; Pomahacova, M.1; Brezina, S.2; Cervena, L.1; Jelenek, J.1
1Charles University in Prague, Faculty of Science, Dpt. of Applied Geoinformatics and Cartography, CZECH REPUBLIC; 2The Krkonose Mountains National Park Administration, CZECH REPUBLIC

The aim of this paper is to evaluate suitability of WorldView-2 imagery for grassland associations and their management practices classification in the model area of The Krkonose Mountains National Park (Giant Mountains). The classification is based on the legend compiled by a botanist and includes eight classes (five grassland assotiations classes and three management classes):

1. Cut grass
2. Pastures
3. Meadows with dominat Nardus stricta
4. Degraded meadows
5. Meadows with dominant Vaccinium sp. and Caluna sp.
6. Meadows with dominant Trisetum flavescens, Dactylis glomerata or Alopecurus pratensis
7. Waterlogged meadows
8. Valuable anthropogenic associations
9. Mulched meadows

In order to eliminate the effects of other types of land cover on the classification accuracy, a mask of grasslands was created. The mask was created based on the ZABAGED vector data (Fundamental Base of Geographic Data scale 1:10,000) in combination with image unsupervised classification. The significance of spectral bands of WorldView-2, as well as significance of selected vegetation indices and components from Principal Component Analysis (PCA) for distinguishing particular classes of grassland vegetation (together 24 variables) were evaluated using discriminant analysis. 14 input channels/parameters based on the results of discriminant analysis were used for classification (all the WorldView-2 bands and the vegetation indices NDVIGreen, NDVIBlue, NDVI, NDVIYellow, SAVI, RVI bands). Classifications using neural networks method and maximum likelihood method were performed in ENVI software, and the results were compared. This approach was used for two images, one was acquired in August and second in September 2012. The accuracy assessment was based on the training data set. All the overall accuracies were better than 80 % but generally the better results were achieved by the neural network classification method. Interesting differences between the results of classification based on image acquired in August and in September were recorded. Higher accuracy of grassland associations classification was achieved based on the image acquired in August while better differentiation of particular management types was accomplished by classification of the image acquired in September.