Popis článkuObjektovo orientovaná klasifikácia drevinového zloženia na digitálnych leteckých snímkach zosuvného územia.
[Object-oriented classification of tree species in digital aerial photos of landslide area] 195 - 205.
|Objektovo orientovaná klasifikácia drevinového zloženia na digitálnych leteckých snímkach zosuvného územia
|Miroslav Kardoš, Alžbeta Medveďová, Šimon Supek, Martina Škodová
The aim of the presented paper was to perform an automatic object-oriented classification of tree species composition on the landslide area near Ľubietová (Central Slovakia) using color infrared orthophotos and true color orthophotos with a spatial resolution of 0.2 m, 0.5 m and 1 m. Achieved results then could be used to evaluate the effect of vegetation on slope stability. Species classification was performed using Definiens Professional software. The best classification result of the CIR images was achieved using segmentation of the scale 40 and 0.2 m pixel size, with an overall accuracy of 53%. Salix alba, Salix fragilis and Pinus were best classified according to the KHAT index and classification accuracy. Using higher scales of segmentation, classification accuracy reaches smaller values. The best classification accuracy of 44% was achieved in the orthophoto with 0.5 m pixel at the scale of segmentation 20. The overall classification accuracy of the image with 1 m pixel size reached the lowest values. The best classification result of RGB images is 50% at the segmentation scale 30 and 0.5 m pixel size. The highest classification accuracy of almost 84% achieved in this case the Salix class, but at the 0.2 m pixel size and the scale of segmentation 20. The combination of standard nearest neighbour classifier and membership functions in the CIR orthophoto with 0.2 m pixel size has improved the overall classification accuracy from 53 to 56%. In the Salix alba class, the classification accuracy increased by 53% and in the Salix fragilis class almost by 20%. The classification of a separate class Pinus in the CIR image with 0.2 m pixel size using membership functions reached almost 84% accuracy. In sum we can evaluate the classification results as the average, and in some cases very good.