Landsat TM image segmentation for delineating geological zone correlated vegetation stratification in the Kruger National Park, South Africa

详细信息    查看全文

• 作者：Christopher Munyati^a ; ^{cmunyati@csir.co.za} ; ^{chrismunyati@yahoo.co.uk} ; Thihanedzwi Ratshibvumo^b ; Jason Ogola^c
• 关键词：Vegetation ; Habitat mapping ; Geology ; Soil fertility ; Remote sensing ; Image classification
• 刊名：Physics and Chemistry of the Earth, Parts A/B/C
• 出版年：2013
• 出版时间：2013
• 年：2013
• 卷：55-57
• 期：Complete
• 页码：1-10
• 全文大小：1251 K

文摘

Image classification approaches are widely used in mapping vegetation on remotely sensed images. Vegetation assemblages are equivalent to habitats. Whereas sub-pixel classification approaches potentially can produce more realistic, homogenous habitat maps, pixel-based hard classifier approaches often result in non-homogenous habitat zones. This salt-and-pepper habitat mapping is particularly a challenge on images of savannas, given the characteristic patchy texture of scattered trees and grass. Image segmentation techniques offer possibilities for homogenous habitat classification. This study aimed at establishing the extent to which established, field surveyed and geology-related vegetation types in South Africa¡¯s Kruger National Park (KNP) can be reproduced using image segmentation. Rain season Landsat TM images were used, selected to coincide with the peak in vegetation productivity, which was deemed the time of year when discrimination between key habitats in KNP is most likely to be successful. The multiresolution segmentation mode in eCognition 5.0 was employed, object classification accomplished using the nearest neighbour (NN) classifier, using object texture and training area mean values in the NN feature space.

Compared to delineations of the vegetation types of KNP on a digital map of the vegetation zones that was tested, image segmentation successfully mapped the zones (overall accuracy 85.3 % , K^ = 82.7 % ) despite slight shifts in the location of vegetation zone boundaries. Maximum likelihood classification (MLC) of the same images was only 37 % accurate (K^ = 24.2 % ). Whereas the vegetation zones resulting from MLC were non-homogenous, with considerable spectral confusion among the vegetation zones, image segmentation produced more homogenous vegetation zones, comparably more useful for conservation management, because realistic and meaningful habitat maps are important in biodiversity conservation as input data upon which to base management decisions. Image segmentation appears to be a useful approach in mapping savanna vegetation.