Object-based mapping of native vegetation and para grass (Urochloa mutica) on a monsoonal wetland of Kakadu NP using a Landsat 5 TM Dry-season time series

Abstract

This paper evaluates the use of multi-temporal Landsat 5 TM for object-based classification of native wetland vegetation and the perennial aquatic weed para grass within Kakadu National Park, Northern Territory, Australia. Using identical training data and segmentation, a nearest-neighbour classification produced from a four-image (dry season) time-series was compared with four ‘single-date’ classifications produced from the individual images of the same series. A 15-class vegetation map generated from the multi-date classification produced an overall accuracy of 82 percent (kappa = 0.80). This was an average increase in accuracy of 25 percent (kappa = 0.28) compared to single-date classifications. The multi-date image composite also improved segmentation quality and spectral separability of vegetation classes. Reliable maps of wetland vegetation, potentially useful for strategic conservation, can be produced by integrated, object-based, analysis of multi-temporal Landsat.