Discrimination of Eucalyptus canopy from airborne linescanner imagery using Markov random field modelling

Abstract

Remotely sensed vegetation classification would be performed more effectively if the areas of interest, such as forest canopies, could be isolated from the rest of the image (i.e. the background). This background is typically made up of features such as soil, understorey vegetation, shadows, trails and manmade structures. In order to discriminate between Eucalyptus canopies and background features from high resolution airborne linescanner data collected over the Mount Eccles national park (south-eastern Australia), a supervised Markovian texture modelling approach was developed. To account for the spatially random nature of eucalypt canopies, the texture corresponding to the canopy area was modelled using a parametric Markov random field. The band used for this processing was the near infrared channel which proved useful in highlighting the vegetation owing to its photosynthetic activity. A probability map locating all textures in the image, which were similar to that of a training sample, was produced and then thresholded. The simulations and the validation procedure suggest that the extraction of the entire eucalypt canopy is possible using the algorithm. This provides a means of pre-processing high resolution airborne data to create masks for further image analysis, and in particular for forest health mapping applications.