Image processing of multi-frequency polarimetric airsar data to characterise the vegetation communities of the tropical savannas in Northern Australia

Abstract

The aim of this thesis was to develop suitable image processing techniques for multi-frequency polarimetric AirSAR data to characterise the vegetation communities of the tropical savannas in northern Australia. Major steps have been taken that overcome some of the obstacles inherent to such data products. Currently, analysis of such data is severely restricted due mainly to the phenomena of speckle and variation of backscatter signatures which are dependent on the varying incidence angles.

This thesis focused on the use of AirSAR data for estimating bio-physical parameters and achieving a satisfactory land cover classification of vegetation communities in the study area developing methodologies appropriate to this data product. Firstly, the problem of speckle was addressed by investigating the optimum spatial resolution for discriminating the land cover and thus establishing an appropriate filter to use for the suppression of speckle. Secondly, a new method to correct for the effect of incidence angle variation was developed. These two steps were instrumental to assess the correlation between bio-physical parameters of woody vegetation and the AirSAR data, and for the development of a classification methodology that can be implemented using standard image processing software.

The investigation of the optimal resolution for the delineation of vegetation communities in the study area was found to be between 20 and 27m. The resultant averaging process to reduce the data product to this resolution has been shown to have the additional benefit of substantially reducing the effect of speckle. A new method developed for this project allowed the reduction of the effect of variation in incidence angle. The correction method developed applies a histogram equalization to the lines of constant incidence angle and was shown to be effective through an evaluation against an existing land cover classification of the study area.

No empirical evidence was found for a relationship between above ground woody biomass and AirSAR backscatter. However, the land cover classification of the AirSAR data was successfully implemented using standard image processing techniques. Differentiation between species and tree density was possible where there is a significant difference in the vegetation structure. The classification accuracy is very similar to that achieved by a land cover classification using Landsat TM data. This is particularly encouraging as the AirSAR data achieved far greater accuracy in the delineation of Floodplain vegetation and the Pine plantation, two cover types for which the Landsat data registered a low classification accuracy due to the overlap with grassland and woodland communities respectively. A synergism of optical and AirSAR data has the potential to produce a more detailed and more accurate classification than either data set on its own.

Date of Award	2000
Original language	English
Supervisor	Waqar Ahmad (Supervisor) & Greg Hill (Supervisor)