|Institution:||University of New South Wales|
|Department:||Biological, Earth & Environmental Sciences|
|Keywords:||Multi; Object; Vegetation; Mapping; Delineation; Hyperspectral; Crown; Fuel load; Fire; Australian; Scale; Remote sensing; Lidar|
|Full text PDF:||http://handle.unsw.edu.au/1959.4/54474|
An underlying premise of any segmentation method is that spectral similarity and thematic similarity are synonymous. This assumption holds true for image objects at an individual tree crown scale and they can be classified with a degree of accuracy. However, at coarser spatial scales, a large patch of vegetation can encompass a variety of thematic attributes. Mapping native vegetation using remote sensing suffers from an inability to make meaningful predictions through a change in scale. I propose that heterogeneous vegetation needs to be analysed across multiple scales to categorise it as a vegetation community. A multi-scale, object-based, hierarchical approach was introduced to generalise floristic data collected at the plot scale to a vegetation community map using remote sensing. This framework uses the cover and abundance of classified tree crown objects to inform the classification of larger patches of vegetation. Community scale image objects can then be named using the same hierarchical framework used by ecologists in plant ecology. Machine learning classification algorithms and patch scale segmentation algorithms were reviewed and benchmarked for this application. A crown delineation algorithm was formulated as well as a new way to combine lidar with optical imagery. The scope of this thesis was limited to three sensors: the HyMap hyperspectral airborne scanner, small footprint lidar, and the multi-spectral SPOT-5 satellite. To ensure that the results are relevant, the fieldwork for this thesis was based largely on operational standards. The result was a vegetation map classified on cover and abundance of dominant crown species. The extra resources required for individual tree crown surveys and the difficulty of analysis in highly diverse ecosystems are the main limitations. Vegetation structure was assessed by quantifying forest fuel load using remote sensing. The correlation between field derived attributes and vegetation indices was stronger when narrow band hyperspectral vegetation indices were used. Small footprint lidar successfully penetrated the canopy and offered quantitative information about the structure of the understorey. However, the total fuel load assessed in the field was dominated by leaf litter component in wet forest, which was problematic to quantify with remote sensing.