2. Methods
4.2 Class performance and confusion rates
For both classification results, mean accuracies are similar between the two sensors. Class- specific mean accuracies are very similar between the two results, with only two classes, dry woodland and Themeda grassland showing significantly higher accuracies in the WorldView- 2 result compared to the Landsat ETM+ results. In addition, overall mean accuracy was also shown to be significantly higher in the WorldView-2 result.
The poorest performing class in both classification results was the grassland complex class, with mean classification accuracies of 54.82% for Landsat ETM+ and 56.26% for WorldView- 2. This class also has the highest standard deviation of all classes in both sets of results. The poor performance and high variability observed for this class is likely due to the broad definition of the class in terms of species composition and dominance. The grassland complex is the most loosely defined of the lowland native grassland classes, and serves as a generalised class accounting for any native grassland not dominated by Poa labillardierei or Themeda triandra (Kitchener and Harris, 2013). Typically, the community is dominated by Danthonia species, but the formal community benchmark states that other grassland species may be dominant or co-dominant (Kitchener and Harris, 2013).
In the Landsat ETM+ results, misclassification of the grassland complex class is primarily due to confusion with the woodland class, and secondarily with the Poa class. Confusion between native grassland classes and the woodland class is expected given the high degree of community intergrading. In the WorldView-2 result, confusion is again greatest with the woodland class, but the second highest rate of confusion is with the agriculture class. Confusion
with other native grassland classes is very low, and likely due to the increased number of bands in the red edge region of the spectrum, which are better able to detect varying phenology between communities. The high level of confusion with the agriculture class is likely due to the smaller segmentation scale employed for the WorldView-2 dataset. Due to the fact that the grassland complex and agricultural classes often share class borders, any potential misplacement of segment boundaries is likely to manifest as misclassification between the two classes.
The best performing class in both sets of results is the agricultural class. This class has the advantage of being the most distinct in terms of its spectral and textural properties. The majority of species incorporated in this class are introduced, and have planophile canopies. All other classes are also composed of native vegetation, which exhibits typical sclerophyllous adaptations and morphology, while the introduced agricultural species do not. These differences result in the agricultural class having distinctive reflectance properties associated with high photosynthetic rate, greenness, and water content. Additionally, the agricultural land in the study site is managed much more intensively than the native areas. Such management practices include the use of irrigation, fertilisation, and the sowing of both crop and pasture species. These practices result in distinctive textural properties, such as clear row marks from sowing, that clearly differentiate them from more native areas where growth is less constrained. The dry woodland class is the second highest performing class for both results. ANOVA results indicate that there is a significant improvement in classification accuracy for this class in the WorldView-2 data over Landsat ETM+ results. Confusion values are similar between the datasets. The exception to this, however, is the grassland complex class, which has a reduction in confusion for the WorldView-2 result. The definition for the woodland class is quite broad, as it covers three forested land cover classes as originally identified in the TLC data. Two of the original classes, DPO and DVG, are Eucalyptus dominated woodland variants, with low levels of floristic diversity in their understories (Kitchener and Harris, 2013). The remaining class, NBA, is an open woodland dominated by Acacia or Bursaria species over a dense grassy understorey, typically of Themeda triandra or Danthonia sp. (Kitchener and Harris, 2013). The classification frequency counts observed for the woodland class in figure 2.6 indicate that the areas with higher rates of misclassification are associated primarily with the known distribution of the NBA class, as shown in figure 2.2. The overlap in dominant species between the NBA class and the native grassland classes, coupled with the sparseness of the tree cover associated with this type of woodland, is the source of significant confusion within the classification results for both datasets.
The Poa grassland class shows good performance in both sets of results, with no significant difference in accuracy between the two datasets. Confusion for this class is primarily with the dry woodland and agricultural classes. Misclassification with the woodland class is due to reasons discussed previously, such as overlapping constituent species between classes. Both the agricultural and Poa classes have similar phenological cycles, as species are primarily C3
species. At the time of acquisition, both classes were entering a period of senescence over the warmer summer months, which may have resulted in confusion due to lower rates of
photosynthetic activity. Additionally, some of the area covered by the Poa class is actively managed through the use of irrigation and fertiliser, though not as extensively as the designated management practices in both classes may be an additional source of confusion due to increased variance in class spectral properties. There is very little observable confusion with the Themeda class which occurs almost exclusively in unmanaged areas of the property. The final native grassland class, Themeda, has the strongest classification performance of the three lowland native grassland types, and additionally has significantly improved performance in the WorldView-2 results. Confusion for this class is primarily with the woodland class, as per reasons discussed previously. For the Landsat ETM+ result, confusion with the remaining three classes occurs at similar levels for each, while in the WorldView-2 result, confusion with the grassland complex and agriculture classes is greatly reduced. The main differentiating factor for the Themeda class is its varying phenology and photosynthetic pathway as it is a C4
species. Themeda triandra also exhibits a very characteristic red colouration, which, coupled with the increasing growth rate of the species at the time of data acquisition, may be a reason for its improved performance in the WorldView-2 result, given the increased number of bands in the red and red-edge regions of the spectrum.