6. Thesis Context
4.3 RF classification and validation
The areas of misclassification within the results are nearly identical to the areas misidentified in the predicted class extents derived from the segmentation assessment process. The final community extents obtained from the three different segmentation scales are very similar, however, there is some disagreement between the results for MC and the other two results. The disagreements in extent is primarily associated with the Danthonia class, particularly surrounding the southern boundary of the lagoon. The results obtained for the MSS and TRO results delineate this area well, however this is not the case for the MC results, in which the area is significantly overgeneralised. Additionally, there is still significant overprediction of the Acacia class in the north-westerly and south-westerly corners of the scene in all results. The primary source of inaccuracy in the classification results across both validation tests is confusion between the Danthonia and Wilsonia classes. The confusion in this case is only in one direction, in that a large proportion of the Danthonia class is erroneously classified as Wilsonia, while there is very little misclassification of Wilsonia as Danthonia. As the two communities intergrade extensively, the establishment of discrete reference objects for segmentation validation was very difficult. Inaccuracy in the creation of these reference objects is likely to be the case of the poor overall accuracy obtained for the Danthonia class, however, the results of Chapter 3 indicated that there is significant difficulty in differentiating the two classes based solely on spectral properties. As the two classes additionally occur in the same area, primarily on low-lying saltpan, the DSM and slope variables are not likely to increase separability between the two classes. The Danthonia class exhibits significantly different final classification accuracies between the validation performed using the reference objects and the validation based on the image transects. The significant increase in classification accuracy in MS and TRO results based on the image transects over the results for the reference segments indicate the need for highly detailed fine-spatial scale validation datasets in order to accurately assess community intergradation. Additionally, the segmentation scale factors predicted as optimal for the Wilsonia and Danthonia classes are very similar, indicating similar spatial scales of occurrence.
Based on the validation using the reference segments, the most accurate results were obtained for TRO segmentation. The MSS segmentation obtained similar, but slightly lower final classification accuracies. The scale factors used to generate the MSS inputs are all close to a value of 400. The classification accuracies for the MC segmentation are also similar to the accuracies of the MSS and TRO segmentations. When the MC segmentation is validated against the transects, it achieved the highest overall classification accuracies of any of the results. However, the degree of overgeneralisation present in this result is concerning, and therefore it can be determined that the scene-wide optimal segmentation scale is best predicted using the TRO approach, rather than MC approach.
5 Conclusions
This chapter presents the results of an RF classification approach for identifying lowland native grassland communities in the Tasmanian Midlands using high spatial and spectral resolution
UAS imagery. Class-specific optimal segmentation scales were predicted and subsequently tested on the data. The results of the segmentation assessment indicated very similar optimal segmentation scales for classes, with scale factors ranging between 250 and 500. This gives a range in mean object size between 29.17 m2 at scale factor 250, and 93.28 m2 at scale factor 500. Scene-wide segmentation scales were also predicted using the two approaches identified in Chapter 4. Overall, the scene-wide prediction of optimal segmentation scale derived for TRO produced a very similar prediction of class extent to MSS result generated using the class- specific optimal scale factors. The behaviour of the various component metrics used to calculate the final optimal segmentation scale Sq is in keeping with the behaviour seen in the case studies presented in Chapter Four. There is a tendency for the scale predictions identified by U give an oversegmented result, as was also seen in Chapter Four. The final predictions based on Sq, however, are generally very good, and indicate that the segmentation assessment method is capable of performing well in environments with a high degree of community intergradation.
Classification accuracies obtained for the three input segmentations are very similar for both validation results. Overall, the optimal scale factor predicted using TRO scale factor of 400 performed the best across the reference object validation with a mean accuracy of 72.4%, while the scale factor prediction based on MC performed best based on the transect validation with a mean accuracy of 93%. The results indicate that there is potential inaccuracy in the training points, particularly in regards to the Danthonia and Wilsonia classes. The higher classification accuracies obtained for the transect validations indicate that accurate assessment of community gradients requires the collection of high spatial frequency field observations over a large area. The small extent covered by the two transects means that assessment of communities in this manner is limited for this result. Future studies could benefit from the creation of multiple transects with closely spaced observations to aid in a more robust assessment of classification results.
6. Thesis Context
This chapter presents an approach to lowland native grassland mapping that combines the findings of previous chapters. Accuracies achieved in this chapter indicate that there is a need for high spatial resolution datasets in all stages of analysis. This means that training and validation data need to be collected using field-based approaches in order to provide accurate assessment of classification results. The findings of this chapter reiterate previous findings, in that there is significant difficulty differentiating C3 dominated lowland native grassland species
from other native C3 native vegetation types. The importance of topographic variables within
the classification workflow is also reiterated in these findings. This chapter also provides a case-study showing the potential applications of snapshot hyperspectral sensors for ecological research applications. This chapter has shown that these new sensors can function well, and provide high quality datasets for analysis.
Chapter 6
Conclusions
The aim of this thesis was to develop remote sensing methods for identification and mapping of endangered lowland native grassland communities in the Tasmanian Midlands region. The key goal was to identify methods capable of community identification that could contribute to frequently updated maps that can be employed for long-term monitoring of both status and extent. As current mapping methods are based primarily on manual digitisation of aerial photography, there was a need to assess the applicability and utility of various remotely sensed data sources in order to determine the optimal approaches for community mapping. This raised a number of questions regarding the selection of datasets. Classification tests were performed on different sources of remotely sensed data in order to determine the applicability of different spatial and spectral sensor resolutions for lowland native grassland community mapping. Additionally, the selection of an optimal scale of analysis has been frequently highlighted as an issue for accurate mapping of lowland native grassland communities. Here, a method for the identification of optimal image segmentation scale is proposed in order to mitigate this issue. The methods investigated, developed, and employed in this thesis have demonstrated that remote sensing is a viable option for lowland native grassland community identification and mapping provided that the spatial and spectral resolution is appropriate. This thesis has identified several key factors that contribute to success in community mapping, which can be used to produce accurate maps of lowland native grassland extents.
6.1 Multispectral approaches