Object-based remotely sensed image classification for identification of lowland native grassland communities in the Tasmanian Midlands

In 2012, construction begun on Tasmania's largest irrigation project; the Midlands Water Scheme. It was determined, however, that construction and subsequent operation of the scheme could potentially negatively impact endangered lowland native grassland communities found in the region. Both the Australian government and Tasmanian State government issued a mandate for increased monitoring of these communities, however no formal program has been undertaken. Concerns were also raised about the appropriateness of the current vegetation mapping methods used in the State, as they rely on manual image interpretation techniques. Such techniques are time consuming and often produce subjective results, therefore the need for a new remotely sensed approach relying on semi-automated techniques was required. The aim of this thesis is to investigate the utility of remote sensing as a means of providing frequently updatable maps of lowland native grasslands. The first objective of this study was to classify Tasmanian lowland native grassland communities from moderate spatial and spectral resolution satellite imagery and compare these classification results to traditional field mapping techniques. The aim of the classification was to distinguish three types of lowland grassland communities (Poa, Themeda, and grassland complex), and broader agricultural and woodland classes. An object-based image classification was undertaken on segmented Landsat ETM+ and WorldView-2 datasets using 50 random forest models trained using random subsets of reference points generated from field samples collected by the Tasmanian Land Conservancy. Validation was performed using the reciprocal points not used to train the models. Resulting average accuracies were moderately high, ranging between 55-88% for the Landsat ETM+ results, and 56-87% for the WorldView-2 result. The currently existing community map (TASVEG) was also evaluated, and found to have comparatively poorer accuracy for all classes. Analysis of Variance Results (ANOVA) indicated a significantly higher accuracy of the WorldView-2 result compared to the Landsat ETM+ result for the dry woodland and Themeda grassland classes, but no other statistically significant differences in classification accuracy were detected. The results of the first study indicated the need for higher spectral resolution datasets. Therefore, the second objective of this study was to determine whether discrimination of lowland grassland communities was possible based solely on spectral properties. A field campaign was undertaken to collect data using an ASD handheld-2 spectrometer. The dataset was resampled to match the broadband resolution of Landsat OLI and WorldView-2 to compare results from narrowband and broadband approaches. Spectral signatures were classified using a random forest classifier at their full spectral resolution as well as spectrally convolved broadband equivalents to simulate coarser spectral resolutions of different sensors. ANOVA results indicated that classification accuracy for the Themeda class was highest when using a reduced narrowband model in which correlated bands had been removed. The analysis also indicated that significant problems in differentiating between Danthonia grasslands and Themeda grasslands at all spectral resolutions. Variable importance measures indicated strong separability in wavelengths associated with pigment decomposition, photosynthetic rate, and water content. Confusion rates and variable selections showed that differentiation between all classes was improved using high spectral resolution signatures and by grouping vegetation based on photosynthetic pathway. Uncertainty within image segmentation can be particularly problematic in heterogeneous environments with indistinct class boundaries. Therefore, the third objective of this study was to devise a method capable of predicting class-specific optimal segmentation scale. A new method of segmentation assessment was developed that evaluated segmentation performance using a combined geometric and thematic assessment. Two trials of the index were run, the first in an urban environment with clear object boundaries, and the second in the native grassland study area in which vegetation communities intergraded significantly. Both trials indicated successful prediction of optimal segmentation scales, and the optimal segmentation scale parameters identified by the new optimisation algorithm resulted in significant improvements in class delineation and characterisation. The fourth objective was to combine the findings of the previous three chapters in order to optimise approaches to lowland native grassland community mapping. A 15 cm, 20 band hyperspectral orthomosaic was acquired using an Unmanned Aircraft System (UAS). The spectral regions measured by the sensor corresponded to those identified previously as being of key importance. Segmentation assessment was performed using reference data acquired from transects and at additional randomly placed field plots. The UAS orthomosaic was segmented based on class-specific optimal scale predictions and classified with a random forest model. Overall, the methods outlined in the study provide a targeted approach to lowland native grassland mapping capable of providing reliable community maps at a range of scale levels. Accuracies for remotely sensed results at the landscape scale significantly improve on those from traditional mapping techniques, and therefore remote sensing is deemed to be a viable mapping approach for lowland native grasslands. In summary, the analysis undertaken in this thesis shows that lowland native grassland communities can be accurately identified using remote sensing techniques. The results obtained here provide several key findings that illustrate the importance of data selection. Key spectral regions for differentiation of communities were identified, and can be used to improve variable and dataset selection in future analysis. This thesis contributes significantly to the broader field of grassland mapping and monitoring as it provides important case-studies proving that community level classification is possible at varying spatial scales. Finally, this thesis contributes to the field of object-based image analysis, and the prediction of optimal segmentation scale. The prediction of class-specific and scene-wide optimal segmentation scale is a novel development in the field. Additionally, the use of thematic accuracy in conjunction with spatial accuracy to determine optimal segmentation scale is a new technique developed in this thesis. Overall, the results of this thesis provide important findings that can be used to further guide conservation and targeted management of lowland native grassland communities.