Abstract
The ecological applications of remote sensing have progressed over time, largely as a result of improvements in remote sensing technology. Increases in both the spatial and temporal resolution of available imagery has allowed ecological phenomena acting on finer scales to be resolved and analysed. The proliferation and technological advancements of remotely piloted aircraft systems (RPAS) represents a new avenue for remote sensing in ecological applications, facilitating the possibility of capturing very high resolution imagery ( ~ 5 cm pixel size) over relatively large areas (~ 100 ha). Very high resolution RPAS imagery, and its derived products, is well suited to analysing both the structure and spatial distribution of plant life forms. RPAS based vegetation mapping has predominantly occurred in agricultural applications, with fewer instances of ecologically focused RPAS vegetation mapping, especially in New Zealand. In order to address the gap, this study addressed three primary research questions which aimed to develop and assess a method to use RPAS data to map tall tussock grass in a mountain grassland environment, and to test its utility of this for multi-scale integration with other data sources. Chapter 2 considered the question: Can RPAS imagery and derived high resolution surface model be used to accurately detect and map the spatial distribution of tall tussock grasses? Chapter 3 considered: How does tall tussock distribution and density vary in regard to environmental drivers such as topography, elevation and snow cover at the sub-catchment scale? Finally chapter 4 considered: How can RPAS imagery be used to refine landscape scale mapping of tall tussock cover? Data collection was undertaken in a sub-catchment of the Leopold Burn, Pisa Range, New Zealand; a site chosen as it was already in use as a test site for a wider study investigating the hydrological dynamics of the Clutha River catchment, and exhibited substantial spatial variation in tussock cover.
In order to inform the tall tussock mapping process, a controlled pilot study was carried out at Okia Reserve, Otago Peninsula, New Zealand. Comparison was made between how plant life forms and artificial objects were represented in RPAS data, which confirmed that vegetation exhibited a signal that was detectable in RPAS data. In order to refine this signal for the Leopold Burn study site, RPAS data was first processed using a digital surface model (DSM) differencing approach. The signal was further refined using a numerical threshold and a sequence of majority filtering operations. The map was validated at each stage by visually comparing it to a high resolution ortho-photomosaic, and numerically comparing the number of tall tussocks observed by the map to the number manually counted on the ortho-photomosaic. The final version of the map achieved a R2 value of 0.72 and a Root Mean Square Error of 15.72 tussock / 100 m2 when comparing the number of tall tussock observed with number manually counted. Once complete, the map could then be leveraged to create a raster layer measuring the density of tall tussock / 100 m2. This representation of tall tussock density was analysed in regard to environmental drivers such as topography, elevation and snow cover. Tall tussock density was found to be significantly lower (p < 0.001) on slope faces than in gullies, and significantly lower in snow free areas than in persisting snow covered areas as recorded on the 9th September 2017.
To demonstrate the integration of the tussock distribution map with other data sources, the maps of tall tussock distribution and density were used as a ground-truth to refine satellite imagery, as part of a scaling up approach to mapping tall tussock distribution for the entire Pisa Range. Sentinel 2 A & B satellite imagery was used to create a normalized difference vegetation index (NDVI) image of the Pisa Range. NDVI values that were representative of tall tussock grass were identified and refined using the tussock distribution and density maps. These NDVI values were then applied to the entirety of the Pisa Range, to create a landscape scale map of tussock distribution. The resulting distribution map was compared to an existing range of tall tussock grass, taken from the New Zealand Land Cover Database (LCDB). Ultimately RPAS imagery and derived products provide a means to remotely map the distribution and density of tall tussock grassland at a range of spatial scales. The methodology outlined in this thesis has proven to bring significant advantages over field based approaches to analysing tall tussock grassland, in terms of both accuracy and efficiency.