Object-based image analysis of ultra-fine spatial resolution imagery acquired over a saltmarsh environment by an Unmanned Aircraft ASystem (UAS)

Unmanned Aircraft Systems (UAS) are an emerging technology in the field of remote sensing. Two fundamental differences of UAS when compared with traditional aerial remote sensing platforms are the operational altitude and payload capacity. The lower operational altitude of UAS generates ultra-fine spatial resolution data (< 10 cm). The small size of most UAS platforms allows scientific research groups to transport and operate the platform within small focused study areas. However, a small size also places physical limitations on UAS sensor payload carrying capacity. This requires a compromise between sensor functionality, cost, and weight. Sensor feature reduction or miniaturisation achieves this compromise but at the cost of data quality. This thesis advances UAS remote sensing through an exploration of the development, scale analysis and application of ultra-fine spatial resolution UAS data. Two sites of remnant cold temperate saltmarsh vegetation in Tasmania were selected to assess UAS remote sensing. Frequent salt water spray and tidal inundation within saltmarsh create a saline gradient that limits the establishment of larger canopy species. This has resulted in the dominance of salt and water-logging tolerant herbaceous and small woody shrub species. Despite the harsh environmental conditions, the combination of land wash-off and tidal inundation both readily supply and redistribute nutrients, creating one of the most environmentally productive environments. Measuring the finescale vegetation distribution and productivity of cold temperate saltmarsh vegetation requires the ultra-fine spatial resolution data of UAS. In this study, a sensor correction methodology was designed and implemented to reduce the effects of noise and distortion in the 6-band multispectral miniature multiple camera array (mini-MCA) produced by Tetracam. This methodology includes techniques for sensor noise reduction using dark offset subtraction, vignetting correction through at field look-up tables, and lens distortion correction by implementing the Brown-Conrady model. The sensor correction framework is demonstrated through a real-world application on UAS-derived saltmarsh data. Chapter 2 demonstrates that sensor noise and distortions can be satisfactorily corrected in 6-band Tetracam mini-MCA data acquired from a small multirotor UAS. Once image data are constructed, the next challenge lies in deconstructing the complex ultra-fine spatial resolution UAS data to derive meaningful information. The increased resolving power of UAS data provides spatial measurements of image features at scales previously too small to distinguish. This results in increased spatial complexity as finescale structural variation becomes measurable. A key challenge is to disassemble and simplify this fine-scale variation for the extraction of information. This is achieved through two frameworks that provide a meaningful spatial generalisation using image texture models and geospatial object-based image analysis (GEOBIA). Image texture is defined as the replications, symmetries and patterns in tonal structure. Image texture models are used to quantify the tonal structure in a local neighbourhood into a single, statistical measure. The large number of available texture models and parameters, as well as the dependence of texture on image scale and context, complicates the optimal selection of image texture measures. In Chapter 3, a texture selection methodology is introduced to provide a rapid, broad assessment of image texture. The texture selection framework is illustrated using a 6-band multispectral dataset of a saltmarsh site. Four texture models are investigated: a simple first-order kernel, the greylevel co-occurrence matrix (GLCM), local binary pattern operator (LBP), and wavelets. Using image subsets, 693 texture measures are extracted from seven vegetation and nonvegetation groundcover classes. A random forest ensemble classifier was used to quantify the relative class-specific importance of individual texture measures. A correlation threshold was used to remove highly correlated, less important measures before forward inclusion was used to identify the minimum optimal number of texture measures. The number of required texture measures was linked with class spectral variation, with spectrally complicated classes requiring more measures. The performance of the measures was tested across the entire image, with a recorded improvement of 17.2% in overall classification accuracy with the inclusion of selected texture measures. GEOBIA extends traditional pixel-based analysis through the segmentation of imagery into meaningful objects. The results of the initial segmentation determine the units of analysis, and their accuracy is therefore paramount to the entire analysis. As with texture, image segmentation is dependent upon image structure and content. In Chapter 4, a methodology is presented utilising image subsets to identify class-specific relative scales of image segmentation through identifying under- and over-segmentation. Reference objects were used to compare image segmentation results against a meaningful real-world abstraction. Under-segmentation was tested using spatial area metrics, and was quanti fied on a class-by-class basis whenever a subset recorded 100% omission in labelling. Over-segmentation was identified by extracting the statistical properties of objects and then testing the separability using a random forest model. The insuficient spatial generalisation of over-segmentation resulted in reduced class separability. Furthermore, spatial accuracy was limited by classification accuracy, as the need of spatial generalisation to achieve class separability required suitably large objects. It was found that this dependence upon objects for spatial generalisation could be reduced through the incorporation of texture measures. Chapter 5 explores the scale potential of ultra-fine spatial resolution data. Field-level biomass modelling relies upon the construction of allometric models for the rapid estimation of biomass based upon easily measurable plant characteristics. Allometric modelling is regarded as the most accurate approach for estimating plant biomass, but its extension to remotely sensed data has been limited by data resolution. Coarser data resolution may limit or exclude the ability to measure the parameters required of plant allometric biomass models. The potential of ultra-fine resolution UAS data to measure allometric parameters is presented in Chapter 5, which is focused on fine-scale shrub biomass. Field derived allometric relationships are used to deconstruct shrub structure through image segmentation. Allometric parameters derived from the shrub components are then used to estimate biomass. This thesis demonstrates a methodology to develop and analyse UAS remotely sensed data, illustrating the scale potential of ultra-fine spatial resolution data. The increased complexity of fine-scale variability is a recognised problem associated with the improved resolving power of image data. This variability is a central challenge for UAS remote sensing and the analysis of the ultra-fine data scale it generates. By developing a clear methodology to construct and meaningfully disassemble ultra-fine resolution UAS data, this thesis provides a foundation which provides broader access to the novel scale niche that UAS measurements fill.