Evaluating the use of remote-sensing to locate weeds in Kosciuszko National Park

The spread of invasive species in sensitive ecosystems is a major environmental problem, resulting in significant hazards to local and regional environmental and socioeconomic facets. Traditional field-survey based management, particularly in adverse terrain, is difficult – inhibited by climactic conditions, worker safety, and time. There is a strong demand for remote-sensing to mitigate these issues, improving efficiency and effectiveness, and increasing coverage ability. The aims of this thesis are to investigate the potential use of remote-sensing in weed management and specifically evaluate the use of it to two locate weed species in Kosciuszko National Park (KNP). The evaluation was performed through the collection and processing of floral spectra to formulate a spectral library, which was then analysed for statistical uniqueness, utilising machine-learning classification algorithms. Results of the Random Forest (RF) discriminability analysis presented an overall separability accuracy of 70%. This was then resampled to simulate the discriminability for commercial drone and multi-spectral satellite sensors, giving accuracies of 59% and 63%, respectively. This preliminary analysis is promising, and builds the foundation for future multispectral research for weed management in KNP, and provides a foundational methodology for spectral pre-assessment in general.