Mapping invasive plants using RPAS and remote sensing

The ability to accurately detect invasive plant species is integral in their management, treatment, and removal. This study focused on developing and evaluating RPAS-based methods for detecting invasive plant species using image analysis and machine learning and was conducted in two stages. First, supervised classification to identify the invasive yellow flag iris (Iris pseudacorus) was performed in a wetland environment using high-resolution raw imagery captured with an uncalibrated visible-light camera. Colour-thresholding, template matching, and de-speckling prior to training a random forest classifier are explored in terms of their benefits towards improving the resulting classification of YFI plants within each image. The impacts of feature selection prior to training are also explored. Results from this work demonstrate the importance of performing image processing and it was found that the application of colour thresholding and de-speckling prior to classification by a random forest classifier trained to identify patches of YFI using spectral and textural features provided the best results. Second, orthomosaicks generated from multispectral imagery were used to detect and predict the relative abundance of spotted knapweed (Centaurea maculosa) in a heterogeneous grassland ecosystem. Relative abundance was categorized in qualitative classes and validated through field-based plant species inventories. The method developed for this work, termed metapixel-based image analysis, segments orthomosaicks into a grid of metapixels for which grey-level co-occurrence matrix (GLCM)-based statistics can be computed as descriptive features. Using RPAS-acquired multispectral imagery and plant species inventories performed on 1m2 quadrats, a random forest classifier was trained to predict the qualitative degree of spotted knapweed ground-cover within each metapixel. Analysis of the performance of metapixel-based image analysis in this study suggests that feature optimization and the use of GLCM-based texture features are of critical importance for achieving an accurate classification. Additional work to further test the generalizability of the detection methods developed is recommended prior to deployment across multiple sites.