Multitemporal Analysis of Wetland Species using Multisensor UAV data - By Kevin Musungu

The Steenbras Nature Reserve near Cape Town, part of the Cape Floral Kingdom, contains seep fynbos wetlands that are both ecologically vital and vulnerable. To support their monitoring and conservation, this study developed a UAV-based workflow for mapping wetland species using multispectral imagery and machine learning. The research addressed four key questions: (1) how to optimize UAV multispectral data for species-level mapping, (2) which classifiers perform best for delineating fynbos species, (3) what the optimal season for species discrimination is, and (4) whether different UAV sensors can produce consistent results. Key vegetation indices included the Normalized Green-Red Difference Index (NGRDI), Normalized Difference Red Edge Index (NDRE), and Chlorophyll Index Red Edge (CIRE), as well as RG, Log Red, and Log Red Edge. Six classifiers consisting of Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbours (KNN), Naive Bayes (NB), Artificial Neural Networks (ANN), XGBoost, and Treebag were tested to classify twelve dominant wetland plant species. RF achieved the highest accuracy at 97%. The findings highlight that late winter is the best time of year for mapping these species, offering the clearest spectral separability among them. The study also showed that sensor characteristics significantly affect classification accuracy, even within the same season. These results contribute a robust and transferable workflow for high-resolution species-level monitoring in complex wetland ecosystems.