Browsing by Subject "Unmanned Aerial Vehicles"

Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Open Access
    Evaluating the suitability of UAV data for mapping dominant plant species in a heterogenous fynbos seep wetland
    (2025) Musungu, Kevin; Shoko, Moreblessings; Smit, Julian; Dube, Timothy
    Traditional methods for mapping plant species necessitate fieldwork and labour-intensive estimation of proportionate cover of the species under study. However, the inundated nature of wetlands makes fieldwork significantly difficult, costly, and prone to inaccurate estimations. In comparison, remote sensing technology offers a less resource-intensive approach, and multitemporal observations can enable species identification and monitoring. Among the remote sensing sensors, unmanned aerial vehicles (UAVs) have gained global prominence as affordable platforms for vegetation inventory studies. However, the potential use of UAVs for Fynbos wetland inventory has not been explored. This study used multispectral UAV photography from Parrot Sequoia and Micasense RedEdge- M multispectral cameras to discriminate eleven wetland Fynbos plant species in a seep wetland located in the Steenbras Nature Reserve in the Western Cape province of South Africa. The UAV multispectral data was gathered over six dates (August 2018, October 2018, December 2018, February 2019, April 2019, and February 2020) spanning three seasons and used to extract the multitemporal spectral signatures of the plant species. Then, critical spectral indices were identified based on the plant spectral signatures and ensemble feature selection. Of the twenty-seven indices assessed, the Visible Atmospherically Resistant Index (VARI), Modified Soil Adjusted Vegetation Index 2 (MSAVI2) and two indices developed in this study, namely, Red Green Vegetation Index (RG) and Log Red Edge (LogRed) were found to be pertinent for the classification. Three machine learning classifiers comprising Random Forest, K Nearest Neighbour, and Support Vector Machines were used to classify the different plant species across all dates using a dataset consisting of only spectral bands and another consisting of key spectral bands and indices. Classification accuracies improved when spectral indices were integrated with the spectral bands. Random Forest proved the most reliable, with generally better overall and per- class accuracies than the other machine learning classifiers. Lastly, the study assessed the effect of seasonal variability on the per-class performance of the machine learning classifiers and identified Spring as the optimum time of year for the classification of most of the plant species. This study highlights the potential of UAV data for inventory in heterogenous Fynbos wetlands.