The remote sensing of papyrus vegetation (Cyperus papyrus L.) in swamp wetlands of South Africa.
Papyrus (Cyperus papyrus .L) swamp is the most species rich habitat that play vital hydrological, ecological, and economic roles in central tropical and western African wetlands. However, the existence of papyrus vegetation is endangered due to intensification of agricultural use and human encroachment. Techniques for modelling the distribution of papyrus swamps, quantity and quality are therefore critical for the rapid assessment and proactive management of papyrus vegetation. In this regard, remote sensing techniques provide rapid, potentially cheap, and relatively accurate strategies to accomplish this task. This study advocates the development of techniques based on hyperspectral remote sensing technology to accurately map and predict biomass of papyrus vegetation in a high mixed species environment of St Lucia- South Africa which has been overlooked in scientific research. Our approach was to investigate the potential of hyperspectral remote sensing at two levels of investigation: field level and airborne platform level. First, the study provides an overview of the current use of both multispectral and hyperspectral remote sensing techniques in mapping the quantity and the quality of wetland vegetation as well as the challenges and the need for further research. Second, the study explores whether papyrus can be discriminated from each one of its coexistence species (binary class). Our results showed that, at full canopy cover, papyrus vegetation can be accurately discriminated from its entire co-existing species using a new hierarchical method based on three integrated analysis levels and field spectrometry under natural field conditions. These positive results prompted the need to test the use of canopy hyperspectral data resampled to HYMAP resolution and two machine learning algorithms in identifying key spectral bands that allowed for better discrimination among papyrus and other co-existing species (n = 3) (multi-class classification). Results showed that the random forest algorithm (RF) simplified the process by identifying the minimum number of spectral bands that provided the best overall accuracies. Narrow band NDVI and SR-based vegetation indices calculated from hyperspectral data as well as some vegetation indices published in literature were investigated to test their potential in improving the classification accuracy of wetland plant species. The study also evaluated the robustness and reliability of RF as a variables selection method and as a classification algorithm in identifying key spectral bands that allowed for the successful classification of wetland species. Third, the focus was to upscale the results of field spectroscopy analysis to airborne hyperspectral sensor (AISA eagle) to discriminate papyrus and it co-existing species. The results indicated that specific wavelengths located in the visible, red-edge, and near-infrared region of the electromagnetic spectrum have the highest potential of discriminating papyrus from the other species. Finally, the study explored the ability of narrow NDVI-based vegetation indices calculated from hyperspectral data in predicting the green above ground biomass of papyrus. The results demonstrated that papyrus biomass can be modelled with relatively low error of estimates using a non-linear RF regression algorithm. This provided a basis for the algorithm to be used in mapping wetland biomass in highly complex environments. Overall, the study has demonstrated the potential of remote sensing techniques in discriminating papyrus swamps and its co-existing species as well as in predicting biomass. Compared to previous studies, the RF model applied in this study has proved to be a robust, accurate, and simple new method for variables selection, classification, and modelling of hyperspectral data. The results are important for establishing a baseline of the species distributions in South African swamp wetlands for future monitoring and control efforts.