Browsing by Author "Luck-Vogel, Melanie"

Now showing 1 - 4 of 4
  • Thumbnail Image
    Item
    Assessment of accuracy : systematic reduction of training points for maximum likelihood classification and mixture discriminant analysis (Gaussian and t-distribution)
    (CONSAS Conference, 2018-09) Ritchie, Michaela; Debba, Pravesh; Luck-Vogel, Melanie; Goodall, Victoria
    Remote sensing provides a valuable tool for monitoring land cover across large areas of land. A simple yet popular method for land cover classification is Maximum Likelihood Classification (MLC), which assumes a single normal distribution of the samples per class in the feature space. Mixture Discriminant Analysis (MDA) is a natural extension of MLC which can be used with varying distributions and multiple distributions per class, which simplifies the classification process tremendously. We compare the accuracies of MLC and MDA (using a Gaussian and t-distribution) as the number of training points are systematically reduced in order to simulate varying reference data availability conditions. The results show that the more robust t-distribution MDA performs comparatively with the Gaussian MDA and that both outperform MLC when sufficient training points are available. As the number of training points increases the MDA accuracies increase while the MLC accuracy stagnates. At very low numbers of training samples (ranging from 22 to 169 dependent on the class), there is more variability in terms of which method performs best.
  • Thumbnail Image
    Item
    National coastal assessment & coastal climate change vulnerability assessment: Implications for the future
    (2019-09) Luck-Vogel, Melanie
    In this presentation the legacy of the coastal flood and erosion risk assessments conducted from the National Coastal Assessment project and the Coastal Climate Change Vulnerability Assessment is explained and the way forward is lined out.
  • Thumbnail Image
    Item
    Using remote sensing in support of environmental management : a framework for selecting products, algorithms and methods
    (Elsevier, 2016-08) De Klerk, Helen Margaret; Gilbertson, Jason; Luck-Vogel, Melanie; Kemp, Jaco; Munch, Zahn
    Traditionally, to map environmental features using remote sensing, practitioners will use training data to develop models on various satellite data sets using a number of classification approaches and use test data to select a single ‘best performer’ from which the final map is made. We use a combination of an omission/commission plot to evaluate various results and compile a probability map based on consistently strong performing models across a range of standard accuracy measures. We suggest that this easy-to-use approach can be applied in any study using remote sensing to map natural features for management action. We demonstrate this approach using optical remote sensing products of different spatial and spectral resolution to map the endemic and threatened flora of quartz patches in the Knersvlakte, South Africa. Quartz patches can be mapped using either SPOT 5 (used due to its relatively fine spatial resolution) or Landsat8 imagery (used because it is freely accessible and has higher spectral resolution). Of the variety of classification algorithms available, we tested maximum likelihood and support vector machine, and applied these to raw spectral data, the first three PCA summaries of the data, and the standard normalised difference vegetation index.We found that there is no ‘one size fits all’ solution to the choice of a ‘best fit’ model (i.e. combination of classification algorithm or data sets), which is in agreement with the literature that classifier performance will vary with data properties.We feel this lends support to our suggestion that rather than the identification of a ‘single best’ model and a map based on this result alone, a probability map based on the range of consistently top performing models provides a rigorous solution to environmental mapping.
  • Thumbnail Image
    Item
    Vegetation mapping in the St Lucia estuary using very high-resolution multispectral imagery and LiDAR
    (Elsevier, 2016-05) Luck-Vogel, Melanie; Mbolambi, C.; Rautenbach, K.; Adams, J.; van Niekerk, L.
    This paper examines the value of very high-resolution multispectral satellite imagery and LiDAR-derived digital elevation information for classifying estuarine vegetation types. Satellite images used are fromtheWorldView-2, RapidEye, and SPOT-6 sensors in 2mand 5mresolution, respectively, acquired between 2010 and 2014. Ground truthing reference is a GIS-derived vegetation map based on field data from 2008. Supervised maximum likelihood classification produced satisfactory overall accuracies between 64.3% and 77.9% for the SPOT-6 and the WorldView-2 image, respectively,while the RapidEye-based classifications produced overall accuracies between 55.0% and 66.8%. The reasons for the misclassifications are mainly based on the highly dynamic environmental conditions causing discrepancies between the field data and satellite acquisition dates rather than technical issues. Dynamics in water levels and salinity caused rapid change in vegetation communities. Further, weather impacts such as floods and wind events caused water turbidity and led to bias in the reflective properties of the satellite images and thus misclassifications. These results show, however, that the spatial and spectral resolution of modern very high-resolution imagery is sufficient to satisfactory map estuarine vegetation and to monitor small-scale change. They emphasise, however, the importance of synchronisation of ground truthing data with actual image acquisition dates in these highly dynamic environments in order to achieve high classification accuracies. The results also highlight the importance of ancillary data for accurate interpretation of observed classification discrepancies and vegetation dynamics.