Optimising multiresolution segmentation: delineating savannah vegetation boundaries in the Kruger National Park, South Africa, using Sentinel 2 MSI imagery

Image segmentation is useful for mapping vegetation communities since it aggregates pixels into homogenous zones. Multiresolution segmentation, a bottom-up multi-scale segmentation algorithm, is one of the most widely used and successful algorithms. The algorithm requires image band weights, scale parameter, shape, and compactness values to be specified prior to segmentation. The persistent challenge is how to identify these values, which are often determined by trial and error experimentation. This paper aimed at determining pre-segmentation image analyses that can inform the specification of the multiresolution segmentation parameter values, in order to optimise mapping of savannah vegetation community boundaries on high spatial resolution images. The vegetation boundaries in the 56 land types of the Kruger National Park (KNP), South Africa, were targeted for segmentation. Sentinel 2 Multi-Spectral Instrument (MSI) imagery acquired during the peak vegetation vigour period was used. The high (10 m) spatial resolution green, red, and near-infrared bands were selected for use. The KNP’s large size required a mosaic of successively acquired image frames. Multiresolution segmentation was performed using eCognition Developer 9. Pre-segmentation principal component analysis (PCA) revealed that none of the bands had superior data dimensionality. Therefore, equal band weights of 1 were specified. The original 16-bit top-of-atmosphere reflectance data showed that the vegetation communities had high within-community variance. Texture enhancement of the reflectance data using 3 × 3 kernel variance showed that the vegetation communities were more distinguishable by texture than the untransformed reflectance data. A heterogeneity-based scale parameter value of 388, and a high shape (‘texture’) value of 0.9 at the expense of the reflectance (‘colour’) value of 0.1, were specified. The scale parameter value was determined by averaging object coefficient of variation values on the texture image bands. Pre-segmentation fieldwork revealed transitional vegetation boundaries, necessitating the low compactness value of 0.1 (i.e. high smoothness value of 0.9). From the segmented objects, samples per respective vegetation community were specified for k-nearest neighbour (kNN) classification (k = 1). All of the 100 land type polygons were subsequently delineated, with overall mapping accuracy of 86.2%. Fuzzy membership of border objects successfully reproduced the transitional vegetation boundaries. The successful delineation of the savannah vegetation communities indicated that pre-segmentation PCA and analysis of potential objects’ variance-based texture can provide guidance on parameter values to specify for the inherently iterative multiresolution segmentation.