Predictive modelling of wetland occurrence in KwaZulu-Natal, South Africa

  • Jens Hiestermann GeoTerraImage, Pretoria, South Africa
  • Nick Rivers-Moore Centre for Water Resources Research, University of KwaZuluNatal, Pietermaritzburg, South Africa
Keywords: ancillary data, Bayesian network, logistic regression, probability, wetland mapping


The global trend of transformation and loss of wetlands through conversion to other land uses has deleterious effects on surrounding ecosystems, and there is a resultant increasing need for the conservation and preservation of wetlands. Improved mapping of wetland locations is critical to achieving objective regional conservation goals, which depends on accurate spatial knowledge. Current approaches to mapping wetlands through the classification of satellite imagery typically under-represents actual wetland area; the importance of ancillary data in improving accuracy in mapping wetlands is therefore recognised. In this study, we compared two approaches – Bayesian networks and logistic regression – to predict the likelihood of wetland occurrence in KwaZulu-Natal, South Africa. Both approaches were developed using the same data set of environmental surrogate predictors. We compared and verified model outputs using an independent test data set, with analyses including receiver operating characteristic curves and area under the curve (AUC). Both models performed similarly (AUC>0.84), indicating the suitability of a likelihood approach for ancillary data for wetland mapping. Results indicated that high wetland probability areas in the final model outputs correlated well with known wetland systems and wetland-rich areas in KwaZulu-Natal. We conclude that predictive models have the potential to improve the accuracy of wetland mapping in South Africa by serving as valuable ancillary data.

  • Abstract 87
  • PDF 154
  • EPUB 27
  • XML 39
Views and downloads are with effect from 11 January 2018