An investigation into mapping wetlands using satellite imagery : the case of Midmar sub-catchment.
A suitable methodology for mapping wetlands in South Africa has not been agreed upon. This investigation aimed at developing a methodology for the accurate and efficient delineation of wetland areas using satellite imagery and other relevant spatial datasets. Both summer and winter LANDSAT ETM+ satellite imagery covering the study area of the Midmar sub-catchment were processed using various image classification techniques. These included the supervised, unsupervised and level slicing classifications. The accuracy of each technique was tested against the only existing verified wetland dataset that covers the study area. A ground truthing exercise was also undertaken. The different classification techniques resulted in different classification accuracies when compared to the verified wetland dataset. Accuracies for the different classification techniques were as follows: unsupervised 20 class classification (summer) 55%, (winter) 39%, unsupervised 255 class classification (summer) 71%, (winter) 47%; supervised classification (summer) 65%, (winter) 41%; level slicing classification (summer) 65%, (winter) 45%. The inaccuracies could mostly be attributed to a change in land cover as there seems to be an overall loss of wetland areas. However, the ground truthing exercise resulted in higher classification accuracies especially with unsupervised 255 class classification. This study concluded that LANDSAT ETM+ satellite imagery was useful for detecting wetlands areas during summer by using a fine classification technique (255 class). A finer classification technique is also suited for the detection of both large and small wetland areas. Major recommendations include: the use of summer imagery in a high rainfall period; the unsuitability of using winter imagery due to the spectral confusions created; the use of high resolution satellite sensors (SPOT) for monitoring purposes while lower resolution sensors (LANDSAT) should be used for mapping; the increased use of topographical modelling for wetland detection; the use of an appropriate scaled land cover database and the use of field verification exercises for comparing classifications.