Mapping rural savanna woodlands in malawi: a comparison of maximum likelihood and fuzzy classifiers.

Changes in land cover system represent a key variable in managing and understanding the environment, as well as driving many environmental assessment mechanisms such as hydrological models for large river basins water budgeting. Remote sensing can provide information on the spatial pattern of land cover features, but analysis and classification of such imagery primarily suffers from the problem of class mixing within pixels. To reflect the actual land cover conditions rigorously and well defined, statistical algorithms have to ´bridge the gap´ between legend requirements and the input satellite imagery. While studies have been done using maximum likelihood and fuzzy classifiers in forestry, urban planning and savannah woodlands, appropriate methods to map land cover distributions in savanna woodlands associated with rural settlements are yet scarce. The distribution of savanna woodlands, rural residential areas (especially grass-thatched housing) and cultivated/grazing areas within the Shire River catchment in Malawi, represent classes which have similar spectral signatures (especially during the dry season). They occur in similar environments and are often in adjacent or mixed stands. Two classification methods i.e. purely using a maximum likelihood classification and when improving this classification using a contextual fuzzy convolution filter were assessed to map land cover dynamics of the Shire River catchment using Landsat 7 ETM+. With respect to classification methodologies and the ability to correctly identify land cover features, accuracies (before and after applying the filter) were compared and tested for the catchment´s hydrological modelling. Spatial characteristics of the catchment, digital elevation data, precipitation and the Landsat mapped land cover data were derived and exported into a geographic information system (GIS) to provide thematic data layers from which to delineate hydrologic response units (HRUs). Eight detailed land cover classes w- ere mapped for the Shire River catchment. The hierarchical legend structure determined by the Food and Agriculture Organization (FAO) Land Cover Classification System (LCCS) was used to label land cover variables. The purely maximum likelihood statistical classifier accurately mapped individual classes in more detail which could not be discriminated using fuzzy convolution filter. The spatial scale for land cover parameterization can play a significant role in how specific land surface hydrological processes are simulated.