Knowledge based multi-source, time series classification: A case study of central region of Kenya

Abstract

Land use Land cover (LULC) time series mapping for large areas have greatly benefited from the availability of medium multispectral resolution imagery. While medium and low resolution data is greatly affordable, it presents some challenges during its classification such as defining discrete land cover classes, selecting adequate training areas and the mixed pixels. This research applied knowledge based classification with variables of Principal components, digital elevation model, Normalised Difference Vegetation Index (NDVI) and slope in order to overcome the problems of defining discrete land cover classes and selecting adequate training areas with Landsat imagery. The first three Principal components of data epochs 1995, 2002, and 2010 were investigated through factor loading and histogram density slicing to develop an optimum threshold values to differentiate among the following classes: clear water, turbid water, salty muddy water, rocks, dense forest, light dense forest, grass, bare soils, silts/sand rocks and crop covers. NDVI, Digital Elevation Model (DEM) and slope were also incorporated to differentiate among vegetation covers and map water covers. The overall accuracies obtained were 89.6%, 88.8% and 87.8% and kappa coefficients of 0.88, 0.87 and 0.86 for the years 1995, 2002 and 2010 respectively. The change detection analysis showed competing land uses in forest cover and crop versus grass lands and bare lands. Water cover remained almost unchanged with changes of less than 0.1%.