Accuracy assessment between different image classification algorithms using LandSat-7 of Abuja

Abstract

What image classification does is to assign pixel to a particular land cover and land use type that has the most similar spectral signature. However, there are possibilities that different methods or algorithms of image classification of the same data set could produce appreciable variant results in the sizes, shapes and areas of the classified polygons, which is capable of misleading the user. This study attempts a comparative analysis of four (4) image classification methods (maximum likelihood, parallepiped minimum distance and fisher’s methods) to experiment which method best classifies the 2001 Landsat-7 ETM 345 imagery of part of Abuja Nigeria. The results of the classifications were displayed in qualitative palette of Idrisi-32 and it was discovered that, each cover type which includes, closed canopy vegetation, water, open vegetation, urban built-up, rock outcrops and bare ground were best classified by the maximum likelihood method, followed by the minimum distance method. This conclusion was made based on the prior knowledge of the land cover and land use characteristics in the study area. It is recommended that, this experimental approach be used for different data set from various sensor platforms to provide a standard guideline for specific applications by users in different parts of the world, particularly in Africa where data and the relevant software are often scarce.