Comparison of land cover image classification methods

Abstract

In remote sensing, many methods have been developed for image classification. In this study, three of the methods namely Maximum Likelihood classification (MLC), Backpropagation Neural Network classification (BPNN), and Sub Pixel classification (SP) are used to classify a Landsat ETM+ image of the Ejisu-Juabeng district of Ghana into seven land cover classes and the results are compared. The seven classes identified were forest, forested wetland, open woodland, water, non-forested wetland, grassland and urban. In the comparison, the top 20 (80%-100% composition) per land cover class from the SP is used against the MLC and BPNN classification. The results show that of the two hard classifications (MLC & BPNN), BPNN gave a better result with an overall accuracy of 92.5 % compared with that of MLC with an accuracy of 78.95%. The SPclassification however, gavemixed results although forland cover classessuch asforest and forested wetland that are homogeneousin nature,the representationsweregood.Over alltheBPNNclassificationgave thebestrepresentationofthe landcover classesinthe studyarea