KNOWLEDGE BASED SUB-PIXEL CLASSIFICATION FOR LAND COVER MAPPING

Abstract

Population expansion in urban areas has a direct impact on the use of land for transporta-tion, education, recreation, and places excessive demand on housing and existing infrastructure. It necessitates the urban planning for efficient use of land. Accurate land use land cover infor-mation is required for these types of projects. Remote sensing data are useful for the generation of these land use land cover maps, which can be achieved•by the process called image classifica-tion. In real world it is found that each pixel represents a heterogeneous area on the ground. For this reason it has been proposed that fuzziness should be accommodated in the classification procedure so that pixels may have multiple or partial class membership. To preserve the infor-mation present inside a pixel sub pixel classification was introduced. Among the most popular techniques are artificial neural networks, mixture modeling, Fuzzy c-means and Possibilistic c-means classification. This dissertation work presents the results of an experiment carried out to extract land use Iand cover classes of the two different study areas. The first study area is parts of Pune, Raigarh and Mumbai districts and second study area is a part of Calcutta. In this project, the sub pixel classification using fuzzy membership concept was applied for multispectral satellite images. To study the effort of varying across resolution, area of two land cover such as pond, res-ervoir and forest have extracted from MODIS, AWIFS, and LANDSAT data sets and relatively compared with LANDSAT, LISS-Ill, and QUICKBIRD data sets. ERDAS Imagine image analysis software from Leica Geo systems Geo-spatial imaging, LLC was used. In this dissertation the signature for classification is derived based on the component that is common to the training set pixels called the material of interest (MOI) using the fuzzy mem-bership functions rather than typically form a signature by combining the spectra of all training set pixels for a given feature. And the knowledge base is prepared using the ancillary data. The generalized decision tree is generated for the production of rules. An inference engine is devel-oped to classify the pixels furthermore and to improve the accuracy. The moderate resolution images are used as the data to classify and finer resolution imag-es are used as the reference data. The source image and reference image are classified using a soft classification and fuzzy error matrix is used to summarize the accuracy assessment infor-mation. With the usage of the knowledge base the land cover classes which are classified are increased and, it is observed that as the resolution increasing the accuracy of classification is also improved. Keywords: urban planning, land use land cover, sub pixel classification, fuzzy membership, spectral signature, material of interest, knowledge base; inference engine.