An integrated classification approach for remote sensing data incorporating fuzzy neural networks, GIS and GPS

Although satellite remote sensing techniques are potentially quite valuable in monitoring Earth's resources over the large area, the utility of traditional systems has been limited by a number of factors. Low classification accuracy is among these factors.;Low classification accuracy can be due to factors beyond the scope of this thesis, such as atmospheric effects. However, this is mainly the result of the fact that landcover classification methodologies still depend largely on traditional techniques using only spectral information. This traditional approach suffers from many problems, including an information loss problem, lack of incorporation of additional data and information, and those problems associated with only using a single classifier. In order to handle these problems simultaneously and to improve classification accuracy, an integrated classification approach incorporating a fuzzy neural network and modified maximum likelihood classifier, along with Geographic Information System and Global Positioning System technologies is proposed in this thesis.;The tests of the proposed classification approach are performed in two different ecological regions in Wisconsin. By the results of rigorous statistical tests, the overall classification accuracy of the proposed approach is improved significantly over that of the traditional classification approach in both study areas. In addition to this conclusion, the information loss problem is effectively solved by using fuzzy pixel representation and the process of combining two classifiers in the new approach. The problem of the lack of incorporation of ancillary data is handled appropriately by incorporating a GIS layer directly into the fuzzy neural network classifier. With this approach, spectrally inseparable classes can be discriminated correctly and, thus, classification accuracy is improved significantly. Combining two classifiers using a gating network is effective in solving the problem of the inability to use multiple classifiers in the traditional classification approach. This combined approach improves classification accuracy over either of the individual classifiers alone. Therefore, we can conclude that an improvement of overall classification accuracy can be accomplished by effectively resolving problems inherent in the traditional classification approach.