Accuracy of neural network classifiers in humid tropical areas / by Genong Yu.

Tese (Ph.D.)--Indiana State University, 2004.

This research develops a better understanding of the use of neural networks in the context of tropical land cover characterization using multiple scale remotely sensed data. Results of the research contributed to developing a robust neural network classifier suitable to use for humid tropical land cover classification in areas such as Altamira, Brazil. Research was focused upon the following: (1) performance of neural networks with different ancillary data - such as DEM (digital elevation model) and its derivatives; (2) effect of spatial scale on neural network classifiers using remote sensing data from IKONOS, Landsat TM, ASTER, and MODIS; (3) accuracy of neural networks in extracting linear features from remotely sensed data; and (4) examination of different approaches in combining multiple classifiers. Comparative studies at the level of concept principles, algorithm, implementation, and experiments between neural network classification approaches and conventional methods were conducted. An object-oriented design paradigm with C# programming language was used to implement all the analytical procedures, among which were multiple layered perceptron (MLP) and Fuzzy ARTMAP that represented neural network classifiers. For comparison with neural networks, conventional image analytical procedures were implemented, such as the maximum likelihood classifier (MLC) for classification, multiple linear regression and multiple logistic regression for regression, and Canny edge detector and Hough transform for linear feature extraction. Various compound classifiers based on different combining strategies were constructed, including stacked generation schemes, voting ensemble, and Bayesian ensemble. Experiments were carried out to test the different levels of error sensitivity in these compound classification systems in order to select the combining strategy that delivers the best performance. In addressing the four primary research questions identified above, the following were discovered: (1) the MLP had the best performance among the experimental classifiers when ancillary data were used; (2) scale problems and aggregation problems were identified and illustrated by a series of upscaled image pyramids; (3) the supervised neural network linear feature detector was more robust than the conventional edge detection and spectral classification that combined local spatial information and spectral information; and (4) the MLP, when used within a combined context of multiscale image classification to generalize classification output, had good results in most cases with up to 2-20% improvement of Kappa accuracy on difference scales compared to those of maximum likelihood classifier (MLC), although training was slow. In terms of combining strategies, the Bayesian ensemble using sum rule was selected as the optimal combining strategy to be used with the multiscale context-based classifier. The major contribution of this research was the development of a new classifier, the multiscale context-based classifier (MCC), based on the hybrid of scale-theory, neural networks, mathematical morphology, and image understanding. This new classifier, in comparison to traditional approaches, achieved improved overall classification accuracy of five major land cover classes found in a humid tropical study area near Altamira, Brazil.