Title

Modularizing backpropagation neural networks for multisource spatial data modeling and classification

Date of Completion

January 1995

Keywords

Physical Geography|Agriculture, Forestry and Wildlife|Remote Sensing|Artificial Intelligence

Degree

Ph.D.

Abstract

Applications of artificial neural networks (ANN) in remote sensing and multisource spatial data classification have been frequently reported in the past several years. In the previous research, backpropagation ANN (BPANN) has commonly been applied. This popularity primarily revolves around the ability of backpropagation paradigm to learn complicated multidimensional mapping. Other ANN paradigms, however, are seldom applied and reported. Technically, the demand of development of efficient ANN architectures for handling multisource spatial data still remains. In multisource spatial data land cover classification, while more information can be supplied by those data, noise, redundancy and confusion may also be introduced. If artificial neural network paradigms with modular functions or competition mechanisms can be developed, the information process for each data source will be decomposed, and the contribution of each data set will be separately evaluated. Therefore, the advantages from each data source will be discovered and efficiently employed to perform the classification. To reach this goal, other architectures of neural network paradigms, such as modular artificial neural network (MANN) and learning vector quantization (LVQ) network, are alternatives requiring further investigation.^ In this research, three network paradigms of BPANN, MANN, and LVQ were developed and evaluated in multisource spatial data land cover classifications. Traditional maximum likelihood classification (MLC) was also compared in classification performance. Multisource spatial data in different combinations were classified by the three ANN paradigms, respectively. Comparable land cover classification results were achieved by BPANN and MLC. Both MANN and LVQ networks achieved better classification results than MLC. It is concluded that all three artificial neural network paradigms can be applied in high dimensional multispectral, multitemporal, multisource spatial data classification. With a modular design and competition mechanism, the MANN and LVQ networks performed well and are the reliable alternatives in multisource spatial data land cover classification. ^