Wang Yeqiao
Improving land cover classification and mapping accuracy has long been an important objective for remote sensing and spatial data investigators. It is generally recognized that improvements to classification can be made by coupling the information contained in multispectral, multitemporal, and multisource spatial data with human analyst’s expertise and methodology. A direct application of conventional statistically based classification approaches, however, may not be appropriate because of the differences in measurement scales and violation of distribution assumptions for multisource spatial data. Therefore, a distribution free and scale free classification scheme is desirable. Artificial Neural Network (ANN) technology is one of the optimal choices for this type of application. In this paper, two ANN models of backpropagation (BP) and modular ANN (MANN)were discussed and applied in land cover classification. Conceptually, ANN models have the potential to operate similarly to the way in which human image interpreters do. ANN features corresponding to the synapses, neurons, and axons of the brain are input weights, processing elements, and output paths. The BP model consists of three layers. The input layer represents the input data patterns, the output layer represents the classification scheme, and the hidden layer is responsible for information transformation and calculation. MANN model consists of several subnets or individual backpropagation networks (referred to as “local experts”). Each subnet is responsible for interpreting a subset of the input data. In multisource spatial data cases, while the unique spectral information from each season can be employed by individual subnets, spectral mixings among different land cover categories in single date remote sensing data can be filtered. MANN model provide the mechanism to improve the classification performance of simple BP model. Multisource spatial data combinations on (1) multispectral, multitemporal Landsat TM data, (2) multispectral, multitemporal Landsat TM and transportation data, and (3) Multispectral, multitemporal Landsat TM and illumination model data, were classified by the two ANN models, respectively. It concluded that both BP and MANN models performed well in high dimensional multisource spatial data classification. With modular and competition mechanism, the MANN model performed better than the BP model.