Evaluating Neural Networks and Evidence Pooling for Land Cover Mapping

The diversity of data sources, analysis methodologies, and classification systems has led to a number of new techniques for monitoring land-cover change. However, this wide choice means that it is difficult to know which solution to choose. A system capable of integrating the results of different analyses and applying them to land-cover mapping would therefore be extremely useful. This study investigates the use of evidence pooling and neural networks in land-cover mapping. Neural networks were used to classify land-cover using evidence from spectral (Landsat-7 ETM1), textural, and topographic information. Mapping was performed using combinations of evidence source and evidence pooling techniques. The best performance was achieved using all available information with a method that summed evidence directly instead of categorizing it. While the methodology failed to reach the level of accuracy recommended elsewhere, a comparison of the number of classes used with other methods showed that the system performed better than these approaches.