| Land cover maps are one of the primary kinds of geospatial information provided by remote sensing. Land cover classification is essential for terrestrial ecosystem modeling and monitoring, as well as climate modeling and prediction. During the last decade, artificial neural network (ANN) classifiers have been used increasingly in land cover classification and detection of land cover change using remote sensing data. However, understanding of the behavior and characteristics of ANNs lags behind that of conventional techniques such as the maximum likelihood classifier. The focus of this dissertation is improved use of ANNs, particularly ARTMAP, for land cover mapping using remote sensing. Four topics are pursued.; First, an ARTMAP classifier increases the accessibility of ARTMAP to users in remote sensing, and a new set of visualization tools aids interpretation of the internal dynamics of ARTMAP. Second, a hybrid classification approach uses two nonparametric classifiers, ARTMAP and decision trees, to produce a spatially explicit uncertainty metric which can be provided with thematic maps. Validation of this approach using two data sets demonstrates that classification accuracy is strongly related to confidence levels. Third, a new pruning technique combines prediction accuracy and instance counting. This pruning algorithm leads to 2 to 5% improvement in classification accuracy in tests using three Landsat TM data sets. The pruning technique also reduces the category proliferation problem in ARTMAP. Fourth, a new ARTMAP model for multisensor image fusion estimates subpixel land cover proportions from coarser resolution images (MODIS) based on training from finer resolution images (Landsat TM). This approach builds multiscale representations of land cover such that diverse processes can be examined at appropriate spatial scales. ARTMAP consistently performs better than conventional linear mixture models for estimating subpixel fractions. The overall benefits of the research are greater accessibility to the ARTMAP architecture for remote sensing, improvements in ARTMAP via a new pruning algorithm, development of spatial uncertainty measures based on a hybrid classification scheme and an ARTMAP model for multisensor fusion. |
