| Wetland conservation and management requires inventory and monitoring of wetlands and their adjacent uplands. Satellite remote sensing has several advantages for monitoring wetland resources, especially for large geographic areas. This dissertation summarizes the literature on satellite remote sensing of wetlands, including what classification techniques were most successful in identifying wetlands and separating them from other land cover types. Then the use of different features and classification methods were evaluated for their effect on classification accuracy in two large study areas with heterogeneous landcover types, the Twin Cities Metropolitan Area in Minnesota, USA (TCMA) and the Kizilirmak Delta on the Black Sea coast of Turkey. The features examined were all derived from the Landsat MSS or TM imagery: texture, landscape metrics, and indices. The classification methods were conventional per pixel maximum likelihood, three contextual techniques, maximum likelihood with 3 x 3 majority filter, ECHO, and a hybrid segmention method, and one artificial neural network technique, the feedforward multilayer perceptron with backpropagation for training. The use of multitemporal Landsat TM imagery, June, September, and April image dates, and indices were the set of features that gave the highest classification accuracy for the TCMA. Tests with indices highlighted some opportunities for improving wetland classification. For the Kizilirmak Delta, texture features in addition to the MSS bands, gave the highest classification accuracy. For the TCMA, the hybrid segmentation method generally gave the highest overall classification accuracy, followed by followed by the per pixel maximum likelihood with 3 x 3 majority filter, and the ECHO classifier. For the Kizilirmak Delta, the ANN classifier had the best performance. The ANN classifier generally had equal or better classification accuracy than the maximum likelihood classifier, for the full training set and also for a minimal training set, indicating that ANN classifiers do not necessarily require more training data than conventional methods. Varying the number of hidden nodes in the one hidden layer did not have a large effect on classification accuracy. However, composition of the training data set had a large impact on the performance of the ANN classifiers as did the number of epochs for training. |
