Application of neural networks for wetland classification in RADARSAT SAR imagery

The purpose of this study was to evaluate the ability of backpropagation neural networks to delineate forested and open wetlands and to distinguish between wetland categories using RADARSAT SAR data. To accomplish this objective, a multi-temporal dataset of RADARSAT images was used to evaluate the utility of the neural network approach for monitoring wetland vegetation communities and to detect seasonal changes in the Lac Saint-Jean region (Quebec, Canada). In order to accomplish this task, several parameters must be supplied, including the number of hidden nodes, learning, training, and ancillary data, such as textural information. To improve the neural classification performance, several techniques have been tested. In this way, a new methodology is developed for selection of training data sets and development of neural network structure. The advantages of neural networks for extracting information from radar backscattered energy are discussed with respect to classification accuracy