Adaptive edge enhancement in SAR images training on the data vs. training on simulated data

Edge detection and edge enhancement in SAR images is, due to the speckle effect, not so easily achieved. Here we consider edge enhancement as a classification problem, i.e. we segment an image in several edge classes and a no edge class. Thus supervised classification techniques become available. We propose an artificial neural network approach and interpret the output as a 'grade of being an edge pixel'. For training of the network we applied two training strategies: (1) selection of training samples from the data in a supervised way and (2) artificial creation of training samples based on speckle statistics using a speckle simulation algorithm. Both strategies are applied on a data set acquired by DLR's E-SAR in the L-band. The outputs of the edge enhancement process are compared among each other and with the RoA edge detector.