Transcoding-based pre-training of semantic segmentation networks for PolSAR images

Pre-training deep neural networks uses a proxy task to learn representative features that are transferable to a different problem. We investigate how semantic segmentation networks for PolSAR images benefit from pre-training on a transcoding task which translates PolSAR data into optical images. We compare multiple approaches ranging from a simple regression network up to a cycle-GAN. All pre-training methods lead to significant gains in classification accuracy. Surprisingly, the cycle-GAN as the most sophisticated architecture evaluated here leads to the worst results. The best results are achieved by the conditional GAN but closely followed by the much simpler regression network.