Context: SAR images can be acquired all time (day and night and whatever the weather) which is a real advantage for tropical areas. Nevertheless the signal is affected by speckle noise which induces strong fluctuations. Hydrological networks usually appear as dark linear features on these images since they are rather smooth compared to the wavelength.
Objective: This project is dedicated to hydrological network detection on SAR images. The idea is to train a network on similar images of retina blood vessels. Indeed the shape and structure of the blood vessels are very close to the ones of hydrological networks. The U-net network will be trained with a modified datasets including speckle noise and then tested on real SAR images. The following steps have to be addressed: database modification to include speckle noise; training of the network (with or without transfer learning); evaluation on simulated and real SAR images.
Supervision: Florence Tupin, Nicolas Gasnier
https://github.com/orobix/retina-unet
