Road Segmentation in SAR Satellite Images with Deep Fully-Convolutional Neural Networks

Remote sensing is extensively used in cartography. As transportation networks expand, extracting roads automatically from satellite images is crucial to keep maps up-to-date. Synthetic Aperture Radar (SAR) satellites can provide high resolution topographical maps. However roads are difficult to identify in SAR images as they look visually similar to other objects like rivers and railways. Deep convolutional neural networks have been very successful in object segmentation, yet no method was developed to extract entire road networks from SAR images. This letter proposes a method based on a Fully-Convolutional Neural Network (FCNN) adjusted for road segmentation in SAR images. We study two approaches, binary segmentation and regression, intolerant and tolerant to prediction errors, respectively. The segmentation consistency is improved by applying Fully-connected Conditional Random Fields (FCRFs). We also share insights on creating a suitable dataset to facilitate future research. Our FCNN model shows promising results, successfully extracting 57% of the roads in our test dataset. We find out that the erosion effect of the FCRFs can effectively remove incoherent predictions, but is detrimental to road interconnections. The predicted roads have smooth borders yet oscillating shapes, hence regularization would help improving their straightness and connectivity.