Road Segmentation in Aerial Imagery by Deep Neural Networks with 4-Channel Inputs

Roads are the essential modules in various applications. Road segmentation from aerial imagery is a significant research topic in the area of remote sensing. Extracting road information with high quality is very significant for many applications. Deep learning framework has been proved to be an effective approach in many image analysis and detail extraction techniques. In this paper, images with multiple colour channels are used to train the low complexity deep Res-UNet model architecture. Res-UNet model is trained with combination of boundary and Lovasz softmax loss functions. Experimental results on the public dataset indicates that proposed method with 4-channel inputs provides better results.