A Deep Learning Approach for Road Detection in Satellite Images Using U-Net

Road extraction from satellite photos is important because it greatly increases the efficiency of mapgeneration, which makes it potentially useful for emergency (rescue) systems and car navigation systems alike. Road network regions are identified by segmenting the picture, then properly identifying the roads using a decision-making and continuity approach, and lastly identifying line segments or roads using vectorization. There are many geospatial applications, such as routing software, infrastructure planning and mapping are depended on road network. The extraction rate of road networks has been greatly enhanced by automatic and semi-automated methods. This paper presents a deep learning model for road detection. The model uses U-Net architecture for image segmentation and extraction of roads. Publically available datasets are used to train the model. Various hyper parameters such as batch size, learning rate, and epochs are fines tuned and various optimizers are observed for their performance. When proposed model is compared to several well-known procedures, it is discovered that the strategies suggested thus far have significant gaps.