AUTOMATIC ROAD DELINEATION USING DEEP NEURAL NETWORK

Road extraction from high resolution satellite imagery has been a challenging task. The problem has been attempted by many people employing different methods and techniques and many have been able to solve it to a large extent. The novelty of this paper is to reach the end goal of providing a final product which can be used to generate semantically meaningful applications like vehicle detection, vehicle counting and determining the size of vehicle on the road. In this paper, an approach of road delineation in high resolution multi-spectral satellite imagery is proposed using Deep Neural Networks to generate a road binary mask. The binary mask comprising of objects is further processed with image processing techniques. Whereas to reduce the non-road objects, which are classified as road, object attributes such as object size and shape are used. The refined objects are converted into a shape file of road. Various challenges faced along the way and some useful observations and algorithmic strategies to achieve the end goal have been discussed in this paper.