A Comparison of Convolutional Neural Network Architectures for Road Classification from Satellite Images

Abstract

The classification of roads from a satellite image has a scope for high level research due to the variations in topology. So it is quite challenging to classify roads from satellite imagery in a realistic manner. The importance and utility of classifying roads from a satellite image having high resolution can help in the navigation of roads, revision of geographic information systems, in the area of emergency rescue applications and so on. The problem that arises is to identify and differentiate a road segment from its background. The difficulty for classifying a road is the existence of trees, manmade constructions like buildings particularly in an urban landscape. The different types of sensors used in satellites can also have an impact on identifying roads due to its variations. What is required is a fast and accurate method for extraction of roads. This proposed work deals with extraction of a road network from satellite images having high resolution. This work focuses on estimating a particular pixel in the image is an element of a road or not using an enhanced Convolutional Neural Network (CNN) approach. The advent of various frameworks of CNN has made this work realistic. The work proposes a new approach for making data sets for this complex problem and has concluded with a feasible solution for the problem