Deep Learning based Framework for Semantic Segmentation of Satellite Images

Abstract

A sharp increase in the availability of satellite imagery data sets available in the recent past has made the explanation of all those data with a challenging problem. Retrieving useful information and insights from images taken by satellite or any other aerial imagery system requires a good understanding of the information present in the images itself. This work focuses on the design and implementation of an automated model to extract semantic maps of waterways, roadways, buildings to track urban cities growth of satellite imagery. As it is a machine learning problem, a deep neural network is implemented and evaluated experimentally. Open source, publicly available frameworks and datasets are used in this work. The result of this work is a pre-processing image model, which enables the user to take an input image with different quality and resolution and apply the semantic segmentation on it. U-Net is used as the underlying architecture for this project, because of many advantages and functionalities it provides, as an example, we can say less computation time, which can help starters in the field of AI and computer vision to build their prototype and models fast without the need of too much hardware (GPU, Processor, and RAM). To get a better result and decrease the amount of pressure on machine, FCN (Fully Connected Neural Network) is used along with U-Net. As FCN has skip layer function and it also makes the network to learn faster and give more accurate results than CNN due to normalizing the pooling layers which in CNN it makes the network to lose lots of information about the data. The primary purpose of this work is to implement Semantic Segmentation on satellite images taken from urban areas, SpaceNet dataset provides such a data. The data selected for this project is from Khartoum (Capital of Sudan), it is a developing and deserted city. The result of this work will help in managing resources such as agriculture, natural energy, and water, checking on natural disasters and responding to it, like floods, earthquakes, and tsunamis checking on environmental use such as deforestation and monitoring urban development.