Sensor Data Analysis and Management

9 Object Detection in Satellite Images Using Modified Pyramid Scene Parsing Networks

Akhilesh Vikas Kakade¹, S Rajkumar² (Corresponding Author), K Suganthi³, and L Ramanathan⁴

¹ SAP Labs, Bangalore2,3,4 Vellore Institute of Technology, Vellore

9.1 Introduction

The phenomenal growth in the variety, the improved accessibility, and the global availability of satellite imagery have resulted in dramatic improvements in our understanding of planet Earth. Such an understanding is required in situations ranging from emergency operations of activating resources during disasters to routine processes like observing the impacts of global warming. However, there is still a great limitation to these developments. The limitation is the assumption that detecting features or objects of interest in satellite images can be easily done either manually or with partial help of computers, that is, semi-automatically. On the one hand, this assumption puts a tremendous burden on the experts who are responsible for detecting and identifying such objects of interest. On the other hand, there have been spectacular improvements in processing capacities of the processing units and great advancements in computer vision with help of machine learning technologies such as deep learning through deep neural networks. It is then natural to think about utilizing the hardware and logarithmic advancements in automating important or significant objects in satellite images. This identification, if automatic, accurate, ...