Fundamentals and Methods of Machine and Deep Learning

18Case Studies: Deep Learning in Remote Sensing

Emily Jenifer A. and Sudha N.*

SASTRA Deemed University, Thanjavur, India

Abstract

Interpreting the data captured by earth observation satellites in the context of remote sensing is a recent and interesting application of deep learning. The satellite data may be one or more of the following: (i) a synthetic aperture radar image, (ii) a panchromatic image with high spatial resolution, (iii) a multispectral image with good spectral resolution, and (iv) hyperspectral data with high spatial and spectral resolution. Traditional approaches involving standard image processing techniques have limitations in processing huge volume of remote sensing data with high resolution images. Machine learning has become a powerful alternative for processing such data. With the advent of GPU, the computation power has increased several folds which, in turn, support training deep neural networks with several layers. This chapter presents the different deep learning networks applied to remote sensed image processing. While individual deep networks have shown promise for remote sensing, there are scenarios where combining two networks would be valuable. Of late, hybrid deep neural network architecture for processing multi-sensor data has been given interest to improve performance. This chapter will detail a few hybrid architectures as well.

Keywords: Deep learning, remote sensing, neural network, satellite image, fusion

18.1 Introduction

Remote sensing ...