Earth observation (EO) involves the collection, analysis, and presentation of data in order to monitor and assess the status and changes in natural and built environments. This technology has many applications including weather forecasting, tracking biodiversity, measuring land-use change, monitoring and responding to natural disasters, managing natural resources, monitoring emerging diseases and health risks, and predicting, adapting to and mitigating climate change. This book shows how cutting-edge technologies such as artificial intelligence, including neural networks and deep learning, can be applied for processing satellite data for Earth observation. One of the objectives of this book is to explain how to develop a set of libraries for the implementation of artificial intelligence that could overcome some limits and encompass different aspects of research, ranging from data fusion to speckle filtering. In the first part, the authors introduce remote sensing concepts and deep neural networks and convolutional neural networks. In the second part of the book, they present the main tools used for image processing, several simulations and the data processing of specific case studies as well as the testing of related datasets. The book ends with conclusions, open questions and future works and perspectives for artificial intelligence techniques applied to future satellite missions. The book will be of interest to researchers focusing on using machine learning tools to process remote sensing data - particularly satellite data - for Earth observation. The book can also be used as a guide for researchers in many other fields of research who are interested in using ML techniques to process data and get reliable outcomes so they can make informed decisions for their specific objectives.
Inspec keywords: neural nets; learning (artificial intelligence); radar imaging; synthetic aperture radar
Other keywords: estimation theory; remote sensing; neural nets; radar imaging; filtering theory; mathematics computing; synthetic aperture radar; learning (artificial intelligence); artificial intelligence; speckle
Subjects: Data and information; acquisition, processing, storage and dissemination in geophysics; Education and training; General and management topics; Other topics in statistics; Geophysical techniques and equipment; Instrumentation and techniques for geophysical, hydrospheric and lower atmosphere research; Radar equipment, systems and applications; Computer vision and image processing techniques; Machine learning (artificial intelligence); Textbooks; Neural nets; General electrical engineering topics
This chapter introduces the topic on AI4EO (Artificial Intelligence for Earth Observation). Powerful trends in digital and sensing technologies are rapidly increasing, thereby also transforming the world of EO. In particular, extraordinary developments in information and communication technologies (ICT), including the Internet, cloud computing and AI, are giving rise to radically new ways of storing, distributing and analysing big data about our planet. This 'digital' revolution is also accompanied by a 'sensing' revolution that is delivering unprecedented amounts of data on the state of our planet and its changes. Europe is leading this sensing revolution in space through the Copernicus initiative and the corresponding development of the Sentinel missions, which monitor our planet on an operational and sustained basis. In addition, an emerging trend, referred to as New Space in the United States or Space 4.0 in Europe, is rapidly appearing through the expanding commoditisation and commercialisation of space. In particular, the increased capabilities and the rapidly declining costs of building and launching small satellites are allowing new EO actors - including start-ups, ICT giants and other kind of actors - to enter the space business. Consequently, innovative constellations of standardised small satellites are delivering new data on our planet with high spatial resolution and high temporal frequency.
This chapter introduces the principles underlying satellite remote sensing, highlighting the analysis and manipulation of data acquired by the related platforms. In particular, the satellites of the ESA (European Space Agency) Copernicus programme are taken as a reference for a detailed description of the chapter, because of the ESA policy to make Copernicus data freely available. As a consequence of this choice, several case studies will be presented in the other chapters of the book, where data acquired by the Sentinels of the Copernicus programme are mainly used.
The aim of this chapter is to introduce the reader to the concepts of artificial intelligence (AI), a branch of computer science attempting to build machines capable of intelligent behaviour, as well as its subdisciplines, machine learning (ML) and deep learning (DL). By highlighting the differences between ML and DL and tracing the steps that led to their developments, this chapter explores the ability of a machine to learn instead of being explicitly programmed. This chapter focuses on AI and its related disciplines.
In this chapter the authors present the theory and mathematics behind artificial neural networks (ANNs) as the starting point for understanding more complex networks that will be presented in the next chapters. This chapter also includes related concepts such as activation functions, gradient descent, optimisers, initialisers, normalisation and dropout, just to list a few.
In this chapter we introduce the convolutional neural network theory including concepts such as convolution operator, kernel, stride, padding and pooling.
In this chapter we describe how to create and manage a good satellite imagery dataset. We first focus on a classic way to get data, and then explain how to create a script to get it automatically, showing the main steps through code snippets and pseudo code blocks. Finally, a description of how to annotate this data will be given, thus generating the EO (Earth Observation) training datasets. Authors will mainly focus on Sentinel data from the Copernicus programme.
This chapter introduces the reader to the implementation and training of simple deep learning model. The chapter includes examples that have been developed in Python, using the Keras library build on top of TensorFlow.
In this chapter, the problem of image classification, target detection, and objects segmentation is addressed. The chapter is full of case studies, in which both the dataset implementation procedure and the neural network architecture are presented.
In this chapter we address the problem of generating satellite images, of a particular domain, using data of another domain. The description of the type of network used will be followed by a case study in which optical images are generated from synthetic aperture radar (SAR) data.
This chapter presents the problem of speckle filtering. After an introduction to the problem, the authors propose a case study on speckle filtering from Sentinel-1 images, using a Deep Learning approach.
In this wrap up chapter we present possible perspectives of using AI (Artificial Intelligence) on board future satellite missions. Conclusions are briefly given at the end. Some of our experiments will be presented and a first prototype introduced in this chapter. Moreover, considerations on the multidisciplinary aspects of the topics will be highlighted and other research fields listed, where the proposed AI and ML (Machine Learning) algorithms could be exploited.