Shevtsova O. Information technology for pre-processing and classification of multi-time satellite images of high spatial resolution

The dissertation research solved an important scientific and applied problem of increasing the accuracy, level of automation and speed of recognition and classification of images of high spatial fragmentation by developing information technologies and processing using machine learning. The dissertation consists of an introduction, four chapters, a conclusion, a list of sources used and appendices. The full volume of the dissertation is 165 pages; list of used sources of 136 titles, 2 appendices. The work is illustrated with 40 drawings and contains 12 tables. In the introduction, the relevance of the topic is substantiated, the research objectives and tasks are formulated, the research methods are defined, and a general description of the work and the structure of the dissertation are given. The scientific novelty, practical significance of the conducted research, and the author's personal contribution are determined, the reliability of the obtained results is assessed, and information about publications and the results of the approbation and implementation of the work is provided. The first chapter examines the current state of the development of methods and technologies for the preliminary processing and classification of high spatial resolution satellite data, as well as the existing mathematical and software tools used for developing such technologies. Methods based on artificial intelligence, modern deep learning approaches, and the latest trends in the development of artificial neural networks are analyzed. An analysis of the current state of the problem of preliminary processing and classification of high spatial resolution Earth remote sensing images and a review of the literature revealed the inefficiency of existing methods, leading to numerous classification errors, and identified several unresolved issues, particularly the lack of automated information technologies that allow real-time classification of multispectral, multitemporal high spatial resolution satellite images. Existing solutions are generally costly, labor-intensive, and involve manual decryption. Conclusions and the formulation of the task regarding the need to develop a comprehensive information technology for the preliminary processing, analysis, and classification of satellite images are presented. In the second chapter, an analysis of the problem and modern approaches to processing extensive Earth remote sensing data is carried out. Basic concepts and definitions used for its solution are provided. The properties of the data organization are described, allowing to predict the result of performing certain operations in the structure on its elements, using their arrangement without performing a computational algorithm. Conclusions about the influence of properties and methods of working with the structure were obtained. The effectiveness of the method of optimizing the main characteristics of big data processing based on the application of a tuple structure of data organization, which allows to reduce the amount of processed information, to increase the speed of data search and processing while preserving their respective values and reliability, is proposed and proven. The effectiveness of an IaaS solution for remote sensing data flow processing based on deep learning and Kubernetes and Apache Airflow cloud technologies, hosted on the Google Cloud Platform, is proposed and proven. The proposed algorithm is presented in the form of a directed acyclic graph in an IaaS application. The mentioned cloud technologies are used for a better representation of the working process, which implements a complex system of parallel execution of computationally difficult tasks of high spatial resolution image processing.