Semantic Segmentation of Vegetation in Remote Sensing Imagery Using Deep Learning

In recent years, the geospatial industry has been developing at a steady pace. This growth implies the addition of satellite constellations that produce a copious supply of satellite imagery and other Remote Sensing data on a daily basis. Sometimes, this information, even if in some cases we are referring to publicly available data, it sits unaccounted for due to the sheer size of it. Processing such large amounts of data with the help of human labour or by using traditional automation methods is not always a viable solution from the standpoint of both time and other resources. Within the present work, we propose an approach for creating a multi-modal and spatio-temporal dataset comprised of publicly available Remote Sensing data and testing for feasibility using state of the art Machine Learning (ML) techniques. Precisely, the usage of Convolutional Neural Networks (CNN) models that are capable of separating different classes of vegetation that are present in the proposed dataset. Popularity and success of similar methods in the context of Geographical Information Systems (GIS) and Computer Vision (CV) more generally indicate that methods alike should be taken in consideration and further analysed and developed. For the first part, we will start off with an overview of the current problem space and the argument necessity of studying such techniques. Following this, Chapter 2 begins by providing extended details regarding the current state of research for GIS starting from characteristics of Remote Sensing data to how the acquisition of such data is performed and to additional preprocessings that could be required. Within Chapter 2 we also present informations about the inner workings of ML with respect to current heuristics that are used in the literature for solving various CV tasks. Chapter 3 showcases the approach taken for developing an appropriate multi-modal spatio-temporal Remote Sensing dataset publicly available data. Some results on experimenting with ML techniques for training Convolutional Neural Networks (CNN) on the task of semantic segmentation of vegetation related classes are also presented for validating the usability and relevance of the aforementioned dataset. Chapter 4 illustrates the results of training such models on the proposed dataset on how they perform on solving the task of semantic segmentation of vegetation.