A Federated Learning Framework and System for Land Mapping with Deep Learning and Visualisation

Land use mapping, a critical domain in urban planning and environmental management, has witnessed unprecedented advancements driven by data-driven techniques. This thesis focuses on a multifaceted exploration of land use mapping techniques, by using classical machine learning and deep learning algorithms. Through investigation, this study introduces a novel framework that combines classical machine learning and cutting-edge deep learning methods to address the practical issues in land use mapping.
This thesis starts with presenting the background and significance of enhanced classification techniques. By bridging the gap between research and application, this study concentrates on improved prediction accuracy, interpretability, and decision-making in land use mapping. The diverse challenges inherent to the field are explored, covering the way for new solutions. A comprehensive review of land use mapping algorithms sheds light on classical machine-learning techniques alongside deep-learning methodologies. Visualisation techniques are used to visualise data exploration, network architecture, and model training. This thesis presents a novel framework with its architecture, requirements, and functional modules. The proposed framework employs algorithms such as embedding, SVM, CNN, and CapsNet.
In addition, a system built upon the proposed framework has been implemented through federated learning, achieving both local and cloud implementations. This system enables local servers distributed across various geographic locations to contribute to the mapping process. Each local node retains control over its data while participating in the learning process. This approach not only preserves the confidentiality of sensitive information but also encourages collaboration among diverse datasets from different regions without the need for central data aggregation. On the other hand, the cloud-based implementation serves as a centralised hub for coordinating the federated learning process. This setup facilitates the aggregation of model updates from the distributed local nodes. By consolidating the collective knowledge from these nodes, the cloud implementation enables a better understanding of land use patterns on a larger scale. It ensures scalability, allowing the framework to handle extensive datasets and complex mapping tasks efficiently. Moreover, the exploration of visualisation techniques enhances the framework's flexibility, improving interpretability and guiding decision-making processes.
The experimental evaluations have been conducted on different platforms (Windows, Windows Subsystem Linux (WSL), Ubuntu Linux, Google Colab, and Rocky Linux) against the five datasets to evaluate the framework's performance and accuracy. These five datasets represent a diverse spectrum of land image types and complexities. Each dataset represents varied land image structures, sizes, and characteristics, providing valuable insights into the framework's adaptability and robustness across different data landscapes.
Except for experimental evaluations, we integrated the framework into real-world applications for its practical utility. The real-world deployments showcase the framework's adaptability and efficacy. The framework not only displayed its technical proficiency but also demonstrated its resilience in meeting the real-world demands of practical environments.
In conclusion, this thesis presents findings, contributions to the field, and outlines future research. This study aims to provide a guide for researchers, practitioners, and scientists engaged in land use mapping. Provided with comprehensive algorithms, a dynamic framework, and practical applications, it equips decision-makers, urban planners, and researchers with the tools to navigate the growing urban landscape with precision and insight.