Variational Autoencoders for Heterogeneous Data Integration: Applications in Remote Sensing, Fusion, and Anomaly Detection

Abstract

This sandwich thesis comprises a comprehensive survey of Cognitive IoT and remote sensing systems, followed by three technical contributions that advance the state-of-the-art in data compression, multi-modal fusion, and anomaly detection. The increasing integration of the Internet of Things (IoT) and remote sensing systems has created an unprecedented need for efficient data processing, transmission, and integration. These systems often rely on heterogeneous data (spanning modalities such as numerical measurements, textual information, and imagery) each with unique characteristics and structures. While effective at reducing data size, traditional data compression and processing techniques often fail to retain the contextual and relational information required for downstream analytical tasks. This limitation is particularly acute in resource-constrained environments, where computational power, bandwidth, and energy are restricted. This thesis explores Variational Autoencoders (VAEs) as a unifying framework to address these challenges. VAEs provide a mechanism for encoding complex, multi-modal data into low-dimensional latent representations that are simultaneously compact, efficient to transmit, and inherently structured for interpretability. The overarching goal of this research is to establish a methodology for representing information such that heterogeneous data can be processed, compressed, and fused seamlessly. The research is organized around three key objectives: (1) developing and fine-tuning VAE architectures that generate compressed latent spaces optimized for direct classification and reconstruction, minimizing the reliance on reconstructive processing while preserving interpretability, (2) investigating the capacity of VAEs for multi-modal data fusion by combining disparate data types, such as Synthetic Aperture Radar (SAR) and optical imagery, into a unified latent representation, and (3) evaluating the potential of VAE-derived latent spaces for anomaly detection, particularly in applications where identifying critical events or failures is essential. These results collectively underscore the potential of VAEs not only as tools for compression but also as versatile foundations for diverse analytical and predictive tasks across varied datasets. In the broader context of remote sensing and IoT, these methods align well with the overarching theme of the thesis to increase system efficiency through multi-level intelligence and distributed computing. By leveraging compressive sensing and latent representations, these approaches facilitate reduced data transmission and enhanced computational efficiency, supporting the development of scalable architectures for data-rich applications in IoT and remote sensing environments. The results also demonstrate that compressive VAEs generate rich latent spaces, enabling their dual use for direct downstream tasks and reconstruction as well as for data fusion and anomaly detection. This implies that deploying VAEs for compression on edge devices could fundamentally transform data transmission workflows. Rather than transmitting raw data, edge devices could send compressed, machine-learning-interpretable representations, reducing bandwidth requirements while preserving essential information for analysis and data fusion. This approach not only enhances efficiency but also lays the groundwork for intelligent, resource-aware systems capable of performing complex, real-time tasks through distributed and interpretive data handling. This thesis highlights the transformative potential of VAEs for addressing the critical challenges associated with processing and fusing heterogeneous data. By leveraging their inherent flexibility and capacity for structured representation, VAEs provide a scalable, interpretable, and resource-efficient approach for data-intensive applications in IoT. cognitive IoT (CIoT) and remote sensing. The findings lay a foundation for future research into compressive neural networks and their broader applications in intelligent systems.