Multimodal Remote Sensing Data Classification using Semi-Supervised Variational Autoencoder

Abstract

In recent years, the rapid expansion of machine learning has inevitably led to the integration of artificial intelligence into diverse scientific disciplines, where machine learning techniques have played a pivotal role in revolutionizing the processing and analysis of large-scale datasets. This integration has significantly transformed the field of remote sensing.

This thesis contributes to this evolving landscape by presenting a comprehensive investigation into the classification of multimodal remote sensing data using semi-supervised Variational Autoencoder architectures. Variational Autoencoders have eme-- rged as a powerful tool for uncovering the underlying patterns and structures inherent in data, showing significant potential in semi-supervised learning.

The architectural innovation proposed here incorporates a latent feature-level fusion strategy into the Variational Autoencoder framework, enabling the seamless integration of multiple modalities within the realm of remote sensing. Through a series of extensive experiments conducted on dataset representing rural area, we demonstrate the critical impact of encoder selection and latent space dimensionality on classification performance.The semi-supervised Variational Autoencoder models outperformed traditionally used methods such as Support Vector Machines and Random Forests, not only in terms of metrics but also in qualitative performance and uncertainty assessment.

Furthermore, this study provides insights into the strengths and limitations associated with data-level fusion and latent feature-level fusion strategies. As we test the capability of the proposed architectures on progressively larger dataset of urban area, we gain a deeper understanding of the importance of qualitative analysis, which reveals valuable insights about the performance of each fusion strategy.

As we navigate the complex landscape of multimodal data analysis, the framework proposed in this thesis not only offers valuable insights into remote sensing but also opens up exciting possibilities for creative solutions and applications across a spectrum of scientific domains.