Detection of Fake Multispectral Remote Sensing Images Using Spectrum-Based Deep Learning and Subspace Learning Techniques

This is the latest version of this item.

Abstract

The rapid advancement of image generation and manipulation techniques has significantly increased the prevalence of highly realistic fake imagery, posing critical challenges for applications that rely on the integrity of remote sensing data. Detecting such manipulations is particularly difficult due to the heterogeneous nature of forgery mechanisms, which introduce fundamentally different and often non-overlapping artefacts. This thesis investigates the detection of fake remote sensing images from a representation-driven perspective. Challenging the conventional assumption that all manipulated images share common detectable patterns, this work demonstrates that different manipulation types—specifically GAN-generated images and copy-move forgery (CMF)—exhibit distinct characteristics that require fundamentally different feature representations. Through systematic analysis using frequency-domain (FFT), wavelet-based, and data-driven decomposition methods, it is shown that detection performance is strongly dependent on the alignment between feature representation and manipulation characteristics. To explore this dependency, multiple models are employed, including representation-specific ResNet architectures and the interpretable Geo-DefakeHop framework. Experimental results reveal that detection models do not fail randomly; instead, their failures are structured and directly linked to the type of representation they employ. Models trained on specific manipulation types exhibit limited generalisation due to a fundamental mismatch between the learned features and the underlying artefacts. To address this limitation, a representation-aware hierarchical ensemble framework is proposed. The framework integrates multiple specialised models through a conditional decision mechanism that adaptively selects and combines representations based on input characteristics. Unlike conventional ensemble methods, this approach explicitly accounts for representation compatibility, enabling the system to leverage complementary strengths while reducing conflicting predictions. As a result, the proposed framework achieves more robust and balanced performance across diverse manipulation types without requiring prior knowledge of the forgery type. Overall, this work establishes that fake image detection is inherently a multi-characteristic and representation-dependent problem. It provides both empirical evidence and a principled framework for transitioning from single-model detection to adaptive, representation-aware systems, offering improved robustness, interpretability, and generalisation in remote sensing image forensics.

Available Versions of this Item

• Detection of Fake Multispectral Remote Sensing Images Using Spectrum-Based Deep Learning and Subspace Learning Techniques. (deposited 07 May 2026 13:34) [Currently Displayed]