Exploiting graph-based data representations for multispectral remote sensing image completion

Abstract

Multispectral images from satellites are an indispensable source of information for monitoring the planet’s surface. However, such data frequently features gaps caused by cloud cover, sensor malfunctions or incomplete scene coverage. These gaps undermine the continuity and reliability of downstream analyses, creating the need for effective reconstruction methods. At the same time, the dynamic nature of Earth observation creates a demand for methods that are flexible and transferable across tasks, sensors and resolutions. This thesis establishes a novel graph-based framework for multispectral image reconstruction. The framework features a new representation for multispectral data, where images are modelled using graphs that encode spectral similarity relationships between pixels. This representation provides the foundation for the development of the proposed GraphProp method, a reconstruction algorithm that completes missing values using diffusion processes over the spectral-similarity graph. By propagating observed spectral information through the graph structure, GraphProp infers missing values directly from the relationships that are established from the observed data and encoded within the graph structures. The proposed approach is evaluated extensively across a wide range of reconstruction tasks, such as cloud gap imputation and sensor failure, using data from multiple multispectral sensors. The results demonstrate that a single algorithm, without retraining or adjustment, can be applied effectively across different reconstruction tasks and settings. In all cases, GraphProp consistently achieves state-of-the-art reconstruction accuracy, matching or surpassing leading interpolation-based, low-rank and deep learning approaches. In addition to accuracy, the method has been deliberately designed for practical deployment. Presented results show that the graph-based propagation can be solved efficiently, delivering runtimes that are orders of magnitude faster than low-rank methods and considerably more resource-efficient than deep learning models. By eliminating training overheads, GraphProp avoids the energy and financial costs associated with large-scale model development. These costs have been quantified, highlighting the efficiency and scalability of the proposed method for operational image reconstruction. The thesis advances both the theoretical and practical state of the art in image completion, demonstrating that a single framework can deliver accurate, robust and efficient reconstruction across diverse tasks and sensors. In doing so, GraphProp establishes a general solution for reconstructing gaps in multispectral imagery, supporting downstream analyses.