Quantum Machine Learning for Earth Observation: Methods, Applications and Future Directions

Earth Observation (EO) plays a key role in climate monitoring, urban planning, and natural resource management, producing increasingly large and complex datasets. While traditional Artificial Intelligence techniques, particularly classical Machine Learning and Deep Learning, have significantly improved the analysis of EO data, they are beginning to encounter important limitations in terms of computational cost, scalability, and model efficiency. This doctoral thesis investigates Quantum Computing as a potential paradigm to address these challenges, focusing on the integration of Quantum Machine Learning (QML) methods within EO workflows. Given the current constraints of Noisy Intermediate-Scale Quantum (NISQ) devices, the research does not aim to replace classical approaches but rather explores hybrid quantum-classical architectures based on a co-design perspective. The proposed framework introduces and evaluates several hybrid models designed for different EO tasks. First, Quanv4EO leverages quanvolutional neural networks that use parameterized quantum circuits as local feature extractors, demonstrating promising results in land cover classification, building segmentation, speckle filtering, and water quality monitoring across multiple satellite datasets. Second, hybrid Quantum Graph Neural Networks are developed to model long-range spatial dependencies and complex global relationships, showing faster convergence when applied to the forecasting of large-scale geophysical indicators such as the Oceanic Niño Index. Finally, Quantum Diffusion Models integrate quantum layers within generative diffusion frameworks to support the synthesis and augmentation of EO data, improving the semantic consistency and realism of generated satellite imagery. Experimental results indicate that, although a definitive quantum advantage over state-of-theart classical deep learning models cannot yet be claimed, hybrid architectures provide an attractive balance between performance and model complexity. In particular, quantum modules offer higher expressivity per parameter, enabling substantial reductions in the number of trainable parameters while maintaining competitive accuracy. Overall, the thesis outlines a practical research direction for the evolution of QML for EO, moving it from an exploratory research area toward a more systematic and operational component of future satellite data analysis systems.