This project advances the development and fine-tuning of Geospatial Foundation Models (GFMs) tailored for Earth Observation data. By processing massive multi-spectral and high-resolution temporal satellite image time series, these models are designed to significantly improve critical downstream applications, including climate monitoring, land-use classification, and rapid disaster response. A core scientific objective of this next phase is transitioning our representation learning paradigm from fixed Self-Supervised Learning (SSL) datasets to robust continual learning approaches. Because Earth observation data is inherently a continuous temporal stream, enabling our models could dynamically learn from incoming multi-modal data. Furthermore, we are actively exploring and integrating novel diffusion-based architectures to enhance the generative capacities and complex feature extraction of our geospatial models. Led by PI Andrea Nascetti and supported by three PhD candidates, our group will build upon large-scale training strategies that we have successfully validated. Our team has a proven track record of establishing rigorous evaluation protocols in this domain, most notably through the development of a robust benchmark dataset for the geospatial community (Pangaea-bench: https://github.com/VMarsocci/pangaea-bench). Furthermore, our recent state-of-the-art results directly validate our expertise in both diffusion mechanisms and foundation model scaling, serving as the solid baseline for our next phase of continual learning development: References: Diffusion-Based Representation: SatDiFuser Model, Published at ICCV 2025 (https://github.com/yurujaja/SatDiFuser) Recent Foundation Model Advancements: Accepted at CVPR 2026 (https://arxiv.org/pdf/2603.02522)