This Habilitation à Diriger des Recherches (HDR) synthesizes an eight-year research trajectory at the critical intersection of Artificial Intelligence (AI), Geospatial Analytics, and image processing. The core objective is to overcome methodological barriers to translate massive, complex spatio-temporal data into actionable intelligence for sustainable development and environmental governance. The work establishes a unified; end-to-end framework organized around three scientific axes that define novel contributions in GeoAI. The first axis, Advanced Spatio-Temporal Modeling for Forecasting Dynamics, introduces novel hybrid deep learning architectures—integrating functional decomposition, temporal attention, and spatial graph reasoning—to effectively model and robustly predict non-stationary, multivariate environmental dynamics. This significantly enhances predictive accuracy for critical variables like drought indices and vegetation monitoring, enabling proactive resource management. The second axis, Representation Learning and Multimodal Data Fusion, addresses data complexity by proposing advanced representation learning techniques. Leveraging autoencoder and probabilistic approaches, these methods generate temporally stable and semantically rich embeddings from multimodal EO data, significantly improving high-resolution land-use/land-cover mapping and change detection (e.g., desertification monitoring), with a strong focus on model robustness and interpretability. The final axis, Reproducible and Scalable Software Frameworks for GeoAI Analytics, ensures societal relevance through FAIR and R\&R principles, translating core methodological innovations into accessible, open-source platforms and decision-support systems (DeepWealth, SMETool). This commitment guarantees the trustworthiness and ethical integration of GeoAI into policy-making. Collectively, this HDR articulates a scientific vision for the design and validation of adaptive, interpretable, and fully reproducible GeoAI systems essential for resilience modeling. The contributions directly advance Sustainable Development Goals (SDGs) 1, 2, 6, 11, and 15, simultaneously charting critical research avenues toward multimodal learning fusion, online learning, explainable AI, and geospatial digital twins.