High-resolution optical satellite imagery is a critical data source for geoscientific studies of urban development, forest dynamics, agricultural systems, and transportation infrastructure. However, its practical use is often limited by cloud occlusion, atmospheric interference, and sensor-related artifacts, leading to large missing or corrupted regions. These challenges are particularly severe in tropical and monsoon-influenced regions within the Asia–Oceania domain. Recent advances in Generative AI, especially diffusion-based models, have shown strong potential for image restoration. Nevertheless, their application to very high-resolution satellite imagery remains constrained by high computational cost and limited adaptation to heterogeneous geospatial structures. In this work, we propose a Generative AI framework based on kernel-adaptive diffusion for high-resolution satellite image inpainting that explicitly incorporates spatial context and land-cover variability. Building upon a kernel-adaptive optimization (KAO) formulation, our method operates in a latent diffusion space, enabling spatially adaptive reconstruction of missing regions while preserving large-scale geospatial coherence. Unlike conventional diffusion-based approaches that require dataset-specific retraining or extensive post-processing, the proposed framework dynamically modulates kernel behavior during the denoising process, allowing the model to respond to local structural and semantic heterogeneity across urban, forest, and agricultural landscapes. To improve robustness under large-area cloud occlusions, we further introduce a latent propagation mechanism that transfers structural information from valid regions into missing areas through a forward–backward diffusion process. This design enhances reconstruction stability in complex environments such as dense urban areas, fragmented road networks, and mixed natural–anthropogenic landscapes. Experiments on representative satellite datasets demonstrate consistent improvements in spatial continuity, boundary preservation, and perceptual realism compared with existing diffusion-based restoration methods. Overall, this work highlights the potential of kernel-adaptive Generative AI as a scalable and practical tool for improving data completeness in geoscientific remote sensing applications. Related implementation resources are available at https://kaopanboonyuen.github.io/KAO/.