Multimodal remote sensing image clustering is fundamental for autonomously discovering semantic structures in unlabeled heterogeneous Earth observation data. While current anchor graph-based approaches offer computational efficiency, they face three key limitations: suboptimal anchor selection, inflexible deterministic fusion strategies, and insufficient spatial coherence preservation. To tackle these limitations, we propose Uncertainty-aware Deep Anchor Graph learning (UDAG), which jointly learns modality-specific anchors and a unified graph via encoder-decoders in a latent space. Crucially, we develop an uncertainty-aware dynamic fusion strategy that adaptively weights modality contributions based on their fine-grained reliability estimates. Furthermore, to enhance spatial consistency, we incorporate an isotropic total variation regularization into the anchor graph construction. Extensive experiments on multimodal remote sensing benchmarks demonstrate that UDAG consistently achieves superior clustering accuracy compared to state-of-the-art methods.