Motivation: Spatial multi-omics can delineate tissue architecture more faithfully than any single modality, but unsupervised spatial domain detection remains difficult because different modalities exhibit distinct sparsity patterns, noise characteristics, and graph structure. Existing methods already couple spatial and multimodal signals in different ways, yet spot-level cross-modal dependencies can still be underexploited. We therefore present SPAMO, an unsupervised framework that couples adaptive dual-graph encoding with interaction-aware multimodal fusion.Results: SPAMO adaptively balances spatial adjacency and feature similarity during unimodal encoding, refines feature graphs during training, and performs cross-modal interaction before gated fusion, while regularizing the shared embedding with complementary global and local structural objectives. Across Human Lymph Node, Mouse Brain, and simulated benchmarks, SPAMO achieves the strongest overall clustering performance, with the clearest gains on the real datasets. These results support the value of modeling cross-modal interaction and graph structure jointly for spatial domain identification.