Accurately predicting reservoirs containing methane hydrates requires comprehensive interpretation of multi-source data, including seismic, logging, and geochemical measurements, as well as the experience and knowledge of geologists. However, significant heterogeneity in measurement units and semantic rules across different data modalities, combined with the subjective and qualitative nature of expert experience, often results in underutilization of multi-source information and hinders knowledge discovery. To solve these challenges, this study proposes a novel multi-modal data fusion framework that integrates ontology modeling with a cross-modal attention mechanism, using the Qiongdongnan area in the South China Sea as a case study. The framework aims to establish a unified semantic representation system to eliminate data heterogeneity and achieve deep integration of structured, semi-structured, and unstructured geological knowledge. First, an ontology model for methane hydrate exploration is constructed based on the petroleum system theory, encompassing the key elements of source, migration, reservoir, and preservation. Subsequently, natural language processing techniques are employed to automatically extract entity-relation triples from extensive geological literature. Building on this, a graph attention network and cross-modal attention mechanism are introduced to capture multi-scale spatiotemporal correlations among seismic attributes, logging facies, and geochemical anomalies. The extracted triples and reservoir-forming features are then jointly embedded into a knowledge graph. Experimental results demonstrate that the proposed method effectively bridges the semantic gap between data and knowledge, significantly enhancing the interpretability of methane hydrate systems in complex geological settings and offering a new technological pathway for hydrate resource assessment.