In social sensing, conflicting observations are provided by arbitrary humans as sensors with unknown reliabilities, which causes truth hidden in the tremendous observations. To solve these conflicts, truth discovery becomes a useful solution, it identifies the reliability of sources and the credibility of observations without prior knowledge. However, existing works only detect source dependencies through whether there is a copying behavior, without further fine-grained analysis of the propagation pattern. In addition, existing works assume that observations containing multi-modal features are more credible, ignoring the impact of multi-modal semantic consistency on the credibility of observations. In this paper, we develop a Propagation pattern and Semantic consistency Truth Discovery framework (PSTD). We consider different propagation patterns and analyze the sentiment behind the quoting behavior. Then, we calculate the semantic consistency evaluation of each multi-modal observation. Finally, we propose a novel probabilistic graphical model to incorporate all the above considerations into truth discovery and utilize a Gibbs sampling algorithm to estimate the truth. Experimental results on real-world and synthetic datasets prove the effectiveness of our method.