【Purpose/significance】During the disasters, multimodal information reflecting the real-time situation is widely published

in the social media. How to effectively identify informative tweets is important to disasters response.【Method/process】This paper pro⁃poses a social media informative tweets identification framework based on multimodal fusion. Formal features and semantic features of text and text in tweets are used as fusion inputs. Then, a binary classifier to identify informative is constructed through machine learn⁃ing and deep learning. Finally, the role of different tweet features and fusion methods in informative tweets identification are analyzed.【Result/conclusion】The results of informative tweets identification experiment show that the effect of semantic features is better than formal features, and the effect of multimodal fusion are better than single modal such as text or images. At the same time, the best way of identifying informative tweets related disasters is fusing formal and semantic features of text and images.【Innovation/limitation】Considering the formal and semantic features of multimodal information in tweets, we propose a multimodal fusion method of feature layer and decision layer to identify informative tweets in social media. However, the applicability of this method in different disasters needs to be further tested.