In modern society, using sarcastic language in social media has gradually become a trend. Using sarcasm on social media is actually a subtle form of language that, compared to other negative languages, can attract more attention and have more far-reaching effects. As social platforms allow users to publish various types of posts, including text, images, videos, and more, the difficulty of detecting sarcasm on social media is increasing, making it a difficult challenge to improve the accuracy of sarcasm detection. This article proposes a hierarchical multimodal model for detecting sarcasm, which jointly learns three different types of features: textual features, image features, and image caption features, and uses cross-modal attention (CMA) mechanism to fuse these cross-modal embeddings to achieve sarcasm prediction. Compared to existing sarcasm detection methods, the proposed model outperforms mainstream sarcasm detection methods, improving the performance and accuracy of sarcasm detection models. Through experiments, the proposed model has achieved significant results on public datasets, surpassing all benchmark models.