Chinese wheat named entity recognition can obtain relevant entity information in the field of wheat agriculture in recent years, hybrid embedding models have become very popular in Chinese NER tasks, but there are still some challenges. Firstly, existing research methods capture the structural information with the problem of insufficiently comprehensive feature extraction, and the rich morphological information in agricultural texts is not fully exploited. In addition, existing methods focus too much on irrelevant structural features of Chinese characters in feature fusion, which generates noise. To solve these problems, we first create a rich Chinese Wheat Pest and Disease Dataset (CWDPD) and propose an agricultural named entity recognition model (ME-CWNER) based on it. The ME-CWNER model introduces the structural information based on the word model, and improves the semantic representation of the characters by capturing the radical and glyph information of Chinese characters through the bimodal perspectives of text and image, respectively; then, a new structural feature fusion method of Chinese characters is designed, which allows the characters to pay more attention to their own structural features, thus efficiently reducing the noise, and finally, the final features are obtained by using the gated fusion module. Experimental analysis shows that on the specially annotated Chinese wheat pest corpus, the optimal F1 score of the model reaches 89.96%. In addition, the F1 value on the public Resume dataset also reaches 96.02%, which verifies the generalisation of the model.