Aiming to improve the accuracy of entity identification in apple production field, a new Transformer optimization model was proposed. Firstly, in order to address the lack of apple production dataset, a corpus focusing on diseases and pests was constructed based on the knowledge and experience of horticultural experts in related field of apple cultivation. The accuracy of semantic representation of text was improved by combining word vector and character vector. Secondly, since the location information was crucial to text semantics，but the traditional Transformer model lacks the directionality of location information, in order to take advantage of the location features of text, an attention mechanism with direction and distance perception was introduced in Transformer encoder. And the contextual long-distance dependence features of BiLSTM was integrated on average to enhance semantic representation. Lastly, with imposing restrictions on labeling results by conditional random fields (CRF), the Transformer optimization model was obtained. The experimental results showed that the F1 score of the proposed method was 92.66% in Chinese named entity recognition of Apple diseases and pests. It indicated that the method proposed could effectively identify the named entities of apple diseases and pest, and provide a technical means for the accurate and intelligent identification of other agricultural named entities.