With the rapid advancement of information technology, it has become a trend for farmers to address offline agricultural issues through online intelligent question-and-answer systems. Question classification plays a crucial role in question-and-answer systems, as its accuracy directly determines the correctness of the final answers. Traditional single-label text classification models often struggle to accurately capture the precise intent of agricultural queries. Moreover, the lack of large-scale publicly available query datasets about agricultural pest and disease poses a significant challenge to existing research methods. To address these challenges, a hierarchical classification framework for queries about agricultural pest and disease was established based on a tree-like structure. This framework progressively refined the classification from the ambiguity of queries towards precision, aiming to overcome the semantic complexity of agricultural queries. Additionally, adversarial training method was introduced. By constructing adversarial samples and incorporating them into the training of large-scale language models, the model's generalization capabilities were enhanced, while mitigating issues arising from limited training data. Experimental validation conducted on real question-and-answer corpora demonstrated that the proposed method significantly enhanced the classification performance of queries about agricultural pest and disease. The research result can provide an effective means of identifying the intent behind agricultural queries, thereby offering support for advancing agricultural informatization.