In response to the issues of a wide variety of crop types, poor resource information standardization, and low model training accuracy in China, focusing on seven crops: wheat, rice, maize, soybeans, cotton, peanuts, and rapeseed, using parameters like variety, morphology, yield, and quality as indicators, totally 83 crop variety entities were constructed. A manual annotation approach was adopted and an information extraction four-layer network model (BERT-PGD-BiLSTM-CRF) was introduced by incorporating adversarial training techniques. The model utilized the bidirectional encoder representation from transformers(BERT) model, based on a deep bidirectional transformer, as a pre-training model to acquire semantic representations of words and phrases. It employed projected gradient descent (PGD) adversarial training to introduce perturbations to the samples, thereby enhancing model robustness and generalization. Additionally, it leveraged a bidirectional long short-term memory (BiLSTM) network to capture long-distance text information and combined conditional random fields (CRF) to learn label constraint information. Comparing the training results with 18 different information extraction models, the research indicated that the proposed BERT-PGD-BiLSTM-CRF model achieved a precision of 95.4%, a recall of 97.0%, and an F1 score of 96.2%. This suggested that the BERT-PGD-BiLSTM-CRF model, utilizing adversarial training techniques, was effective in extracting crop variety information and also provided a technological reference for agricultural information extraction.