Crossing multiple life stages: fine-grained classification of agricultural pests.

Background: Pest infestation poses a major challenge in the field of global plant protection, seriously threatening crop safety. To enhance crop protection and optimize control strategies, this study is dedicated to the precise identification of various pests that harm crops, thereby ensuring the efficient use of agricultural pesticides and achieving optimal plant protection.

Results: Currently, pest identification technologies lack accuracy, especially in recognizing pests across different growth stages. To address this issue, we constructed a large pest dataset that includes 102 pest species and 369 pest stages, totaling 51,670 images. This dataset focuses on the identification of pest growth stages, aimed at improving the efficiency of pest management and the effectiveness of plant protection. Moreover, we have introduced two innovative technologies to tackle the significant differences between pest growth stages: a Multi-stage Co-supervision mechanism and a Spatial Attention module. These technologies significantly enhance the model's ability to extract key features, thus boosting recognition accuracy. Compared to the industry-leading Vision Transformer-based methods, our model shows a significant improvement, increasing accuracy by 3.67% and the F1 score by 2.49%, without a significant increase in the number of parameters.

Conclusions: Extensive experimental validation has demonstrated our model's significant advantages in enhancing pest identification accuracy, which holds substantial practical significance for the precise application of pesticides and crop protection.