AP162: A large-scale dataset for agricultural pest recognition

Pest image datasets play a crucial role in improving the accuracy of pest recognition models, enabling automated monitoring and precise pest control, and advancing intelligent agriculture . However, existing agricultural pest datasets have limitations in terms of species diversity and the number of samples, and only the appearance of the pest is considered while ignoring professional agricultural knowledge. In this paper, a large-scale dataset, termed AP162, is proposed for the recognition of agricultural pests, which consists of 162 pest categories and 194,700 images. AP162 takes into account both the visual appearance of agricultural pests and their harmful characteristics for crops. Different subset divisions are adopted to explore the appearance features and potential damage to pests from two perspectives. The agriculture-based perspective classifies pests according to crop damage types to support targeted pest control, while the vision-based perspective relies on morphological similarities. Extensive experiments have been conducted on the AP162 dataset utilizing recent deep learning methodologies, thereby establishing it as a new benchmark. This paper proposes a weighted fusion strategy that agricultural domain knowledge is integrated with vision-based fine-grained feature, achieving a recognition accuracy of 89.2% on AP162. It is publicly available at https://github.com/SCNYDX-KL/AP162 . • A new dataset, AP162, with 162 categories and 194,700 images for pest recognition. • AP162 includes agricultural- and vision-based subsets for pest control research. • Experiments show AP162 is challenging and establishes a new benchmark.