Tackling the long-tailed challenge of greenhouse tomato cultivation cycles Recognition: A Sub-Group Guided, Multi-Expert Lightweight Framework

Ruochen Zhang^1,2Jingxin Yu^1,2Lin Han^1,3Cui Huankang^1,4Lichun Wang^1,2Fan Xu^1,2*Xiaoming Wei^1,2*

• ¹National Engineering Research Center for Intelligent Equipment in Agriculture, Beijing, China
• ²Beijing Academy of Agricultural and Forestry Sciences, Beijing, China
• ³Tianjin Agricultural University, Tianjin, Tianjin Municipality, China
• ⁴Xiamen University of Technology, Xiamen, Fujian, China

Greenhouse tomato cultivation cycles recognition is often impeded by the long-tailed challenge, which arises from significant differences in cycle lengths affecting data collection. This study proposes an innovative multi-expert grouping strategy and a lightweight model to address the long-tail recognition problem in greenhouse tomato Cultivation Cycles. We devised a novel approach: the dataset is divided into three groups based on sample quantity-head, balanced, and tail. Expert models are trained on each group, and knowledge distillation is employed to transfer the expertise to a lightweight student model. We enhanced the MobileViT foundation by incorporating multi-scale convolution modules, significantly improving the model's feature extraction capabilities. Experimental results demonstrate that our method outperforms existing state-of-the-art models regarding accuracy, precision, recall, and F1 score while maintaining a meager parameter count (0.95M). Notably, our model excels in handling tail classes, improving accuracy from 79.27% in the baseline model to 93.83%, significantly enhancing the recognition of rare stages. Additionally, our approach achieves remarkable results in minimizing performance disparities across different categories, with the maximum gap being only 3.49 percentage points. This research provides a new paradigm for addressing long-tail recognition issues in agricultural production and opens new avenues for applying lightweight models in complex scenarios. Our method achieves model lightweight while maintaining high performance, offering a viable solution for deploying intelligent systems in real-world greenhouse environments. Keyword：Long-tail recognition；Multi-expert grouping ；Lightweight model； MobileViT； Greenhouse tomato cultivation cycles Fig.1. Greenhouse tomato cultivation cycles Various solutions have been proposed in the academic community to address the long-tailed distribution problem. From the data level, methods such as oversampling (