Tackling the long-tailed challenge of greenhouse tomato cultivation cycles recognition: a sub-group guided,
Ruochen Zhang1, Jingxin Yu1, Lin Han1,2
1National Engineering Research Center for Intelligent Equipment in Agriculture, Beijing, China.
Introduction:
Greenhouse tomato cultivation cycles recognition is often impeded by the long-tailed challenge, arising from significant differences in cycle lengths affecting data distribution. This imbalance hinders accurate recognition, particularly for rare stages, limiting intelligent management in precision agriculture.
Methods:
This study proposes a lightweight framework integrating a novel multi-expert grouping strategy with knowledge distillation. The dataset is divided into three groups (Head, Balanced, Tail) based on sample quantity. Separate expert models are trained on each group. Knowledge distillation then transfers the expertise of these models to a lightweight student model (MSC-MobileViT). MSC-MobileViT enhances the MobileViT foundation by incorporating a multi-scale convolution module to improve feature extraction across different scales, capturing both local details and global structure.
Results:
Experimental results demonstrate superior performance. The framework achieves an overall accuracy of 95.99%, precision of 91.03%, recall of 93.57%, and F1-score of 92.02%, outperforming state-of-the-art models (ResNet50, MobileNetV3, MobileViT variants). Crucially, it excels in handling tail classes, improving accuracy from 79.27% (baseline) to 93.83% for rare stages like "Substrate Soaking" and "Early Production". The maximum performance gap across categories is minimized to only 3.49 percentage points. The student model achieves this high performance while maintaining an extremely low parameter count (0.95M).
Discussion:
The proposed framework effectively addresses the long-tailed recognition challenge in greenhouse tomato cultivation cycles. The multi-expert grouping strategy optimizes learning for different data distributions, while knowledge distillation enables high performance within a lightweight model suitable for edge deployment. The integration of multi-scale convolution significantly enhances feature extraction in complex agricultural scenes. This research provides a new paradigm for long-tail recognition in agriculture and demonstrates the viability of deploying efficient, high-accuracy intelligent systems in real-world greenhouse environments.
