Learning from a few samples to automatically recognize the pests or plants is an attractive and promising study with a low cost of data to protect the agricultural yield and quality. Although there have been a handful of efforts on the few-shot classification in agriculture, none of them involves the task-driven meta-learning paradigm. This study is the first work of task-driven meta-learning few-shot classification in the field of agriculture to our best of knowledge. Specifically, we collected samples from publicly available resources to assemble a comprehensive dataset for the few-shot classification, covering both pests and plants to analyze the single domain or cross-domain. Then, we performed 36 groups of comparison experiments to establish the baselines of testing accuracy. Further, we summarized and explained the effect laws of factors on the few-shot performance, such as N-way, K-shot, and domain shift. In summary, this work can be regarded as a significant reference and the benchmark comparison for the follow-up studies of few-shot learning tasks in the agricultural field.

从少量样品中学习以自动识别有害生物或植物是一项有吸引力且有希望的研究，其数据成本低廉，可保护农业产量和质量。尽管在农业的少数镜头分类方面已经做出了一些努力，但它们都没有涉及任务驱动的元学习范式。这项研究是我们力所能及的任务驱动的元学习几次学习分类在农业领域的第一项工作。具体来说，我们从可公开获得的资源中收集了样本，以建立针对几次拍摄分类的综合数据集，涵盖了害虫和植物以分析单个域或跨域。然后，我们进行了36组比较实验，以建立测试准确性的基准。进一步，我们总结并解释了影响单次射击性能的因素的规律，例如N向，K次射击和域偏移。综上所述，这项工作可以作为农业领域少发学习任务的后续研究的重要参考和基准比较。