The Bemisia tabaci species complex is a group of tropical–subtropical hemipterans, some species of which have achieved global distribution over the past 150 years. Several species are regarded currently as among the world’s most pernicious agricultural pests, causing a variety of damage types via direct feeding and plant-disease transmission. Long considered a single variable species, genetic, molecular and reproductive compatibility analyses have revealed that this “species” is actually a complex of between 24 and 48 morphologically cryptic species. However, determinations of which populations represent distinct species have been hampered by a failure to integrate genetic/molecular and morphological species–diagnoses. This, in turn, has limited the success of outbreak-control and eradication programs. Previous morphological investigations, based on traditional and geometric morphometric procedures, have had limited success in identifying genetic/molecular species from patterns of morphological variation in puparia. As an alternative, our investigation focused on exploring the use of a deep-learning convolution neural network (CNN) trained on puparial images and based on an embedded, group-contrast training protocol as a means of searching for consistent differences in puparial morphology. Fifteen molecular species were selected for analysis, all of which had been identified via DNA barcoding and confirmed using more extensive molecular characterizations and crossing experiments. Results demonstrate that all 15 species can be discriminated successfully based on differences in puparium morphology alone. This level of discrimination was achieved for laboratory populations reared on both hairy-leaved and glabrous-leaved host plants. Moreover, cross-tabulation tests confirmed the generality and stability of the CNN discriminant system trained on both ecophenotypic variants. The ability to identify B. tabaci species quickly and accurately from puparial images has the potential to address many long-standing problems in B. tabaci taxonomy and systematics as well as playing a vital role in ongoing pest-management efforts. [Aleyrodidae; entomology; Hemiptera; machine learning; morphometrics; pest control; systematics; taxonomy; whiteflies.]

中文翻译：

---

基于形态学的烟粉虱隐匿种蛹通过嵌入式组对比卷积神经网络分析识别

烟粉虱物种复合体是一组热带-亚热带半翅目动物，其中一些物种在过去 150 年中已实现全球分布。目前，有几个物种被认为是世界上最有害的农业害虫之一，它们通过直接取食和植物病害传播造成多种损害类型。长期以来被认为是单一可变物种，遗传、分子和生殖相容性分析表明，这个“物种”实际上是 24 到 48 个形态上神秘的物种的复合体。然而，由于未能整合遗传/分子和形态学物种诊断，阻碍了对哪些种群代表不同物种的确定。这反过来又限制了疫情控制和根除计划的成功。以前的形态学研究，基于传统和几何形态测量程序，在从蛹的形态变异模式中识别遗传/分子物种方面的成功有限。作为替代方案，我们的研究重点是探索使用在蛹图像上训练的深度学习卷积神经网络 (CNN)，并基于嵌入式、组对比训练协议作为搜索蛹形态一致差异的手段。选择了 15 个分子物种进行分析，所有这些物种都已通过 DNA 条形码鉴定，并通过更广泛的分子表征和交叉实验得到证实。结果表明，仅基于蛹形态的差异就可以成功区分所有 15 个物种。对于在多毛叶和无毛叶寄主植物上饲养的实验室种群，实现了这种水平的区分。此外，交叉表测试证实了在两种生态表型变体上训练的 CNN 判别系统的普遍性和稳定性。从蛹图像中快速准确地识别烟粉虱物种的能力有可能解决烟粉虱分类学和系统学中许多长期存在的问题，并在正在进行的害虫管理工作中发挥重要作用。[粉虱科; 昆虫学; 半翅目; 机器学习；形态测量学；除害虫; 系统学；分类; 粉虱。] 从蛹的图像中快速准确地提取烟粉虱物种有可能解决烟粉虱分类学和系统学中许多长期存在的问题，并在正在进行的害虫管理工作中发挥重要作用。[粉虱科; 昆虫学; 半翅目; 机器学习；形态测量学；除害虫; 系统学；分类; 粉虱。] 从蛹的图像中快速准确地提取烟粉虱物种有可能解决烟粉虱分类学和系统学中许多长期存在的问题，并在正在进行的害虫管理工作中发挥重要作用。[粉虱科; 昆虫学; 半翅目; 机器学习；形态测量学；除害虫; 系统学；分类; 粉虱。]