Classification of specimens is the important first step to characterize populations and species assemblages. Although species-level classification has been a popular goal, the sex difference and sex ratio are also an important property in ecology and pest control. Here we focus on the images of mixed sex specimens of a stored product pest beetle (Callosobruchus chinensis) and its parasitoids (parasitic wasps; Anisopteromalus and Heterospilus) in various postures and classify them into species and sex, by training supervised machine learning programs: logistic model trees (LMT), random forest, support vector machine (SVM), simple logistic regression, multilayer perceptron and AdaBoost (adaptive boosting). Both object-based features and pixel-based features were extracted from each image. Simple logistic regression, LMT and AdaBoost (employing simple logistic regression as base learner) performed well to classify sexes or species/sexes; average true positive rates (prediction accuracy) of 88.5–98.5% were achieved for within-species sexing of beetles or wasps, 97.3% for two species sexing and 93.3% for three species sexing. For most datasets, the best performed models incorporated both object-based features and pixel-based features. LMT models were identical to simple logistic regression models in most cases. Robust performance and small variation in prediction accuracy of simple logistic regression, irrespective of classification target (sexes or species), was shown, and this is probably because of the efficient feature selection implemented in the algorithm. This study is one of the earliest to classify the gender of insects using machine learning based on still images.

标本分类是表征种群和物种集合的重要的第一步。尽管物种级别的分类已成为一个普遍的目标，但性别差异和性别比也是生态和害虫防治中的重要属性。在这里，我们专注于一个存储的产品害虫甲虫（混合性别标本的图像绿豆象）和它的寄生（寄生蜂; Anisopteromalus和Heterospilus）通过训练监督机器学习程序以各种姿势将其分类为物种和性别：逻辑模型树（LMT），随机森林，支持向量机（SVM），简单逻辑回归，多层感知器和AdaBoost（自适应增强）。从每个图像中都提取了基于对象的特征和基于像素的特征。简单的逻辑回归，LMT和Ada​​Boost（采用简单的逻辑回归作为基础学习者）在对性别或物种/性行为进行分类方面表现良好；甲虫或黄蜂的种内性别鉴定的平均真实阳性率（预测准确度）为88.5–98.5％，两种性别鉴定为97.3％，三种性别鉴定为93.3％。对于大多数数据集，性能最佳的模型同时包含了基于对象的特征和基于像素的特征。在大多数情况下，LMT模型与简单的逻辑回归模型相同。无论分类目标（性别或物种）如何，简单的逻辑回归的鲁棒性能和预测准确性的小变化都可以显示出来，这可能是由于算法中实现了有效的特征选择。这项研究是使用基于静态图像的机器学习对昆虫性别进行分类的最早方法之一。