Identification of representatives of the genus Pulsatilla (Ranunculaceae) and their hybrids using convolutional neural networks

The presence of a limited number of reliable morphological features and a high level of interspecific hybridization in representatives of the genus Pulsatilla Mill. complicates their identification. In this study, the authors made a successful attempt to identify hybrids using various artificial intelligence tools. The initial data included 50 images of basal leaf blades for P. multifida, P. turczaninovii, and 8 for hybrid plants. An established computer vision approach to solving the classification problem using pre-trained convolutional neural networks was applied. First, classifiers based on the ResNet50 model were trained, the F1-score for which was 0.99. Then, it was shown that tools such as Google Teachable Machine (TM) and Personal Image Classifier (PIC) can be used for rapid prototyping of such solutions with virtually no knowledge of artificial intelligence technologies and programming skills. All test samples of P. turczaninovii and P. multifida were classified correctly.