Avaliação de extratores de características para identificação de espécimes de herbário da família botânica Piperaceae

Abstract

Plant identification remains a challenging task due to the large number of species that have not yet been described and the need for in-depth knowledge of taxonomy. Despite the existence of more than 3,500 herbaria worldwide, housing approximately 400 million specimens, the identification process is hindered by factors such as the scarcity of specialists and the slowness inherent in manual work. In this context, Machine Learning models have shown promise for the automation of plant identification. In this study, we used a dataset composed of segmented images of specimens from the Piperaceae family. Feature extraction was performed using the convolutional neural networks VGG19, DenseNet201, and EfficientNet, and the generated vectors were employed in traditional classifiers such as Decision Tree, k-NN, SVM, MLP, and Random Forest. Due to the imbalance in the dataset, the SMOTE data augmentation technique was applied to increase the number of samples from minority classes. The results obtained were satisfactory with regard to feature extraction. DenseNet201 achieved the best performance, with results approximately 8% higher than those reported in previous studies. The application of SMOTE provided an additional gain of about 2%. Nevertheless, the best F1-Score achieved was 58%, a value that does not, by itself, allow the construction of a fully automatic identification system. However, the Top-3 and Top-5 metrics reached approximately 84% and 90%, respectively, indicating the feasibility of developing suggestion tools capable of assisting specialists in the classification of species from the Piperaceae family.