Identification of herbarium specimens: a case study with Piperaceae Giseke family

Although millions of herbarium specimens have been recently digitized, many of them have not yet been properly identified or reviewed. The main reason is that the classification process is manual, slow, and error-prone. Machine Learning techniques are promising alternatives for supporting herbarium plants identification. This paper evaluates feature extraction techniques and classification algorithms to identify herbarium specimens of the Piperaceae Giseke family at the genus level. For the evaluation, we extracted a balanced subset of pre-processed images from the five genera of the Piperaceae family (Manekia, Ottonia, Peperomia, Piper, and Pothomorphe) from the speciesLink repository. Our experiments point to potential support in identifying of herbarium images of the Piperaceae family, mainly for the genera Manekia, Peperomia and Ottonia. The best accuracy was 80.53% achieved by combining MobileNet-V2 and the SVM classifier.