| Abstract Detail
Biodiversity Informatics & Herbarium Digitization McCormick, Shelby [1], Bellis, Emily [2]. TRANSFER LEARNING FOR AUTOMATED PHENOPHASE PREDICTION OF PARASITIC STRIGA. Global climate change is altering the timing of key life cycle transitions (phenology). To monitor changes in phenology over time, digital images of herbarium specimens are an exceptionally rich source of information. For example, as of March 2022, the Global Biodiversity Information Facility (GBIF) contained 35 million digitized herbarium specimens of vascular plants sampled over the last two centuries. Annotating these many images with information reflecting fine-scale phenophase characteristics remains a major challenge that could be automated through application of machine learning approaches. However, guidance for transfer learning, which leverages previously trained deep learning models to improve performance on a new task, currently remains limited. To explore the translatability of phenophase classification models to new taxa, we used transfer learning to improve performance of a model trained on herbarium images of Asteraceae, on images of several representatives of the Orobanchaceae, which differ considerably in floral characteristics. We annotated 500 images comprising three species: Striga hermonthica, S. asiatica, and S. aspera, which are locally known as witchweeds. These parasitic plants are geographically widespread and significantly impact food security across Africa, infesting food crops such as maize, sorghum, sugarcane, rice, and millets. We used a 5-phase scale classification system, and tested model performance on S. asiatica and S. aspera, after transfer learning on S. hermonthica. Output from these models can be used to inform future work seeking to understand how phenology is changing through time across diverse environments in the Striga genus.
1 - Arkansas State University, Comp. Science and Mathematics, 12 Pinehurst Way, Maumelle, AR, 72113, United States
2 - Arkansas Bioscience Institute
Keywords:
Herbarium Digitization
Machine Learning.
Presentation Type: Poster
Number: PBI007
Abstract ID:501
Candidate for Awards:None |
