Assisting species differentiation using hyperspectral imaging
Authors
Vasili Alexander Balios (presenting author) [1]
Samuel Ortega [2]
Karsten Heia [2]
Anna Avetisyan [1, 3]
Kirsten Krause [1]
Affiliations
- Department of Arctic and Marine Biology, UiT The Arctic University of
Email (presenting author): vasili.a.balios@uit.no - Department of Seafood Industry
- Department Scientific
Keywords
Cuscuta; hyperspectral; machine-learning; species classification; Host/parasite differentiation.
Abstract
Cuscuta species, commonly known as dodder, are parasitic plants infecting a broad range of globally significant crops severely impacting yields. Climate change is expected to further expand Cuscuta distribution due to shifts in temperature and precipitation patterns. Effective management of Cuscuta infestations is challenging due to limited resistant crop genotypes, drawbacks of mechanical removal, and restrictions on chemical controls. We evaluate hyperspectral imaging (HSI), coupled with machine learning, to differentiate between host plants (Pelargonium zonale) and parasitic Cuscuta species, including C. campestris, C. monogyna, C. reflexa, and C. platyloba. Images were captured using two hyperspectral cameras, covering visible and near-infrared (VNIR, 407–995 nm) and short-wave infrared (SWIR, 950–2518 nm) spectral regions. Multiple segmentation algorithms were evaluated. NDVI-based segmentation emerged as the most efficient and consistent method, significantly improving classification accuracy. Machine learning models, specifically random forest and neural network classifiers, were trained to classify pixels as either host or parasite, as well as distinguishing among the four Cuscuta species. High classification accuracy was achieved. Confusion mainly occurred among closely related species pairs (C. campestris with C. platyloba and C. reflexa with C. monogyna). Feature importance analysis identified critical spectral bands linked to chlorophyll and carotenoid content, crucial for differentiating host plants from parasites and among parasite species. To optimize data handling, a genetic algorithm and elbow method were employed, significantly reducing the required spectral bands, while maintaining high classification accuracy. These models could be further trained and would be particularly suited for field applications, which could be applied to real-time monitoring and early detection of Cuscuta infestations.
Funding
This research was part of a PhD project financed by UiT - The Arctic University of Norway. Work on Cuscuta in the group of Kirsten Krause was supported by the Norwegian Research Council FriPro grant 301175. Anna Avetisyan was funded by grant CPEA-LT-2016/10092 from the Norwegian Agency for International Cooperation and Quality Enhancement in Higher Education (DIKU).