# Assisting species differentiation and taxonomic classification by hyperspectral imaging: an example from the parasitic plant realm

## TL;DR

This paper shows how hyperspectral imaging and machine learning can accurately identify parasitic Cuscuta plants and their host crops, offering a non-invasive tool for agriculture.

## Contribution

The novel use of hyperspectral imaging combined with machine learning for species-level differentiation of parasitic plants and host tissue.

## Key findings

- Random Forest and Neural Network models achieved ~0.97 F1 scores in classifying host versus parasite and among Cuscuta species.
- Feature selection methods identified key wavelengths linked to chlorophyll and biochemical markers for accurate predictions.
- Hyperspectral imaging with machine learning offers a non-destructive, rapid method for field diagnostics of parasitic plants.

## Abstract

Cuscuta, a genus of parasitic plants, poses a threat to global agriculture by infesting a wide variety of economically important crops and facilitating the transmission of plant viruses. Accurate species identification is crucial for management but is traditionally based on morphological traits that require expert knowledge, limiting accessibility and early detection. Hyperspectral imaging, a technique that captures detailed reflectance information across hundreds of narrow and contiguous wavelength bands, offers the potential to non-invasively monitor plant health with high precision. This study aimed to explore whether hyperspectral imaging, combined with machine learning algorithms, can accurately differentiate between host plant tissue and parasitic Cuscuta species and further distinguish among different species within the genus.

Hyperspectral images were collected in both the visible-near infrared and short-wave infrared ranges, followed by preprocessing and segmentation of plant material from the background. The Normalized Difference Vegetation Index method yielded the most consistent segmentation performance. Random Forest and Neural Network models trained on segmented pixels achieved high classification accuracy and balanced F1 scores of approximately 0.97 in both binary (host versus parasite) and multiclass (species-level) classification. Feature selection using a genetic algorithm and an iterative elbow method successfully reduced the number of spectral bands needed for accurate predictions, identifying key wavelengths associated with chlorophyll content and other biochemical markers.

This study demonstrates the effectiveness of hyperspectral imaging combined with machine learning for identifying and classifying parasitic Cuscuta species. The findings highlight the potential of this approach for rapid, non-destructive field diagnostics and precision agriculture applications. As imaging hardware continues to improve and become more affordable, such integrated systems could be deployed in real-world crop monitoring and management to mitigate the impact of parasitic plants on global food production.

The online version contains supplementary material available at 10.1186/s13007-025-01498-y.