Enhancing plant morphological trait identification in herbarium collections through deep learning-based segmentation

Area of Science:

• Botany
• Computer Science
• Biodiversity Research

Background:

• Digitized herbarium collections are crucial for plant evolution and biodiversity studies.
• Analyzing these collections with deep learning is challenging due to heterogeneous backgrounds.
• Removing non-plant backgrounds is hypothesized to improve algorithm performance.

Purpose of the Study:

• To develop a deep learning method for segmenting plant masks and removing backgrounds from herbarium scans.
• To improve the accuracy of plant trait identification using processed herbarium images.

Main Methods:

• A novel deep learning approach was used to segment plant masks and remove non-plant backgrounds.
• Semi-automatic preprocessing reduced manual effort in dataset preparation.
• The method was evaluated on its effectiveness in image segmentation and subsequent classification tasks.

Main Results:

• Image segmentation achieved a high F1 score of up to 96.6%.
• Segmentation improved classification accuracy by up to 3% and F1 score by up to 7% for plant trait identification.
• The processed images significantly enhanced the performance of classification models.

Conclusions:

• Effective image segmentation is vital for analyzing herbarium scans.
• The proposed method successfully isolates plant elements, improving downstream classification tasks.
• This approach enhances the utility of digitized herbarium collections for scientific research.