Plant Phenotyping from Limited Training Data by Active Learning and Annotation

TMLR Paper7966 Authors

Abstract: AI-based plant phenotyping has significantly expanded the scope, scale, and speed of trait data collection. Phenotyping of epidermal cell patterning is a long-standing target of studies on carbon and water relations due to the important functions of stomata, bulliform cells and veins. Although SD estimation has benefited from AI-based techniques, automated segmentation and analysis of costal zones and bulliform cell regions have received much less attention. One major bottleneck is that AI tools often require manual annotation of large datasets for training, which is labor-intensive and requires domain expertise. In addition, models trained on one species frequently perform poorly on others. In this paper, we propose an automated framework that enables the detection of stomata, costal zones, and bulliform cell regions, leveraging domain knowledge and minimizing the need for extensive training data. We extract features from both topographic and intensity data, separating them using knowledge of spatial structure, specifically the repeatable, linear, periodic arrangement of epidermal cells, and intrinsic cell models to minimize noise. We bootstrap learning by starting with the most structured parts of the images and progressively adding less structured regions. Furthermore, we incorporate active annotation to continually expand the training dataset throughout the learning process. We demonstrate the effectiveness of our method in two areas: (i) stomatal detection and (ii) detection of costal and bulliform zones. Through extensive quantitative and qualitative experimental results on three crop species: Setaria viridis, Sorghum bicolor, and Zea mays, we show that our approach outperforms state-of-the-art segmentation methods in terms of both precision and time efficiency.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Shayok_Chakraborty1

Submission Number: 7966