Dr. Arti Singh: Image-based plant stress phenotyping in the fast changing world of AI

Machine Learning (ML) and its subtype deep learning (DL) approaches are rapidly advancing and are being executed at an exceptional scale in agriculture to generate automated solutions with higher accuracy particularly in phenotyping tasks. To train a robust DL model, we need large quantities of data. With recent improvements in high throughput phenotyping platforms like rovers and drones equipped with various sensors, we are now collecting large amounts of data minimizing data collection bottleneck in plant phenotyping. However, training a supervised ML model with annotated data remains a major bottleneck. To overcome this challenge, semi-supervised and self-supervised ML methods have been used for data annotation. ML algorithms for automated feature extraction from images can circumvent labelling of images and provide better estimation of plant stress severity. Semi-supervised learning saves human hours for labeling tasks, and material resources over supervised learning methods. Further improvement is self-supervised learning, automatically generates labels removing/minimizing human intervention to label samples. In the webinar, we will explore examples of supervised ML models used in plant stress phenotyping problems, and how we are moving towards self-supervised learning to overcome the data annotation challenge with plant stress datasets.