Doing More With Less: A Multitask Deep Learning Approach in Plant Phenotyping.

Area of Science:

• Plant Science
• Computer Vision
• Machine Learning

Background:

• Image-based plant phenotyping is crucial for evaluating plant traits, demanding efficient analysis techniques.
• Deep learning shows promise in plant phenotyping tasks like segmentation and counting.
• Extracting multiple traits simultaneously requires advanced analytical methods.

Purpose of the Study:

• To develop a multitask deep learning framework for simultaneous extraction of plant phenotyping traits.
• To investigate the efficacy of multitask learning (MTL) in improving leaf count prediction.
• To assess the impact of reduced annotation data on model performance.

Main Methods:

• A multitask deep learning framework was designed using a modified pretrained ResNet50 as a feature extractor.
• The framework was trained end-to-end to simultaneously predict leaf count, projected leaf area (PLA), and genotype.
• MTL was employed to leverage related tasks for improved generalization and reduced label dependency.

Main Results:

• The proposed MTL method achieved over 40% improvement in leaf count mean squared error (MSE) compared to single-task networks.
• The MTL framework demonstrated robust performance with up to 75% fewer leaf count annotations.
• Single-task models showed performance degradation with reduced annotations, unlike the MTL approach.

Conclusions:

• MTL offers a powerful approach for simultaneous multi-trait analysis in plant phenotyping.
• The developed framework enhances leaf count prediction accuracy and reduces the need for extensive manual annotation.
• This method facilitates more efficient and data-economical plant phenotyping.