Deep phenotyping: deep learning for temporal phenotype/genotype classification.

Area of Science:

• Plant Science
• Computer Science
• Bioinformatics

Background:

• Accelerating climate-ready crop breeding relies on high-throughput genotype-to-phenotype studies.
• Deep learning, particularly CNNs and LSTMs, excels in visual and sequential data analysis.
• Existing plant phenotyping often uses static images, overlooking dynamic growth patterns.

Purpose of the Study:

• To develop and evaluate a CNN-LSTM framework for plant accession classification using temporal data.
• To leverage LSTMs for encoding dynamic plant behavior and growth as discriminative phenotypes.
• To establish a benchmark dataset for time-series plant image analysis.

Main Methods:

• Proposed a hybrid CNN-LSTM framework for joint feature and classifier learning.
• Utilized LSTMs to analyze temporal information from plant growth sequences.
• Collected and shared a time-series image dataset of Arabidopsis accessions.

Main Results:

• The CNN-LSTM framework demonstrated superior performance compared to traditional hand-crafted features.
• Incorporating temporal information via LSTMs significantly enhanced accession classification accuracy.
• The developed dataset serves as a valuable benchmark for plant phenotyping research.

Conclusions:

• The proposed CNN-LSTM approach offers an effective method for plant accession classification.
• Leveraging temporal data with LSTMs improves classification accuracy in plant phenotyping.
• The framework has potential applications in automated plant production, environmental classification, and disease detection.