Transfer learning for improving generalizability in predicting soybean maturity date using UAV imagery.

Area of Science:

• Agricultural Science
• Plant Breeding
• Remote Sensing

Background:

• Accurate phenotyping is crucial for efficient crop breeding, especially for soybean (Glycine max (L.) Merr.).
• Maturity group significantly impacts soybean adaptation and yield potential, necessitating precise maturity date assessment.
• Transfer learning offers a promising approach to generalize predictive models across diverse environments.

Purpose of the Study:

• To evaluate the effectiveness of transfer learning techniques for predicting soybean maturity dates.
• To improve the generalizability of models using historical data for new breeding trials.
• To assess the impact of fine-tuning sample size on prediction accuracy.

Main Methods:

• Utilized multispectral imagery from unmanned aerial vehicles (UAVs) collected across five soybean breeding trials (2018-2021).
• Employed seven image features as predictors for maturity date prediction models.
• Compared three transfer learning techniques: pre-training and fine-tuning, single-source domain adaptation, and multiple-source domain adaptation.

Main Results:

• The pre-training and fine-tuning transfer learning technique achieved the highest prediction accuracy (R² = 0.74 and 0.79).
• Models demonstrated strong agreement with visual maturity assessments, with root mean square errors of 1.70 and 1.96 days.
• The number of fine-tuning samples had a negligible effect on the prediction accuracy for new datasets.

Conclusions:

• Transfer learning, particularly pre-training and fine-tuning, effectively generalizes soybean maturity prediction models to new environments.
• This approach provides a valuable framework for utilizing accumulated data to enhance practical deep learning applications in crop breeding.
• The findings support the use of historical data and aerial imagery for robust and efficient soybean cultivar selection.