Hyperspectral Technique Combined With Deep Learning Algorithm for Prediction of Phenotyping Traits in Lettuce

Area of Science:

• Plant Science
• Biophysics
• Machine Learning

Background:

• Current plant phenotyping methods for biochemical traits are often destructive and slow.
• Hyperspectral imaging offers a non-destructive way to gather plant biophysical and biochemical data.
• Raw spectral data can be noisy and redundant, impacting traditional analysis model robustness.

Purpose of the Study:

• To develop and evaluate deep learning models for rapid, non-destructive prediction of lettuce phenotyping traits using spectral reflectance.
• To compare the performance of deep learning models against traditional multivariate analysis methods.

Main Methods:

• Two end-to-end deep learning models, Deep2D (2D convolutional neural networks) and DeepFC (fully connected neural networks), were developed.
• Visible near-infrared hyperspectral imaging was used to capture lettuce plant spectra.
• Deep learning models were trained and their performance compared against five multivariate analysis methods (linear and nonlinear).

Main Results:

• Deep learning models achieved high prediction accuracy (R² of 0.9030 for soluble solids content with Deep2D, 0.8490 for pH with DeepFC).
• Deep learning models outperformed all tested multivariate analysis methods in prediction accuracy.
• Deep learning models automatically extracted relevant features, eliminating the need for manual pretreatment and wavelength selection.

Conclusions:

• Deep learning models provide a robust and efficient approach for non-destructive prediction of plant biochemical traits from hyperspectral data.
• These models are well-suited for high-throughput plant phenotyping platforms due to their automation and speed.
• The developed deep learning framework offers a significant advancement over traditional multivariate analysis for plant trait prediction.