Continuous Growth Monitoring and Prediction with 1D Convolutional Neural Network Using Generated Data with Vision...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Traditional crop growth monitoring involves destructive sampling, leading to data discontinuity.
• Real-time, non-destructive monitoring is crucial for dynamic feedback control and immediate crop growth assessment.
• RGB images offer rich, non-destructive data on crop development stages.

Purpose of the Study:

• To develop and validate a non-destructive crop growth prediction method using low-cost RGB images and computer vision.
• To compare the performance of image-to-growth and growth simulation models for predicting crop biomass and leaf area.
• To demonstrate the potential of deep learning for real-time crop monitoring and environmental feedback control.

Main Methods:

• Utilized low-cost RGB images as input for computer vision models.
• Developed and compared two methodologies: an image-to-growth model and a growth simulation model.
• Evaluated the performance of Vision Transformer (ViT) and 1D Convolutional Neural Network (1D ConvNet) for predicting shoot fresh weight, shoot dry weight, and leaf area in lettuce.

Main Results:

• The Vision Transformer (ViT) achieved R² values of 0.89 (fresh weight), 0.93 (dry weight), and 0.78 (leaf area).
• The 1D Convolutional Neural Network (1D ConvNet) demonstrated superior performance with R² values of 0.96 (fresh weight), 0.94 (dry weight), and 0.95 (leaf area).
• High accuracies confirm that RGB images and deep neural networks can effectively interpret crop-environment interactions.

Conclusions:

• Non-destructive growth prediction using RGB images and deep learning is feasible and accurate.
• The 1D ConvNet model showed excellent performance in predicting key growth parameters for lettuce.
• This approach enables growers to enhance resource use efficiency and yield high-quality crops through real-time monitoring and feedback control.