Semisupervised Deep State-Space Model for Plant Growth Modeling.

Area of Science:

• Agricultural Science
• Artificial Intelligence
• Machine Learning

Background:

• Optimal control of crop sugar content is crucial for stable, high-quality fruit production.
• Current methods for measuring sugar content are manual and data is sparse due to limited remote sensing technology.
• Model-based reinforcement learning (RL) offers a promising approach for optimizing crop management through environmental modeling.

Purpose of the Study:

• To develop an advanced plant growth model for precise sugar content control in fruiting crops.
• To address the challenge of sparse sugar content data in plant growth modeling.
• To integrate a novel semisupervised deep state-space model (SDSSM) with model-based RL for enhanced crop cultivation.

Main Methods:

• Proposed a semisupervised deep state-space model (SDSSM), integrating semisupervised learning into a sequential deep generative model.
• Utilized SDSSM for plant growth modeling as an environmental model within a model-based RL framework.
• Employed cross-validation for comparative evaluation using tomato greenhouse cultivation data.

Main Results:

• SDSSM demonstrated high generalization performance by inferring unobserved data and efficiently using sparse training data.
• The model achieved a 38% reduction in mean absolute error for sugar content estimation compared to supervised learning algorithms.
• SDSSM successfully estimated time-series sugar content variations and validated uncertainty.

Conclusions:

• SDSSM shows significant potential for accurate time-series sugar content estimation in plant growth modeling.
• The developed model enhances the optimal control of sugar content for high-quality fruit cultivation using model-based RL.
• This approach effectively addresses data sparsity challenges in agricultural applications.