Process-Based Crop Modeling for High Applicability with Attention Mechanism and Multitask Decoders.

Area of Science:

• Agricultural Science
• Computer Science
• Machine Learning

Background:

• Crop models face compatibility issues due to diverse research scales and methodologies.
• Deep neural networks offer adaptability for complex modeling tasks but haven't been fully integrated into process-based crop models.

Purpose of the Study:

• To develop a process-based deep learning crop model for hydroponic sweet peppers.
• To enhance model adaptability and integration capabilities for agricultural systems.

Main Methods:

• Developed DeepCrop, a process-based deep learning model incorporating attention mechanisms and multitask learning.
• Modified algorithms for regression tasks in crop growth simulation.
• Conducted controlled hydroponic sweet pepper cultivations over two years for data generation.

Main Results:

• DeepCrop achieved the highest modeling efficiency (0.76) and lowest normalized mean squared error (0.18) on unseen data.
• Analysis of attention weights and t-distributed stochastic neighbor embedding supported the model's analytical capabilities.
• Demonstrated superior performance compared to existing accessible crop models.

Conclusions:

• DeepCrop offers a versatile and highly adaptable alternative to traditional crop models.
• The model can analyze complex agricultural systems by processing intricate environmental data.
• Deep learning integration in process-based crop modeling shows significant promise for agricultural advancements.