GAF-ResNet-MHSA: A novel transfer learning method for soil nutrient prediction in small sample datasets.

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Near Infrared Spectroscopy (NIR) is vital for soil nutrient analysis but faces challenges with small sample sizes and regional variations.
• Existing models often exhibit low prediction accuracy and overfitting, limiting their practical application.

Purpose of the Study:

• To develop a novel transfer learning approach, GAF-ResNet-MHSA, for accurate soil nutrient prediction using NIR spectral data.
• To address the limitations of small sample sizes and improve model generalization across different soil regions.

Main Methods:

• Transformed NIR spectral data into Gramian Angular Field (GAF) images.
• Utilized a modified ResNet34 framework incorporating a multi-head attention mechanism (MHSA).
• Applied transfer learning by fine-tuning the GAF-ResNet-MHSA model with target domain samples.

Main Results:

• Initial models on small samples showed limited accuracy (e.g., R² for pH: 0.7575, SAP: 0.7393).
• The GAF-ResNet-MHSA transfer learning approach significantly improved prediction accuracy (e.g., R² for pH: 0.8962, SAP: 0.8110).
• Key performance metrics like RMSEp and RPD showed substantial improvements, indicating enhanced model reliability.

Conclusions:

• The GAF-ResNet-MHSA transfer learning method effectively overcomes accuracy issues in small-sample soil spectral modeling.
• This innovative approach demonstrates practical applicability and potential for efficient spectral transfer learning in soil analysis.