STAR: Soil texture analysis recognizer integrating domain-adaptive transfer learning with NIR spectroscopy

Area of Science:

• Soil Science
• Spectroscopy
• Artificial Intelligence

Background:

• Soil texture is crucial for agriculture and land management.
• Near-infrared (NIR) spectroscopy offers a rapid, non-destructive alternative to traditional soil analysis.
• Current NIR methods face challenges in model generalization and data dependency.

Purpose of the Study:

• To introduce the Soil Texture Analysis Recognizer (STAR), an intelligent NIR-based device for precise soil texture classification.
• To develop a domain-adaptive deep learning strategy to improve model generalization and reduce cross-domain inconsistencies.
• To address limitations in current NIR spectroscopy applications for soil analysis.

Main Methods:

• Development of the Soil Texture Analysis Recognizer (STAR) device.
• Implementation of a transfer learning-based spectral preprocessing method (Transfer Multiplicative Scatter Correction - TMSC).
• Utilizing the Selective Enhanced Transfer Adaptive Boosting (SETAB) framework for enhanced model adaptability.

Main Results:

• STAR achieved 85.0% overall classification accuracy and a Kappa coefficient of 0.78 for 5 soil texture classes.
• Successfully identified previously unseen soil types, including loamy sand (100.0%) and sandy loam (66.7%).
• Demonstrated robust generalization capability and practical utility for soil texture analysis.

Conclusions:

• The STAR device and its domain-adaptive deep learning strategy offer a significant advancement in soil texture analysis.
• The proposed methods provide a feasible solution for bridging deep learning-assisted spectral modeling with real-world applications.
• STAR presents a scalable platform for broader NIR-based soil property measurements.