Luo et al. (2025) STAR: Soil texture analysis recognizer integrating domain-adaptive transfer learning with NIR spectroscopy

Identification

• Journal: Journal of Environmental Management
• Year: 2025
• Date: 2025-12-25
• Authors: Yuchen Luo, Zeyuan Zhang, Siyu Liu, Geng Leng, Wenbo Xu, Xuemei Luo, Yuewu Wang, Zhenwei Xie, Leyun He, Junwei Wang, Hongjin Tong, Nima ZongZong, Wenbo Fu
• DOI: 10.1016/j.jenvman.2025.128378

Research Groups

• School of Resources and Environment, University of Electronic Science and Technology of China, Chengdu, 611731, China
• Sichuan Province Natural Resources Science Academy, Chengdu, 610015, China
• Sichuan Chuan Huan Yuan Chuang Testing Technology Co., Ltd, Chengdu, 611731, China
• Sichuan Province Ecological Environmental Monitoring Station, Chengdu, 610091, China
• Institute of Solid Waste Treatment Technology, Sichuan Academy of Environmental Sciences, Chengdu, 610041, China
• Hongta Tobacco (Group) Co., Ltd, Yuxi, 653100, China

Short Summary

This paper introduces STAR, a compact near-infrared (NIR)-based device for precise soil texture classification, which employs a domain-adaptive deep learning strategy to overcome limitations in model generalization and data dependency. Validated with local soil samples, STAR achieved an 85.0 % overall accuracy and successfully identified unseen soil texture types, demonstrating robust generalization for practical soil analysis.

Objective

• To develop a compact and intelligent near-infrared (NIR)-based device (STAR) for precise soil texture classification, addressing challenges of constrained model generalization, high dependency on annotated data, and cross-domain inconsistencies in NIR spectroscopy.

Study Configuration

• Spatial Scale: Local soil samples from Sichuan Province, China.
• Temporal Scale: Not explicitly defined for the model development, but focuses on current soil conditions.

Methodology and Data

• Models used: Soil Texture Analysis Recognizer (STAR) device, incorporating a domain-adaptive deep learning strategy with Transfer Multiplicative Scatter Correction (TMSC) for spectral preprocessing and Selective Enhanced Transfer Adaptive Boosting (SETAB) framework.
• Data sources: Near-infrared (NIR) spectra collected from local soil samples.

Main Results

• STAR achieved an overall classification accuracy of 85.0 % with a Kappa coefficient of 0.78 across 5 local soil texture classes.
• It successfully identified previously unseen soil texture types: loamy sand with 100.0 % accuracy and sandy loam with 66.7 % accuracy.

Contributions

• Introduction of STAR, a novel compact and intelligent NIR-based device for precise soil texture classification.
• Development of a domain-adaptive deep learning modeling strategy, including Transfer Multiplicative Scatter Correction (TMSC) to reduce spectral distributional shifts and Selective Enhanced Transfer Adaptive Boosting (SETAB) to enhance model adaptability.
• Demonstration of robust generalization capability and practical utility for NIR-based soil property measurement, providing a scalable pathway for real-world applications.

Funding

• Not explicitly mentioned in the provided text.

Citation

@article{Luo2025STAR,
  author = {Luo, Yuchen and Zhang, Zeyuan and Liu, Siyu and Leng, Geng and Xu, Wenbo and Luo, Xuemei and Wang, Yuewu and Xie, Zhenwei and He, Leyun and Wang, Junwei and Tong, Hongjin and ZongZong, Nima and Fu, Wenbo},
  title = {STAR: Soil texture analysis recognizer integrating domain-adaptive transfer learning with NIR spectroscopy},
  journal = {Journal of Environmental Management},
  year = {2025},
  doi = {10.1016/j.jenvman.2025.128378},
  url = {https://doi.org/10.1016/j.jenvman.2025.128378}
}