Soil total nitrogen inversion and interpretability analysis using vis-NIR spectroscopy and transfer learning

The use of Vis-NIR Spectroscopy for soil component inversion has increased, driven by its advantages in non-destructive, large-scale monitoring. However, it often faces challenges in model generalization. Transfer learning, leveraging large existing soil sample datasets, is considered an effective s...

Full description

Saved in:
Bibliographic Details
Main Authors: Ping He, Yu Chen, Xingping Wen, Xiaohua Zhou, Zailin Chen, Zhongchang Sun, Xianfeng Cheng
Format: Article
Language:English
Published: Taylor & Francis Group 2025-08-01
Series:International Journal of Digital Earth
Subjects:
Online Access:https://www.tandfonline.com/doi/10.1080/17538947.2025.2528621
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849224318341349376
author Ping He
Yu Chen
Xingping Wen
Xiaohua Zhou
Zailin Chen
Zhongchang Sun
Xianfeng Cheng
author_facet Ping He
Yu Chen
Xingping Wen
Xiaohua Zhou
Zailin Chen
Zhongchang Sun
Xianfeng Cheng
author_sort Ping He
collection DOAJ
description The use of Vis-NIR Spectroscopy for soil component inversion has increased, driven by its advantages in non-destructive, large-scale monitoring. However, it often faces challenges in model generalization. Transfer learning, leveraging large existing soil sample datasets, is considered an effective solution to overcome the mentioned limitations. This study explores the use of transfer learning, utilizing the LUCAS database, to improve the accuracy of estimating soil total nitrogen (STN) with Vis-NIR spectroscopy in Gejiu, Yunnan, China, addressing challenges in model generalization. It compares spectral preprocessing methods (logR, SNV, MSC) and models (PLS, RF, ResNet) to assess their impact on inversion performance. SHapley Additive exPlanations (SHAP) is employed for model interpretability. Results show that transfer learning with the ResNet model significantly enhances STN inversion, particularly with MSC preprocessing, where the average R2 improves from 0.51 to 0.70. Among the models tested, ResNet with transfer learning outperforms others in accuracy. SHAP analysis identifies key wavelengths −2050, 2459, 2149, 2109, 2410, and 1470 nm – as crucial for predicting STN, which aligns with the observed correlations. This research validates the effectiveness of transfer learning on small datasets, offering a robust solution for STN inversion using Vis-NIR spectroscopy.
format Article
id doaj-art-b5bf1d3de37e4d8dbda9755c768b5a99
institution Kabale University
issn 1753-8947
1753-8955
language English
publishDate 2025-08-01
publisher Taylor & Francis Group
record_format Article
series International Journal of Digital Earth
spelling doaj-art-b5bf1d3de37e4d8dbda9755c768b5a992025-08-25T11:28:44ZengTaylor & Francis GroupInternational Journal of Digital Earth1753-89471753-89552025-08-0118110.1080/17538947.2025.2528621Soil total nitrogen inversion and interpretability analysis using vis-NIR spectroscopy and transfer learningPing He0Yu Chen1Xingping Wen2Xiaohua Zhou3Zailin Chen4Zhongchang Sun5Xianfeng Cheng6School of Fine Art and Design, Kunming University, Kunming, Yunnan, People’s Republic of ChinaInternational Research Center of Big Data for Sustainable Development Goals, Beijing, People’s Republic of ChinaFaculty of Land Resources Engineering, Kunming University of Science and Technology, Kunming, People’s Republic of ChinaYunnan Provincial Bureau of Geology and Mineral Exploration and Development Center Laboratory, Kunming, People’s Republic of ChinaSchool of Earth and Environmental Sciences, Yunnan Land and Resources Vocational College, Kunming, Yunnan, People’s Republic of ChinaInternational Research Center of Big Data for Sustainable Development Goals, Beijing, People’s Republic of ChinaSchool of Earth and Environmental Sciences, Yunnan Land and Resources Vocational College, Kunming, Yunnan, People’s Republic of ChinaThe use of Vis-NIR Spectroscopy for soil component inversion has increased, driven by its advantages in non-destructive, large-scale monitoring. However, it often faces challenges in model generalization. Transfer learning, leveraging large existing soil sample datasets, is considered an effective solution to overcome the mentioned limitations. This study explores the use of transfer learning, utilizing the LUCAS database, to improve the accuracy of estimating soil total nitrogen (STN) with Vis-NIR spectroscopy in Gejiu, Yunnan, China, addressing challenges in model generalization. It compares spectral preprocessing methods (logR, SNV, MSC) and models (PLS, RF, ResNet) to assess their impact on inversion performance. SHapley Additive exPlanations (SHAP) is employed for model interpretability. Results show that transfer learning with the ResNet model significantly enhances STN inversion, particularly with MSC preprocessing, where the average R2 improves from 0.51 to 0.70. Among the models tested, ResNet with transfer learning outperforms others in accuracy. SHAP analysis identifies key wavelengths −2050, 2459, 2149, 2109, 2410, and 1470 nm – as crucial for predicting STN, which aligns with the observed correlations. This research validates the effectiveness of transfer learning on small datasets, offering a robust solution for STN inversion using Vis-NIR spectroscopy.https://www.tandfonline.com/doi/10.1080/17538947.2025.2528621Soil total nitrogenVis-NIR spectroscopytransfer learningspectral preprocessingResNet model
spellingShingle Ping He
Yu Chen
Xingping Wen
Xiaohua Zhou
Zailin Chen
Zhongchang Sun
Xianfeng Cheng
Soil total nitrogen inversion and interpretability analysis using vis-NIR spectroscopy and transfer learning
International Journal of Digital Earth
Soil total nitrogen
Vis-NIR spectroscopy
transfer learning
spectral preprocessing
ResNet model
title Soil total nitrogen inversion and interpretability analysis using vis-NIR spectroscopy and transfer learning
title_full Soil total nitrogen inversion and interpretability analysis using vis-NIR spectroscopy and transfer learning
title_fullStr Soil total nitrogen inversion and interpretability analysis using vis-NIR spectroscopy and transfer learning
title_full_unstemmed Soil total nitrogen inversion and interpretability analysis using vis-NIR spectroscopy and transfer learning
title_short Soil total nitrogen inversion and interpretability analysis using vis-NIR spectroscopy and transfer learning
title_sort soil total nitrogen inversion and interpretability analysis using vis nir spectroscopy and transfer learning
topic Soil total nitrogen
Vis-NIR spectroscopy
transfer learning
spectral preprocessing
ResNet model
url https://www.tandfonline.com/doi/10.1080/17538947.2025.2528621
work_keys_str_mv AT pinghe soiltotalnitrogeninversionandinterpretabilityanalysisusingvisnirspectroscopyandtransferlearning
AT yuchen soiltotalnitrogeninversionandinterpretabilityanalysisusingvisnirspectroscopyandtransferlearning
AT xingpingwen soiltotalnitrogeninversionandinterpretabilityanalysisusingvisnirspectroscopyandtransferlearning
AT xiaohuazhou soiltotalnitrogeninversionandinterpretabilityanalysisusingvisnirspectroscopyandtransferlearning
AT zailinchen soiltotalnitrogeninversionandinterpretabilityanalysisusingvisnirspectroscopyandtransferlearning
AT zhongchangsun soiltotalnitrogeninversionandinterpretabilityanalysisusingvisnirspectroscopyandtransferlearning
AT xianfengcheng soiltotalnitrogeninversionandinterpretabilityanalysisusingvisnirspectroscopyandtransferlearning