| |
|
|
 | Acesso ao texto completo restrito à biblioteca da Embrapa Instrumentação. Para informações adicionais entre em contato com cnpdia.biblioteca@embrapa.br. |
|
Registro Completo |
|
Biblioteca(s): |
Embrapa Instrumentação. |
|
Data corrente: |
09/06/2025 |
|
Data da última atualização: |
09/06/2025 |
|
Tipo da produção científica: |
Artigo em Periódico Indexado |
|
Autoria: |
GUEDES, W. N.; BABOS, D. V.; FREITAS, V. S.; NEIVA, D. K.; SOUZA. D.; MARTIN NETO, L.; MILORI, D. M. B. P.; VILLAS-BOAS, P. R. |
|
Afiliação: |
UNIVERSIDADE DE SÃO PAULO; LADISLAU MARTIN NETO, CNPDIA; DEBORA MARCONDES BASTOS PEREIRA, CNPDIA; PAULINO RIBEIRO VILLAS BOAS, CNPDIA. |
|
Título: |
Robust Soil Total Carbon Prediction Using LIBS: Integrating Expert Knowledge with Machine Learning and External Dataset Evaluation. |
|
Ano de publicação: |
2025 |
|
Fonte/Imprenta: |
Atomic Spectroscopy, v. 46, n. 2, 2025. |
|
Páginas: |
141–149 |
|
DOI: |
www.at-spectrosc.com/as/article/pdf/2025018 |
|
Idioma: |
Inglês |
|
Conteúdo: |
Accurate quantification of total soil carbon (C) is challenging because of the complex composition of soil samples, which include varying levels of organic matter, oxides of Al and Fe, and diverse soil textures. These factors introduce variability and spectral interference, which hinder the measurement accuracy. In laser-induced breakdown spectroscopy (LIBS), the key C emission lines often overlap with those of Al and Fe, complicating the development of reliable calibration models. This study presents a comprehensive machine learning approach for total C quantification using LIBS data. It integrates emission lines identified through the SHapley Additive exPlanations (SHAP) algorithm and literature insights with spectral preprocessing techniques such as baseline correction, peak fitting, and peak area calculations. Although the SHAP algorithm effectively identified crucial emission lines, expert knowledge was vital in refining variable selection and excluding irrelevant data. Two datasets were utilized: one containing 1019 Brazilian soil samples representing various soil types and textures, and another containing 387 samples collected later that served as an external test set. The first dataset was divided into 713 samples for model training and 306 samples for validation. The external test set was used to evaluate the reproducibility and stability of the model. The integration of expert knowledge and machine learning produced a highly efficient model that reduced the number of variables without compromising accuracy. Among the tested algorithms, the Extra Trees regression model excelled, achieving R² values of 0.74, 0.74, and 0.72 for the training, validation, and external test sets, respectively, with root mean square error values of 5.1, 5.1, and 5.0 g kg⁻¹ of total C. This approach outperformed previous machine learning models in terms of accuracy and adaptability, demonstrating its potential applicability beyond soil analysis. This study highlights a pathway for creating robust predictive models across various domains by combining machine learning, spectral processing, and expert insights. MenosAccurate quantification of total soil carbon (C) is challenging because of the complex composition of soil samples, which include varying levels of organic matter, oxides of Al and Fe, and diverse soil textures. These factors introduce variability and spectral interference, which hinder the measurement accuracy. In laser-induced breakdown spectroscopy (LIBS), the key C emission lines often overlap with those of Al and Fe, complicating the development of reliable calibration models. This study presents a comprehensive machine learning approach for total C quantification using LIBS data. It integrates emission lines identified through the SHapley Additive exPlanations (SHAP) algorithm and literature insights with spectral preprocessing techniques such as baseline correction, peak fitting, and peak area calculations. Although the SHAP algorithm effectively identified crucial emission lines, expert knowledge was vital in refining variable selection and excluding irrelevant data. Two datasets were utilized: one containing 1019 Brazilian soil samples representing various soil types and textures, and another containing 387 samples collected later that served as an external test set. The first dataset was divided into 713 samples for model training and 306 samples for validation. The external test set was used to evaluate the reproducibility and stability of the model. The integration of expert knowledge and machine learning produced a highly efficient model that reduced the number ... Mostrar Tudo |
|
Palavras-Chave: |
C sequestration; Global carbon credit; Soil carbon. |
|
Categoria do assunto: |
-- |
|
Marc: |
null
Download
Esconder MarcMostrar Marc Completo |
|
Registro original: |
Embrapa Instrumentação (CNPDIA) |
|
|
Biblioteca |
ID |
Origem |
Tipo/Formato |
Classificação |
Cutter |
Registro |
Volume |
Status |
URL |
Voltar
|
|
|
| Registros recuperados : 1 | |
| 1. |  | MENDES, F. F.; TALES SOUZA TEIXEIRA; GUIMARAES, L. J. M.; MARTINS, K. G.; PARENTONI, S. N.; GUIMARAES, P. E. de O.; OLIVEIRA, K. G. de; REIS, D. P. dos; GOMES, P. H. P. Comportamento de híbridos top-crosses de milho sob estresse hídrico. In: CONGRESSO NACIONAL DE MILHO E SORGO, 29., 2012, Águas de Lindóia. Diversidade e inovações na era dos transgênicos: resumos expandidos. Campinas: Instituto Agronômico; Sete Lagoas: Associação Brasileira de Milho e Sorgo, 2012. p. 2801-2806. 1 CD-ROM.| Tipo: Artigo em Anais de Congresso |
| Biblioteca(s): Embrapa Milho e Sorgo. |
|    |
| Registros recuperados : 1 | |
|
| Nenhum registro encontrado para a expressão de busca informada. |
|
|
