Estimation of soil organic carbon content by Vis-NIR spectroscopy combining feature selection algorithm and local regression method

ABSTRACT Soil organic carbon (SOC) content is a critical parameter for evaluating soil health. However, high redundancy and invalid information in soil hyperspectral data can reduce the accuracy and stability of SOC prediction models. This study developed a global partial least squares regression (P...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Revista Brasileira de Ciência do Solo Ročník 47
Hlavní autoři: Baoyang Liu, Baofeng Guo, Renxiong Zhuo, Fan Dai
Médium: Journal Article
Jazyk:angličtina
Vydáno: Sociedade Brasileira de Ciência do Solo 01.01.2023
Témata:
local calibration
LUCAS 2015 database
soil property
variable selection
ISSN:1806-9657
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:ABSTRACT Soil organic carbon (SOC) content is a critical parameter for evaluating soil health. However, high redundancy and invalid information in soil hyperspectral data can reduce the accuracy and stability of SOC prediction models. This study developed a global partial least squares regression (PLSR) model and a local PLSR model for agricultural soils in the LUCAS 2015 database. Some variable selection methods were combined with the regression models and their effects on prediction accuracy were explored. In addition, when the genetic algorithm is utilized for spectral feature selection, we obtained a more representative spectral subset through a novel coding approach. The results illustrated that the best SOC estimation accuracy was achieved by the local PLSR combined with a coding-improved genetic algorithm (GA), with R2 of 0.71, RMSEP of 5.7 g kg-1, and RPD of 1.87. This study demonstrates that appropriate spectral band selection only slightly enhances the model performance of both global and local regressions, as PLSR models using the full spectrum show similar performance. Local PLSR models consistently outperform global ones using full spectrum or variable selection algorithms.
ISSN:1806-9657
DOI:10.36783/18069657rbcs20230067