Hyperspectral imaging for high-resolution mapping of soil carbon fractions in intact paddy soil profiles with multivariate techniques and variable selection

By Shengxiang Xu, Meiyan Wang, Xuezheng Shi. 2020

60 citations recorded by the source corpus.

Abstract Soil organic carbon (SOC) and its labile C fractions play a central role in soil quality and C cycles. This study aimed to investigate the potential of laboratory-based hyperspectral imaging (HSI) spectroscopy to predict and map SOC and its labile C fractions (e.g., dissolved organic C, DOC; readily oxidizable organic C, ROC; and microbial biomass C, MBC) in soil profiles with a high resolution. The HSI images were captured from 16 intact paddy soil profiles to a depth of 100 ± 5 cm from four typical parent materials. The linear (i.e., partial least squares regression, PLSR) and nonlinear (i.e., artificial neural networks, ANN; cubist regression tree, Cubist; Gaussian process regression, GPR; and support vector machine regression, SVMR) multivariate techniques were compared to assess their ability to map the soil C fractions in the profiles. A spectral variable selection technique (i.e., competitive adaptive reweighted sampling, CARS) was applied to these multivariate models (i.e., CARS-PLSR, CARS-ANN, CARS-Cubist, CARS-GPR, and CARS-SVMR). Overall, the results showed that the nonlinear models performed better than the PLSR models in most cases. All optimized multivariate models with CARS achieved prediction performances similar to the full spectrum models, with high Lin's concordance correlation coefficient (LCC) and low root mean square error (RMSE). CARS-SVMR used only 37–70 spectral variables and took less time-consuming on computations (

Subfields: Analytical Chemistry, Artificial Intelligence, Environmental Engineering

View Paper: https://www.semanticscholar.org/paper/965da091b1b25c88d76174c11f88377688cf3323