Monitoring soil organic carbon in alpine soils using in situ vis‐NIR spectroscopy and a multilayer perceptron

• Published in: Land degradation & development Vol. 31; no. 8; pp. 1026 - 1038
• Main Authors: Chen, Songchao, Xu, Dongyun, Li, Shuo, Ji, Wenjun, Yang, Meihua, Zhou, Yin, Hu, Bifeng, Xu, Hanyi, Shi, Zhou
• Format: Journal Article
• Language: English
• Published: Chichester, UK John Wiley & Sons, Ltd 15-05-2020; Wiley Subscription Services, Inc; Wiley
• Subjects: Algorithms; Climate change; Cores; deep learning; Ecological monitoring; Environmental changes; Environmental monitoring; hyperparameter optimization; Land use; Life Sciences; Machine learning; Multilayer perceptrons; Optimization; Organic carbon; Organic soils; Parameters; Performance evaluation; proximal soil sensing; Qinghai‐Tibet Plateau; Regression models; Soil dynamics; soil monitoring; Soil properties; Soil quality; Soils; Spectra; Spectroscopy; Spectrum analysis; Tibet; proximal soil sensing; deep learning; carbone organique; soil sciences; carbone organique du sol; sciences du sol; hyperparameter optimization; soil monitoring; Qinghai‐Tibet Plateau; organic carbon; carbone du sol
• ISSN: 1085-3278; 1099-145X
• DOI: 10.1002/ldr.3497

Soil quality in alpine ecosystems requires regular monitoring to assess its dynamics under changes in climate and land use. Visible near‐infrared (vis‐NIR) spectroscopy could offer an option, as sampling and transporting large numbers of soil samples in the Qinghai‐Tibet Plateau is extremely difficult. However, the potential for in situ vis‐NIR spectra and the optimal algorithms need to be defined in this region. We have therefore evaluated the performance of a deep learning method, multilayer perceptron (MLP), for in situ spectral measurement of soil organic carbon (SOC) with in situ vis‐NIR spectroscopy in southeastern Tibet, China. A total of 39 soil cores (maximum depth 1 m), including 547 soil samples taken from each 5‐cm depth interval, were collected. The spectra were also measured at each 5‐cm depth interval accordingly. After spectral preprocessing, 4,096 MLP models were generated by taking all the combinations from six parameters defined in the MLP. The 10‐fold‐core cross‐validation showed that MLP had a good performance for in situ SOC prediction, and the best MLP model had an R2 of .92, which were much better than those of the partial least squares regression model (R2 = .80). The results also suggested that the number of epochs, number of neurons, and dropout rate were the most important parameters in the MLP model. We concluded that in situ vis‐NIR spectroscopy coupled with an MLP model has high potential for large‐scale SOC monitoring in the Qinghai‐Tibet Plateau. Our results also provide a reference for rapid hyperparameter optimization using MLP for future soil spectroscopic modeling.
Highlights
We evaluated the in situ measurement of SOC using vis‐NIR spectra.
A multilayer perceptron was used to predict SOC in alpine soils.
Hyperparameter optimization was conducted by grid searching.
A multilayer perceptron had good performance for in situ SOC prediction.
The most vital parameters for a multilayer perceptron model were identified.