Estimation of soil inorganic carbon with visible near-infrared spectroscopy coupling of variable selection and deep learning in arid region of China

Abstract

Soil inorganic carbon (SIC) is the primary component of the soil carbon pool in arid and semiarid regions and strongly impacts the global carbon cycle, ecosystem services, and soil functions. The global climate change and intensify of human activities, could substantially change SIC, which highlights the importance of monitoring SIC. Rapid and accurate estimation of SIC concentration is critical for soil inorganic carbon pool monitoring. Currently, visible near-infrared (Vis-NIR) spectroscopy is a promising technique for estimating SIC via a rapid and cost-effective manner. Thus, in this study, we collected 315 topsoil samples from the Alar Reclamation Area in South Xinjiang, China, and measured their Vis-NIR spectra and SIC content. Then, we used deep learning algorithms, including a one-dimensional convolutional neural network (1D-CNN), two-dimensional convolutional neural network (2D-CNN), long short-term memory network (LSTM), and deep belief network (DBN), combined with variable selection algorithms (particle swarm algorithm (PSO), interval random frog (IRF), competitive adaptive reweighting algorithm (CARS), ant colony algorithm (ACO), and iteratively retaining informative variables (IRIV) to estimate SIC. Results showed that all five variable selection algorithms could effectively extract the featured spectral information of SIC, and reduce the number of spectral variables by >97%, simplifying the model structure. The variable selection algorithm could markedly improve the SIC spectral estimation accuracy, and the corresponding estimation accuracy follows the order: IRF > IRIV > PSO > CARS > ACO. All four deep learning models have high prediction accuracy, and the modeling accuracy of each method follow the order: LSTM > 1D-CNN > 2D-CNN > DBN. The combined IRF and LSTM model achieved the highest estimation accuracy (R2 = 0.93, RMSE = 1.26 g kg−1 in the calibration dataset; R2 = 0.92, RMSE = 1.37 g kg−1 in the validation dataset). This study demonstrated that deep learning combined with variable selection algorithms can detect SIC content quickly and accurately.