A comparison of multiple deep learning methods for predicting soil organic carbon in Southern Xinjiang, China.

[Display omitted] • Potential of spectral estimation of SOC was explored in the arid and semi-arid area. • Deep learning outperformed Random Forest in SOC prediction. • Combinational use of feature selection and deep learning can improve SOC prediction. • The sensitivity of various deep learning algorithms depended on the sample size. Soil organic carbon (SOC) plays an important role in soil functioning and also global C balance. Visible-near-infrared (Vis-NIR) spectroscopy can be regarded as a cost-effective alternative to monitor the SOC content. Previously, application of Vis-NIR spectroscopy in the quantitative estimation of SOC in arid and semi-arid regions has received relatively little attention. Here, three different sample sizes of dataset (i.e., 330, 660, and 990) with SOC contents and Vis-NIR spectroscopy measured in the laboratory were obtained from Southern Xinjiang, China. Eight feature selection methods, including Interval Random Frog (IRF), were used to extract the optimal spectral feature subset. Six deep learning (DL) algorithms (e.g., Long Short-Term Memory Neural Networks, LSTM; Deep Belief Networks, DBN) and one machine learning method (Random Forest, RF) were utilized to relate SOC to spectral predictors. The overall objective of this work was to compare the predicted potentials of seven modeling algorithms combined with eight feature selection methods for spectral prediction of SOC. In addition, this paper also investigated the influence of different calibration sample size on the final modeling accuracy for SOC. Results indicated that the DL algorithms outperformed RF for SOC prediction. Among the six DL approaches, the LSTM model performed the best, while the DBN model performed the worst. The one-dimensional-Convolutional Neural Network (1D-CNN), 2D-CNN, Recurrent Neural Network, and DBN algorithms were sensitive to different sample sizes. For the largest dataset (i.e., 990 samples), four of the eight feature selection methods combined with the DL algorithms could improve the prediction for SOC, relative to the corresponding full-spectrum DL models. Among all models developed for SOC, the IRF-LSTM model achieved the optimal prediction, with the validation R2 of 0.89. Our findings provided both theoretical and technical guidance for the spectral estimation of SOC with the relatively low values in arid and semi-arid area. [ABSTRACT FROM AUTHOR]