Potential of globally distributed topsoil mid-infrared spectral library for organic carbon estimation

Accurate monitoring of soil organic carbon (SOC) is critical for sustainable management of soil for improving its quality, function, and carbon sequestration. As a nondestructive, efficient, and low-cost technique, mid-infrared (MIR) spectroscopy has shown a great potential in rapid estimation of SOC, despite limited studies of the global scale. The objective of this work was to use a globally distributed topsoil MIR spectral library with 33,039 samples to predict SOC using different modeling methods. Effects of nine fractional-order derivatives (FODs) on the predicted accuracy of SOC were evaluated using four regression algorithms (i.e., ratio index-based linear regression, RI-LR; partial least squares regression, PLSR; Cubist; convolutional neural network, CNN). Squareroot transformation to SOC data was performed to minimize the skewness and non-linearity. Results indicated FOD to capture the subtle spectral details related to SOC, leading to improved predictions that may not be possible by the raw absorbance and common integer-order derivatives. Concerning the RI-LR models, the optimal validation result for SOC was obtained by 0.75-order derivative, with the ratio of performance to inter-quartile distance (RPIQ) of 1.85. Regarding the full-spectrum modeling for SOC, the CNN outperformed PLSR and Cubist models, irrespective of raw absorbance or eight FODs; the best-performing CNN model was achieved by 1.25order derivative (validation RPIQ = 6.33). It can be concluded that accurate estimation of SOC using large and diverse MIR spectral library at the global scale combined with deep-learning CNN model is feasible. This global-scale database is extremely valuable for us to deal with the shortage of soil data and to monitor the soils at different geographical scales.

Automated summary

This study used a globally distributed topsoil MIR spectral library to predict SOC using different modeling methods. The results showed that fractional-order derivatives (FODs) improved the prediction accuracy of SOC. The 0.75-order derivative was found to be optimal for ratio index-based linear regression (RI-LR) models, while the convolutional neural network (CNN) model outperformed other models for full-spectrum modeling.