Ensemble and transfer learning of soil inorganic carbon with visible near-infrared spectra

• This study is the first attempt to use a stacking model to predict soil inorganic carbon. • The stacking model outperforms each individual model out of the 10 base models. • Our constructed stacking models show superior transfer capabilities. Soil inorganic carbon (SIC) dominates the soil carbon pools in semi-arid and arid areas globally. Variations in the SIC pool would substantially affect the atmospheric CO 2 concentrations. The rapid and accurate measurement of SIC content using visible near-infrared (Vis-NIR) spectroscopy is of high significance for the management of soil carbon pools in semi-arid and arid regions. Ensemble learning is a novel and advanced modeling approach. However, it has been applied less in soil spectroscopy, and its transfer capability has not been evaluated. Therefore, we hypothesized that the use of the ensemble technique could further SIC prediction accuracy and have a better model transfer capability. In this study, a stacking model was developed using 990 soil samples collected from the Alar Reclamation region in South Xinjiang, China. The stacking model consists of 10 base models (support vector machine (SVM), partial least squares algorithm (PLSR), multi-layer perceptron (MLP), etc.). Two strategies (hyperparameter-adjusted and −unadjusted) were used to transfer the model to other target areas including Shaya and Wensu Counties on the southern border of China. Our results demonstrate that the SIC content could be predicted accurately using the stacking models (R 2 p = 0.81). The stacking model outperformed all the individual models and significantly improved the prediction accuracy of SIC. The R 2 p of the stacking models improved by 0.05–0.21, and the root mean square error (RMSE P ) reduced by 0.33–1.44 g kg −1 . Additionally, the stacking models displayed superior model transfer capability. Compared with direct transfer, the stacking model with fine-tuning of the hyperparameters displayed better model stability and generalization. Moreover, the average R 2 p improved by over 0.09 compared with the stacking model with unadjusted hyperparameters. Overall, stacking ensemble learning is a potential method for predicting SIC with good transfer capabilities. Our results provide new tools and strategies for the accurate estimation of SIC in semi-arid and arid regions.