Enhancing soil organic carbon mapping in Northeast China by incorporating laboratory-based soil hyperspectral measurements into spatial covariates and machine learning.

Information on the spatial distribution of soil organic carbon (SOC) is essential for environmental monitoring and assessment. However, limitations of commonly used environmental covariates and satellite multispectral data constrain the accurate prediction of SOC at large spatial scales. In this study, 647 topsoil (0-20 cm) samples and the corresponding Vis-NIR hyperspectral data were collected from croplands in Northeast China. Machine learning (ML) algorithms, including random forest (RF), extreme gradient boosting (XGBoost) and light gradient boosting machine (LightGBM), were combined with environmental covariates, satellite multispectral covariates, and hyperspectral-informed intermediate variables derived by spatializing the principal components of laboratory-based hyperspectral measurements to predict SOC across croplands in Northeast China. The results showed similar spatial patterns of SOC across the ML models, with high SOC contents in the north and northeast and low contents in the west and southwest. Incorporating multisource satellite data, spectral indices, and moving-window averages of satellite multispectral data improved the prediction performance of SOC, with the R2 of RF model (R2 = 0.55) increased by 11%, in comparison with those using satellite multispectral data alone (R2 = 0.44). While combining environmental covariates and the developed hyperspectral-informed intermediate variables in the RF model yielded the highest R2 of 0.61. Although environmental covariates were the most important predictors of SOC (e.g., mean annual temperature accounts for 32% of the total variable importance), the developed hyperspectral-informed intermediate variable PC2 accounted for the additional 16% importance of the SOC prediction. These results highlight the importance of incorporating laboratory-based soil hyperspectral information into spatial covariates to improve large-scale SOC mapping.