Prediction of Soil Organic Carbon Content Using Sentinel-1/2 and Machine Learning Algorithms in Swamp Wetlands in Northeast China

Abstract

Soil organic carbon (SOC) is a sensitive indicator of climate change, and small changes in the soil carbon pool will affect the carbon balance. Accurate and robust SOC quantitative prediction is of great significance to studying the carbon budget of swamp wetlands and its response to climate change. In this study, a new framework was proposed and assessed for predicting the SOC content based on Sentinel-2 (S2), Sentinel-1 (S1), and the digital elevation model (DEM) together with the extreme gradient boosting with random forest (XGBRF) model. The determination coefficient (R2), root mean square error (RMSE), mean absolute error (MAE), and Lin's concordance correlation coefficient (LCCC) were applied to assess the performances of the models. The results revealed that the prediction performance of the XGBRF regression model was much better than that of extreme gradient boosting and random forest regression models. Compared with single sensor data, using multisensor data to predict the SOC content yielded more accurate results. The XGBRF model based on S1, S2, and DEM fusion yielded the highest prediction accuracy (R2_testing = 0.6639, RMSE = 1.3236 g/kg, MAE = 1.2546 g/kg, LCCC = 0.7621). Regarding the importance of the variables, the S1 and S2 features were major contributors to the SOC content prediction (41% and 52%, respectively), followed by the topographic variables extracted from the DEM (7%). The proposed framework can be used for SOC prediction based on a small sample dataset, and it provides a method for long-term and rapid monitoring of the SOC contents in wetlands.

Keywords

• Extreme gradient boosting with random forest (XGBRF)
• machine learning (ML)
• Sentinel-1/2
• soil organic carbon (SOC)
• swamp wetlands