Integration of machine learning and process‐based model outputs via ensemble Kalman filter enhanced space–time modelling of soil organic carbon in a highly human impacted area

Abstract

Listen

Translate article

AbstractAccurate prediction of soil organic carbon stock (SOCS) dynamics in areas with intensive human activities is crucial for developing sustainable soil management practices and climate change mitigation strategies. This study investigated the spatiotemporal dynamics of SOCS by collecting a total of 1219 topsoil samples in southern Jiangsu Province of China in 1980, 2000 and 2015, and compared the performance of three predictive models: random forest (RF), RothC, and a hybrid model of RF‐RothCEnKF. The hybrid model integrated outputs from the process‐based RothC model and the data‐driven RF model using the Ensemble Kalman Filter (EnKF) for sequential model state updates. Results showed that the three models presented similar spatial patterns of SOCS from 1980 to 2015, with relatively higher SOCS mainly distributed in the areas surrounding Taihu Lake. The mean SOCS change rates estimated by the RF‐RothCEnKF model represented an overall net increase of 0.04 t C ha−1 yr.−1 during that period. The RF‐RothCEnKF model exhibited high prediction accuracy, with an R2 of .52, a mean absolute error (MAE) of 7.38 t C ha−1, and a root mean square error (RMSE) of 9.13 t C ha−1 in 2015. This highlighted the RF‐RothCEnKF's ability to enhance performance when the individual RF model (R2 = .47, MAE = 7.66 t C ha−1, and RMSE = 9.42 t C ha−1) and the RothC (R2 = .13, MAE = 8.77 t C ha−1, and RMSE = 10.87 t C ha−1) fell short. Our findings may not only provide a framework for integrating process‐based and machine learning models to enhance the accuracy and adaptability of SOCS modelling in areas affected by intensive human activities, but also offer some guidance for developing sustainable agricultural practices and carbon management strategies in complex environmental settings.