Predictive Uncertainty Estimation of SOC with Ensemble Algorithms and Bayesian Optimization

• Published in: Eurasian soil science Vol. 58; no. 8
• Main Authors: Xuan-Linh Nguyen, Van Sang, Tran, Pham, Van-Manh, Nguyen, Quoc-Huy, Nguyen, Huu-Duy, Do, Nhung-Thi, Nguyen, Dinh-Hung, Bui, Quang-Thanh
• Format: Journal Article
• Language: English
• Published: Moscow Pleiades Publishing 01.08.2025; Springer Nature B.V
• Subjects: Bayesian analysis; Bayesian theory; Earth and Environmental Science; Earth Sciences; Estimation; Genesis and Geography of Soils; Geotechnical Engineering & Applied Earth Sciences; Independent variables; Land use; Machine learning; Mathematical models; Organic carbon; Probability distribution; Probability theory; Soil properties; Sustainable use; satellite data sets; soil organic carbon; Vietnam; machine learning
• ISSN: 1064-2293; 1556-195X
• DOI: 10.1134/S1064229324604293

Machine learning techniques have become popular in estimating soil organic carbon (SOC) due to their ability to handle complex, nonlinear relationships between soil properties and environmental variables. Previous studies focused on the point estimation of SOC rather than predicting a probability distribution. This paper evaluates the hybrid model that integrated the Bayesian optimizer into NGBoost using satellite data sets for SOC mapping. The investigation was carried out by obtaining about 250 soil samples in the Northeastern region of Vietnam, and the SOC content was analyzed in the laboratory using the Walkley-Black method. We obtained spectral indices from Sentinel 2A and Sentinel 1A as independent variables to predict the SOC contents at these sample locations. The hybrid models were trained in two successive stages: iteratively selecting hyper-parameters using Bayesian and regular model training using cross-validation methods. The tuned NGBoost performs well with RMSE = 2.0201, MAE = 1.5603, and R 2 = 0.674, and it was used to map SOC for a subset within the study area. The information generated from this study will help develop improved and more efficient monitoring techniques for SOC dynamics and promote sustainable land use and management.