Deep Learning Approach With Coupled Weighted Loss Function for Estimation and Prediction of Soil Organic Carbon in China

Abstract

Accurate estimation and projection of soil organic carbon (SOC) density is crucial for understanding the terrestrial carbon cycle and formulating carbon neutrality strategies. The increasing availability of SOC and related environmental data, coupled with advanced prediction models, has opened new opportunities for improving the accuracy of SOC (kg C m⁻²) predictions using data-driven methods. In this study, we developed a deep learning model TabTransformer_WT, by coupling the weighted mean squared error loss function with TabTransformer, to optimise estimation of surface (0–20 cm, SOC_0–20) and profile (0–100 cm, SOC_0–100) SOC in China. Using SOC observations and multi-source environmental covariates, we evaluated model performance through time-series-based 10-fold cross-validation across four periods (1979–1984, 2000–2004, 2005–2009 and 2010–2014) and compared it with machine learning and deep learning models (RF, SVR, CNN-1D, LSTM, RNN and TabTransformer). Our results indicate that TabTransformer_WT achieved the best prediction accuracy, with R² improvements of 8%–37% for SOC_0–20 and 6%–38% for SOC_0–100, and RMSE reductions of 0.31–1.07 and 0.99–2.39 kg C m⁻², respectively. We applied the model to evaluate historical and future spatiotemporal evolution of SOC_0–20 and SOC_0–100 in China. Historical analysis (1979–2023) showed China's soil acted as a carbon sink with annual growth rates of 45 Tg C year⁻¹ for surface and 33.37 Tg C year⁻¹ for profile SOC. Future projections using CMIP6 data revealed slow SOC accumulation under SSP1-1.9 but decreasing trends under SSP2-4.5 and SSP5-8.5 scenarios, with the 0–100 cm layer experiencing the greatest loss (−30.64 Tg C year⁻¹) under SSP5-8.5. This study provides a feasible method for large-scale SOC estimation and insights into SOC evolution under climate change.