Remote estimation of soil organic carbon under different land use types in agroecosystems of Eastern China

Abstract

Soil organic carbon (SOC) plays a key role in soil function, ecosystem services, and the global carbon cycle. Large SOC stocks accumulate in agroecosystems, but the estimates of SOC distribution and magnitude in agroecosystems still have large uncertainties in global products. Based on multiple high-resolution environmental variables (terrain derived from digital elevation model, climate and organism from Landsat 8 OLI) and 369 observed soil samples from 2019 to 2021, two machine learning methods, random forest (RF) and support vector machine (SVM) models, were used to estimate the content and spatial distribution of SOC in agroecosystems of Eastern China. We proposed a hybrid model in which the optimal predictors of SOC were selected for different land use types (cropland, planted forest and grassland), and it aimed to improve the performance of machine learning. Our results showed that (1) the hybrid models performed better than the global models, and the Random forest-Hybrid (RF-Hybrid) model led to the highest prediction accuracy, with a validation R2 of 0.65 and RMSE of 5.76 g kg−1. (2) The SOC content in the agroecosystems of Eastern China among different land use types ranked in the following order: planted forest (19.92 g kg−1) > cropland (17.51 g kg−1) > grassland (17.30 g kg−1). However, in North China, the SOC content in cropland (14.04 g kg−1) was much lower than that in planted forest (17.21 g kg−1) and grassland (17.07 g kg−1), which was caused by excessive mechanized operation. (3) The agroecosystem in Northeast China presented the highest mean SOC content (26.55 g kg−1) due to low temperature. (4) Our estimations (R2 = 0.53, RMSE = 6.74 g kg−1, 30-m resolution) were more detailed and precise than the existing global SOC maps (Soil Grids: R2 = 0.26, RMSE = 12.44 g kg−1, 250-m resolution; HWSD: R2 = 0.06, RMSE = 41.07 g kg−1, 1000-m resolution). The results may improve the accuracy of agroecosystem carbon mapping and contribute to SOC assessments in agricultural ecosystems.