Soil organic carbon stock changes over 40 years in Brazil: a country-scale assessment using soil science-informed machine learning

Brazil is at the forefront of both forest conservation and global food security, and thus, understanding the dynamics of soil organic carbon (SOC) is crucial for both. We developed a spatiotemporal model to predict national-scale changes in SOC stocks to 1 meter depth, over the last 40 years. We utilized a combination of static and dynamic environmental covariates along with more than 50,000 observation points to calibrate a machine learning model. The models achieved satisfactory accuracy, with R 2 values ranging from 0.48 to 0.88 in cross-validation and from 0.18 to 0.31 in external validation. Our findings revealed a significant reduction in SOC stocks in the North rainforest region (Amazon biome), with the largest increase in the Northeast drylands (Caatinga biome). Specifically, net SOC losses occurred in Amazon (-0.020 Pg yr -1 ), Atlantic Forest (-0.002 Pg yr -1 ), and Pampa (-0.0004 Pg yr -1 ). Net gains were found in the Cerrado (0.016 Pg yr -1 ), Caatinga (0.007 Pg yr -1 ), and Pantanal (0.001 Pg yr -1 ). Over the past 40 years, we estimated an absolute gain of 0.80 Pg C in Brazil. Land-use change from forests to anthropogenic uses was the primary driver of SOC stock loss, whereas conversion from pastures to croplands generally led to SOC gains. The efforts to combat climate change in Brazil require reducing deforestation and promoting sustainable agricultural intensification.