Space-time mapping of soil organic carbon through remote sensing and machine learning

Abstract

Pedosphere is the largest terrestrial carbon reservoir. Soil organic carbon (SOC) is a critical attribute for soil quality and crop productivity, being directly linked to climate change mitigation and food security. Brazil boasts a significant agricultural production area and substantial potential for carbon sequestration. Nevertheless, the spatial-temporal distribution of SOC across the country is poorly understood, hindering the implementation of low-carbon agriculture public policies. We aimed to map the spatio-temporal distribution of SOC at from 0.00 to 0.20 cm depth over two periods. We assessed the SOC variation over seven years, generating a time series with five periods, obtaining the average SOC values. The Cubist algorithm was used to calibrate two short period (two years) and a long period (seven years/all period) models for SOC spatial prediction. Remote sensing data and soil particle size distribution maps were used as environmental covariates. We found in validation R2 values of 0.47 and 0.25 for short period models, and 0.34 for the long period model. The SOC content decreased by 54.97 % in the area according to the mapping by short period models and 53.72 % according to mapping by the long-period model. The predicted maps showed the same trend of the database (soil samples with observed SOC values) for the study areas using both short period and long period models.