Optimizing soil carbon and CO2 emission prediction: a integration of machine learning algorithms and VIS/NIR data inputs

Climate change and the intensification of anthropogenic activities affect the dynamics of carbon in the soil, resulting in losses in stock and increased emissions of carbon dioxide (CO₂) into the atmosphere. The balance between the continuous input and output of carbon in the soil, as well as its sequestration from the atmosphere, contributes to the formulation of strategies to mitigate climate change and global warming. Our hypothesis is that it is possible to accurately predict soil CO₂ emissions and soil organic carbon (SOC) stock using hyperspectral sensing and machine learning (ML) algorithms. The objectives of the study were: (i) to predict CO₂ emission and SOC stock using hyperspectral sensor and ML algorithms; (ii) to identify algorithms and dataset inputs with the highest accuracy in predicting CO₂ emission and SOC stock. Samples were collected from three biomes in the State of Mato Grosso do Sul, Brazil: Cerrado, Atlantic Forest, and Pantanal. Within each biome, four land use classes were assessed: agriculture, pasture, eucalyptus plantations, and native vegetation. Data was collected from 100 points distributed in each area within each biome. In all sample point, carbono stock was measured in three deepths (0-10cm, 10-20 cm, and 20-40 cm). In situ soil CO₂ (FCO₂), temperature and moisture measurements were also performed. Hyperspectral data were collected by a sensor in each sample point and then the spectral bands used by MODIS sensor (seven bands) were obtained. Data were submited to ML analysis, in which two input configurations in the dataset were tested: using all the bands provided by the hyperspectral sensor (ALL) and using only the bands used by the MODIS sensor (B). Carbon stock at the three depths, FCO₂, soil temperature and moisture were used as output in datasets. ML models tested were: Artificial Neural Network (ANN), Decision Tree models REPTree and M5P, Random Forest (RF), Support Vector Machine (SVM), and a simple model used as control (ZeroR). Our findings reveal that the use of hyperspectral sensing and ML algorithms enables accurate prediction of CO₂ emissions and SOC stock. The choice of ML model for accuratelly predicting soil CO₂ emissions and carbon stocks is dependent on the input variables used in the datasets, in which SVM provides the highest accuracy when applied to all spectral bands, while RF shows better performance when using the MODIS bands. Therefore, the approach used here can provide large-scale estimates of soil CO₂ emissions and organic carbon stock.