Machine learning soil-environmental impacts on agroecosystems for relating microbial biomass to soil carbon sequestration

Abstract

Background and AimsStoring carbon (C) within soils is significant for maintaining soil-health and reinforces the feedback loop of C loss from soils as CO2 to the atmosphere. Seasonal variation with increased temperatures and inconsistent precipitation as climate change consequences also affect the soil C-sequestration process globally. Soil-health management practices (SHMPs) such as cover crops, crop residues and manures increase organic components as well as soil-organic C (SOC) pool in an agroecosystem. While, soil microbial-biomass (SMB) which is considered as a soil-health metric to understand microbial community response, is still not modelled to relate with SOC and seasonal impacts to identify suitable SHMPs. MethodsCover crops followed by 100% residue addition and combinations of manure +organic-fertilizer were the SHMPs for a winter-wheat system in our field study. Seasonal data regarding SOC, nitrogen, SMB-C and labile-C content of soils were used to machine-learn the system and understand the influence of different drivers on SMB-C. ResultsThe test models based on ‘Multivariate Linear Regression’ could explain 70% of the variability and predicted seasonal-variation as a dominant variant followed by SHMPs and soil-moisture. AdaBoost and Random Forest Models performed better than others if ‘Ensemble Learning’ was used. ‘Feature Importance’ predicted labile-C and aboveground-biomass as the two most important drivers impacting SMB-C. ConclusionsEnsemble Learning’ method of Machine-Learning could be successfully implied to understand the SMB-C in an agroecosystem and set benchmark-strategies for soil-health improvement. 50% manure+ 50% fertilizer with crop-residue could be recommended for maximum labile-C and SOC in surface soil-layers.