Machine learning-based prediction of belowground biomass from aboveground biomass and soil properties

Abstract

Precise and accurate quantification of belowground biomass (BGB) is essential for understanding terrestrial carbon dynamics. Traditional methods for estimating BGB suffer from a number of disadvantages, including inability to resolve differences among plant species, high dependence on Diameter at Breast Height, and destructive sampling. To address these issues, we developed a novel machine learning framework to estimate grassland BGB by integrating vegetation and soil data from 294 plots on China's Loess Plateau. An ensemble model combining XGBoost regression, Gradient boosting regression, Ridge regression, and ElasticNet regression outperformed the individual models, achieving a training R² of 0.623 and a testing R² of 0.502, highlighting its superior ability to identify the complex dependencies of BGB. Integration of key features, including soil organic carbon, plant height, and aboveground biomass, significantly improved the predictive accuracy. Nonlinear BGB–environment interactions are commonly underrecognized in traditional models. The model presented herein advances our ability to assess underground carbon stocks and offers insights into the ecological strategies of grassland species under competitive light conditions. By revealing the multifaceted influences of soil and vegetation on BGB, our research refines the understanding of grassland carbon dynamics. This study marks a precedent for harnessing advanced machine learning in ecological modeling to facilitate more accurate predictions of global change.