Data-driven approach to robust spatio-temporal assessment of carbon fluxes using Earth observation and ground-based data

Abstract

Effective spatial monitoring of carbon fluxes is crucial for implementing climate change mitigation and adaptation measures. This study develops an advanced machine learning (ML) pipeline to assess integral carbon fluxes at regional scales using Earth observation data and ground-based measurements. We aimed to address main limitations of spatial ML assessments associated with ignorance of environmental processes’ physical nature. We propose a training pipeline ensuring prediction robustness and model generalization, introducing influential features and ground truth data selection strategy. This results in a robust mapping tool with uncertainty estimations, supported by Shapley values-based feature importance analysis for interpretability and physical meaning. Our approach utilizes data from 168 FLUXNET stations, NASA POWER meteorological reanalysis, and MODIS satellite observations to train a CatBoost gradient boosting model. The model achieves $R^2$ of 0.76 predicting monthly NEE values with high spatial–temporal coherence, opening possibilities for comprehensive terrestrial ecosystem carbon dynamics assessments.