Integrating AlphaEarth, Sentinel-1 and Sentinel-2 to estimate aboveground carbon stock in the Atlantic Forest

Description

Estimating carbon stocks in tropical forests is essential for monitoring ecosystems and developing climate change mitigation strategies. The objective of this study was to develop and compare Random Forest models for estimating above-ground carbon stocks in the Atlantic Forest, in Brazil, by integrating variables derived from Sentinel-1 (S1) and Sentinel-2 (S2) satellites and the AlphaEarth (AE) model. Initially, predictor variables were subjected to correlation analysis and the Recursive Feature Elimination (RFE) method to select the optimal subset of variables. Subsequently, models were fitted using different datasets. The model consisting of variables combining S1, S2, and AE yielded the best results, with a coefficient of determination (R²) of 0.70, a Root Mean Square Error (RMSE) of 10.80 Mg.ha⁻¹, and a Mean Absolute Error (MAE) of 9.22 Mg.ha⁻¹, estimating carbon stocks between 22 and 100 Mg.ha⁻¹. The integration of variables from multiple sources improved the - predictive performance of the models. AlphaEarth representations accounted for 40% of the variables selected in the best-performing model, demonstrating its potential for applications in carbon estimates in tropical forests. The results show that the integration of conventional geospatial representations and the AlphaEarth Foundation model is a promising approach for mapping carbon stocks and supporting forest conservation and climate change mitigation efforts

Institutions

• Universidade Federal de Lavras

  Minas Gerais, Lavras

Categories

Funders

• Coordenação de Aperfeicoamento de Pessoal de Nível Superior

  Ministry of Education

  Federal District, Brasília

Licence