A multi-source approach combining GEDI LiDAR, satellite data, and machine learning algorithms for estimating forest aboveground biomass on Google Earth Engine platform

Abstract

Monitoring changes in carbon stocks through forest biomass assessment is essential for understanding the carbon cycle. However, acquiring timely and reliable ground measurements poses challenges in creating spatially continuous maps of forests aboveground biomass density (AGB). This study presents a novel approach that combines open-access Global Ecosystem Dynamics Investigation (GEDI) LiDAR data with satellite observation datasets to estimate and map the AGB and to enhance the accuracy and availability of biomass assessments, ultimately contributing to better insights into forest carbon dynamics. The primary objectives of this study are to: 1) evaluate the potential of combining optical and microwave satellite observations for estimating forest aboveground biomass, 2) compare the performance of different machine learning (ML) algorithms—Random Forest (RF), Classification and Regression Trees (CART), Gradient Tree Boosting (GTB), and Support Vector Machine (SVM)—in estimating forest aboveground biomass using the Google Earth Engine platform. Among the selected models, the GTB model demonstrated the best performance, achieving an R² value of 0.77, MAE of 22.27 Mg/ha, and RMSE of 37.78 Mg/ha using Sentinel-2 bands and topographic derivatives. The RF model achieved an R² value of 0.74, value of 0.74, MAE of 31.34 Mg/ha, and RMSE of 39.93 Mg/ha using Sentinel 2 bands, Sentinel 2 Vegetation Indices, Topographic derivatives, and Canopy's height. The RF model followed closely with an R² value of 0.74, value of 0.74, MAE of 31.34 Mg/ha, and RMSE of 39.93 Mg/ha using Sentinel 2 bands, Sentinel 2 Vegetation Indices, Topographic derivatives, and Canopy's height. The CART model achieved an R² value of 0.68, MAE of 33.16 Mg/ha, and RMSE of 39.97 Mg/ha using Sentinel 2 bands, Sentinel 2 Vegetation Indices, Sentinel 1 bands, Sentinel 1 Vegetation Indices, Topographic derivatives, and Canopy's height. Meanwhile, the SVM model performed the worst, with a maximum R² of 0.37, MAE of 55.17 Mg/ha, and RMSE of 73.48 Mg/ha using Sentinel 2 bands, Sentinel 2 Vegetation Indices, Sentinel 1 bands, and Sentinel 1 Vegetation Indices. The results highlighted that integrating Sentinel-2 bands, vegetation indices, topographic data, and canopy height significantly improved model performance. The study emphasizes the importance of multi-source data for enhancing AGB estimation and suggests that the GTB model offers the most reliable predictions. These findings can guide future research in forest biomass mapping using machine learning and remote sensing techniques.