Multi-sensor forest aboveground biomass estimation using GEDI, machine learning, and deep learning techniques

Traditional field-based methods for estimating forest aboveground biomass (AGB) are constrained by limited scalability and spatial coverage, prompting the need for advanced, cost-effective solutions. Recent advancements in machine learning (ML) and deep learning (DL), coupled with remote sensing technologies, particularly NASA’s Global Ecosystem Dynamics Investigation (GEDI) spaceborne LiDAR, enable large-scale AGB estimation by capturing vertical forest structure and integrating multi-source optical data. However, the comparative performance and transferability of ML/DL models, with and without GEDI integration, across varying sensor resolutions remain insufficiently explored. This study bridges this gap through a two-phase analytical framework: (1) evaluating six algorithms, including Random Forest (RF), Gradient Tree Boosting (GTB), Classification and Regression Trees (CART), Deep Neural Networks (DNN), Convolutional Neural Networks (CNN), and Recurrent Neural Networks (RNN), using high-resolution NAIP imagery (0.3 m) and diverse feature sets (spectral indices, texture metrics, and topography); and (2) integrating top-performing models with GEDI-derived metrics to assess AGB estimation efficiency across NAIP, Sentinel-2, Landsat-8, and MODIS data. Model performance was rigorously validated using R 2 , RMSE, and NRMSE metrics, supplemented by SHAP analysis for interpretability. Results identified RF as the most accurate algorithm (R 2 = 0.77, RMSE = 34.85 tonne per hectare (t ha −1 ), NRMSE = 5 %), surpassing DL models (e.g., DNN: R 2 = 0.65). Multi-feature integration (e.g., vegetation indices, texture) significantly improved predictions over single-feature approaches. GEDI integration revealed resolution-dependent efficacy, with NAIP achieving superior accuracy (R 2 = 0.72, RMSE = 24.13 t ha −1 ), while coarser-resolution sensors (Sentinel-2: R 2 = 0.52; Landsat-8: R 2 = 0.47; MODIS: R 2 = 0.16) exhibited progressive declines. SHAP analysis underscored elevation, vegetation indices, and texture metrics as critical predictors. This research establishes a scalable framework for AGB estimation that employs RF-driven workflows and high-resolution data fusion to improve global forest monitoring and carbon stock assessments.