Forest AGB inversion based on GEDI-assisted Sentinel-1/2 data and GWO-PSO optimisation

Forest ecosystems are vital components of the terrestrial biosphere. Accurate estimation of forest above-ground biomass (AGB) is crucial for understanding global carbon stocks. To address the issue of hyperparameter sensitivity in traditional machine learning models for forest AGB estimation and the propensity of conventional search methods to become trapped in local optima, this study proposes a hybrid hyperparameter optimization strategy integrating the Grey Wolf Optimization (GWO) and Particle Swarm Optimization (PSO). The study was conducted in Shawan City, Xinjiang, using data from 177 ground-surveyed sample plots. It combined Sentinel-1 microwave remote sensing data, Sentinel-2 optical remote sensing data, and GEDI L2A and L4A data to extract multi-source features. Among these, GEDI data provided high-precision forest structure information, significantly enriching the vertical information content of the features. The original 85 features were reduced to 19 core features through a multi-stage process. This involved Pearson correlation analysis, RF importance ranking, Recursive Feature Elimination with Cross-Validation (RFECV), and Select From Model (SFM). This significantly increased the information density. For the two traditional machine learning methods, RF and XGBoost, an innovative GWO-PSO hybrid optimization strategy was proposed, combining the global search advantages of the GWO with the local optimization advantages of PSO. Comparative experiments demonstrated progressive improvements: Grid search-RF (R2 = 0.60), GWO-XGBoost (R2 = 0.65), and GWO/PSO-optimized RF-XGBoost ensembles (R2 = 0.73). The proposed GWO-PSO-RF-XGBoost hybrid model achieved superior accuracy (R2 = 0.84, RMSE = 69.46 t·hm− 2, rRMSE = 38.06%, MAE = 52.56 t·hm− 2), significantly outperforming all benchmarks. The research results indicate: 1) The integration of GEDI data provides richer structural information for forest AGB inversion; 2) The GWO-PSO hybrid optimization strategy significantly enhances model robustness; 3) The heterogeneous integration of RF and XGBoost fully leverages their complementary advantages, providing a reliable technical pathway for high-precision forest AGB inversion in arid regions.