Abstract: The increasing dimensionality of feature variables in remote sensing-based estimation of forest aboveground biomass (AGB) necessitates effective feature selection to enhance model accuracy. This study focuses on Nanning City as the research area, utilizing Sentinel-2 data as the remote sensing source. Spectral information from various bands, texture features, and additional factors such as elevation, slope, and aspect were extracted. Three feature selection methods including stepwise regression, bivariate correlation, and random forest were employed to identify modeling variables. Biomass estimation models were established based on CatBoost and random forest (RF) machine learning algorithms. Five-fold cross-validation was applied to evaluate model performance, and the best model was used to complete biomass mapping. The results indicated that among the three feature selection methods, the bivariate correlation method performed the best across three tree types: pine, eucalyptus, and broadleaf species. For Chinese fir forests, the random forest method showed superior performance. Specifically: For Chinese fir forests, the combination of the random forest feature selection method and the RF algorithm was optimal (R²=0.58, RMSE=8.53 Mg·hm^-2). For Masson pine forests, the bivariate correlation method combined with the RF algorithm was the best choice (R²=0.51, RMSE=11.10 Mg·hm^-2). For eucalyptus forests, the bivariate correlation method combined with the RF algorithm yielded the best results (R²=0.56, RMSE=14.91 Mg·hm^-2). For broadleaf forests, the bivariate correlation method combined with the RF algorithm also proved optimal (R²=0.35, RMSE=40.55 Mg·hm^-2). Feature selection methods have a significant impact on the predictive performance of models. Combining feature selection methods with ensemble machine learning algorithms is conducive to improving the estimation accuracy of AGB.

Key words: feature selection, random forest, CatBoost, forest aboveground biomass, machine learning