关键词: GF-1数据, 森林地上生物量, 混合特征选择, 机器学习, 二阶段优化策略

Abstract: Objective In this study, a two-stage remote sensing feature selection method combined with multiple machine learning models was proposed to explore an efficient model for estimating forest aboveground biomass (AGB) using domestic GF-1 satellite data, thereby providing scientific reference for the application of GF-1 satellite data in forestry monitoring.Method A total of 44 remote sensing feature variables, such as spectral features, texture features, and vegetation indices, were first extracted from the GF-1 data. Then a two-stage feature selection strategy was proposed and then applied for feature optimization. In the first stage, the filtered method (Relief) and embedded method (Lasso, RF) were used for initial screening, and in the second stage, the feature subset was further optimized by the wrapper method (RFE) to enhance the model estimation capability. Finally, 5 machine learning models including the K nearest neighbor (KNN), support vector regression (SVR), random forest (RF), gradient boosting (GBRT), gradient boosted regression tree (GBRT) and extreme gradient boosting tree (XGBoost) models were selected and applied for AGB estimation. They were applied for evaluating the match between the different feature selection methods and models and their effects on forest AGB estimation accuracy. According to the results, the optimal strategy was explored for estimating for forest AGB based on GF-1 data.Result 1) The Relief-RFE method achieved the best results in AGB inversion using the XGBoost model with R2 = 0.811, RMSE = 8.45 t·hm^–2, rRMSE = 23.39%. 2) The green light band texture feature (mea-b2) was able to capture spatial distribution information of forest surface cover, such as canopy density, tree distribution, and other spatial structural changes of the forest; The spectral characteristics of the blue light band (b1) were sensitive to changes in chlorophyll and leaf pigment concentrations, and was able to characterize vegetation health status and growth stage information. In the two-stage feature selection of various methods, both of the above two features were selected as core features for forest AGB estimation. 3) Feature selection significantly improved model performance, with XGBoost showing a more pronounced improvement compared to KNN, SVR, RF, and GBRT.Conclusion By comparing the application of different feature selection methods combined with machine learning model in forest AGB estimation, the results have demonstrated the effectiveness of the proposed two-stage hybrid feature selection strategy in forest AGB estimation based on GF-1 data, especially combining it with the XGBoost model, a highly accurate and robust forest AGB estimation strategy can be constructed.

Key words: GF-1 data, forest above-ground biomass (AGB), hybrid feature selection, machine learning, two-stage optimization strategy