森林地上生物量（Above Ground Biomass， AGB）是衡量森林生态系统生产能力的重要参考指标，也是研究地表碳循环和碳平衡的重要组成部分。立足国内高分三号（GF-3）SAR数据，探索不同类型反演模型的适宜性，以提高森林地上生物量的反演精度有着重要意义。以云南省昆明市宜良县花园林场小哨林区西南地区典型针叶林为研究对象，以GF-3 SAR数据为数据源，结合地面样地调查数据将GF-3 SAR数据的4个通道极化后向散射系数和极化分解特征作为森林地上生物量的建模因子；使用参数模型中的多元线性逐步回归（Multivariable Linear Stepwise Regression， MLSR）算法及非参数模型中的K最近邻（K-Nearest Neighbor Method， K-NN）、支持向量回归（Support Vector Regression， SVR）和随机森林（Random Forest， RF）共4种算法，对该研究区域森林AGB进行了反演；并采用皮尔逊相关系数（R²）、均方根误差（RMSE）及总精度（Acc.）3个指标对4种模型的反演结果精度进行了分析。得出结果：多元线性逐步回归模型反演结果的R²为0.37、RMSE为20.70 t/hm²、总精度Acc.为61.85%；K-NN模型R²为0.34、RMSE为20.29 t/hm²、总精度Acc.为62.60%；SVR模型R²为0.33、RMSE为20.95 t/hm²、总精度Acc.为61.39%；RF模型R²为0.35、RMSE为20.40 t/hm²、总精度Acc.为62.40%。通过对比分析形成以下结论：①4种模型中MLSR算法精度相对最高，较适宜于本研究区以云南松为优势树种的针叶林森林AGB反演；②非参数模型中RF算法反演精度略高，但略低于MLSR算法的精度指标；4种模型估测精度总体上偏低，可能与研究区域地形起伏造成的阴影叠掩及抽样调查的样地数据在异质性和代表性上表现欠佳有关。

Forest Aboveground biomass is not only an important reference index to measure the productivity of forest ecosystem， but also an important part to study the surface carbon cycle and carbon balance. Based on domestic GF-3 SAR data， the suitability of different inversion models was explored to improve the inversion accuracy of Forest Aboveground biomass. The typical coniferous forest in southwest of Xiaoshao forest farm， Yiliang County， Kunming City， Yunnan Province was selected as the research object. GF-3 SAR data was used as the data source， and the polarization backscattering coefficient and polarization decomposition characteristics of four channels of GF-3 SAR data were used as the modeling factors of forest biomass. The parametric multivariate linear regression model and the nonparametric k-Nearest Neighbor method （k-NN）， Support Vector Regression （SVR） and Random Forest （RF） models were used to retrieve the Above Ground Biomass （AGB） of forest land in the study area.Pearson correlation coefficient （R²）， Root Mean Square Error （RMSE） and total accuracy （acc.） were used to assess the accuracy of the four models： R² of the multiple linear stepwise regression model was at 0.37， with a RMSE at 20.70 T / hm²， and the total accuracy （acc.） was 61.85%； The k-NN model had a R² of 0.34， with a RMSE at 20.29 T / hm² and the total accuracy acc. at 62.60%； The R² of SVR model was 0.33， the RMSE was 20.95 T / hm²， and the total accuracy was 61.39%； The RF model R² was 0.0.35， the RMSE was 20.40 T / hm²， with the total accuracy acc. at 62.40%. This paper draws the following conclusions. ①The multiple linear stepwise regression algorithm， which belongs to the parametric model， has the highest accuracy， and is more suitable for AGB inversion of coniferous forest with Pinus Yunnanensis as the dominant tree species in this study area. ②In the nonparametric models， the SVR inversion accuracy is slightly higher， but generally lower than the parametric model. The accuracy of the four models is generally low， which may be related to the shadow overlay caused by the topographic relief in the study area and the poor heterogeneity and representativeness of the sample plot data in the sampling survey.