Accurate and efficient determination of stellar atmospheric parameters (T _eff , logg, [Fe/H]) is essential for large-scale sky survey to conduct Galactic archeology and stellar evolution history. This paper proposes a novel data-driven model based on statistical features and Catboost algorithm (SCDD). The model extracts the statistical features of the spectra by windows, and constructs a nonlinear mapping based on the Catboost algorithm between the spectral features and the stellar parameters. After being trained using LAMOST DR6 spectral data set, the SCDD showed excellent results in the estimation of the stellar parameters. In the condition of that the g-band signal-to-noise ratio (S/Ng) is higher than 100, the root mean square errors (RMSEs) of T _eff, logg, and [Fe/H] are 36 K, 0.077 dex and 0.037 dex, respectively. Compared with the StarNet and Cannon2 models, SCDD performs better in mean absolute error and RMSE, which proves its good fitting ability. In addition, this paper compares the parameters estimated by the SCDD with those of APOGEE. The results are in good agreement, which shows that the model we proposed is reliable.

准确有效地确定恒星大气参数（T _eff , log g , [Fe/H]）对于进行大规模天空调查以进行银河考古和恒星演化历史至关重要。本文提出了一种基于统计特征和 Catboost 算法 (SCDD) 的新型数据驱动模型。该模型通过窗口提取光谱的统计特征，并基于Catboost算法构建光谱特征与恒星参数之间的非线性映射。在使用 LAMOST DR6 光谱数据集进行训练后，SCDD 在估计恒星参数方面表现出优异的结果。在g波段信噪比 (S/N g ) 高于 100，T _eff、log g和 [Fe/H] 的均方根误差 (RMSE)分别为 36 K、0.077 dex 和 0.037 dex。与 StarNet 和 Cannon2 模型相比，SCDD 在平均绝对误差和 RMSE 方面表现更好，证明了其良好的拟合能力。此外，本文将 SCDD 估计的参数与 APOGEE 估计的参数进行了比较。结果吻合较好，说明我们提出的模型是可靠的。