Abstract: Estimating stellar atmospheric physical parameters (effective temperature T_\mathrmeff, surface gravity \lg g, and metallicity Fe/H) is a primary task in stellar studies. To meet the demand for analyzing massive stellar spectral data, this study proposes the SCA-Net (Spectral Convolution-Attention Network) model based on publicly available low-resolution observed spectral data from LAMOST DR10 (Large Sky Area Multi-Object Fiber Spectroscopy Telescope Data Release 10). SCA-Net integrates mobile inverted bottleneck convolution, self-attention mechanisms, and multi-scale feature fusion to achieve precise estimation of stellar atmospheric physical parameters. Experiments were conducted on 28913 low-resolution observed spectra, with 8913 used for model training and 20000 for testing. The evaluation metric, Mean Absolute Error (MAE), yielded the following accuracies for the three stellar atmospheric physical parameters: 71.07 K (T_\mathrmeff), 0.042 (\lg \leftg/ \left(\rmcm\cdot \rms^-2\right)\right), and 0.040 dex (Fe/H). The code of this study is available at: https://github.com/HelKai528/SCA-Net.