Accurate extraction of ground points from LiDAR point clouds provides important data for understanding terrain changes and supports decision-making in ecological disaster prevention. Recently, deep learning models have been used to process point clouds directly, with a focus on semantic segmentation of urban scenes using RGB features. However, in complex terrains, using point cloud images to generate RGB features often introduces noise, making high-precision ground point extraction a difficult task. This paper presents a new point-based semantic segmentation network, Multi-KPConv, to overcome these challenges. Unlike methods that rely on color point clouds, Multi-KPConv uses shallow features based on domain knowledge as input to the network. The network employs a multi-dimensional kernel point convolutional architecture to extract high-level semantic features, allowing for better data interpretation. Additionally, the SimAM 3D attention mechanism is integrated to adaptively refine feature contributions and highlight important point features. We evaluate Multi-KPConv on datasets from the Dafosi mining area in China's Loess Plateau and the STPLS3D dataset. Experimental results show that Multi-KPConv outperforms current state-of-the-art models in terms of generalization and robustness, effectively extracting ground points in complex terrain, such as the gully areas of the Loess Plateau.