To address the challenge of accurately estimating the 6D pose of satellites in space environments characterized by significant lighting variations and complex background changes, as well as limited satellite texture features, a novel approach is proposed. This method combines partial convolution (PConv) and large kernel attention (LKA) within the framework of the DenseFusion network. Firstly, improvements are made to the generation of Blender-based rendering datasets and a dedicated simulation dataset for satellite pose estimation is created. Subsequently, the integration of a partial convolution module into the feature extraction network’s encoding section reduces sensitivity to lighting changes and background noise. Finally, to capture weak texture features at different scales on satellite images, a pyramid scene parsing network LKA-PSPNet (Large Kernel Attention Pyramid Scene Parsing Network) is designed. Experimental results demonstrate that this algorithm achieves an ADD-(S) index of 97.6% and 89.2% on both LineMod public dataset and self-built satellite simulation dataset respectively- marking an improvement by 3.3 percentage points and 2.9 percentage points over previous methods- thus validating its effectiveness.