Accurate land cover classification is crucial in opencast coal mining for environmental monitoring and industrial production. Currently, in the field of land cover mapping in opencast coal mining areas, existing methods face challenges such as extensive computational burden, inadequate accuracy caused by multiscale and boundary blurring. This paper proposes a remote sensing image semantic segmentation network, GAMF-Net, to address these issues. To address the challenges of high computational complexity and low accuracy in multiscale recognition for segmentation tasks, GAMF-Net adopts a lightweight backbone network, E-MobileNeXt, which expands the depth of the network while maintaining lightweight characteristics, enhancing the model’s learning capacity for complex land cover object features within the mining area. In addition, this paper proposes a spatial pyramid pooling structure based on cross-layer connections, DS-ASPP, which enhances the model’s receptive field by introducing a strip pooling (SP) module, thereby focusing more on spatial features in remote sensing images. To handle blurred boundaries in remote sensing data, this paper integrates a global attention mechanism module into the decoder. It improves the network’s focus on land object boundaries by capturing global information. This paper constructs a dataset of 5983 remote sensing images from high-resolution data in the opencast coal mining area. Experimental results on the open-pit mining area dataset demonstrate that the proposed GAMF-Net outperforms existing methodsin multiple performance metrics.