LK-UNet: Large Kernel convolution-driven U-shaped network for semantic segmentation of high-resolution Earth surface images

Abstract

Semantic segmentation of remote sensing (RS) images plays an important role in urban planning, environmental monitoring, and agriculture. However, the receptive field of traditional convolutional neural networks (CNNs) is limited, and the model cannot capture the wider context information in the image, resulting in poor segmentation results. Therefore, this paper re-examines the role of large convolution kernels and proposes a new network LK-UNet. First, a U-shaped network driven by a large convolution kernel as the encoder is proposed to increase the receptive field and greatly improve the network’s ability to extract global information. Secondly, the enhanced atrous spatial pyramid pooling (EASPP) module is introduced in the last two stages of the encoder module to aggregate broader contextual information. Finally, in the skip connection part, the feature enhancement module (FEM) is incorporated to augment the network’s ability to capture details and further improve the target segmentation performance. Ablation experiments were performed on the ISPRS Vaihingen to validate the efficacy of each module. At the same time, the proposed method has superior performance compared with the state-of-the-art methods.