Non-local Architecture Net Semantic Segmentation Network for Remote Sensing Images

Remote sensing images play a vital role in geographic information acquisition and resource management. However, challenges such as complex terrain variations, varying object scales, and diverse lighting and weather conditions often hinder the accuracy of edge segmentation and small-object detection in existing methods. Traditional approaches typically rely on stacking multiple convolutional layers to expand the receptive field, which not only increases computational costs but also results in incomplete edge segmentation and poor recognition of small objects.To address these limitations, this paper introduces a Non-Local U-net (NLUnet) model based on the U-net architecture. NLUnet builds upon U-net’s strengths in local feature extraction by integrating a non-local mechanism that captures long-range dependencies, thereby effectively expanding the receptive field. This approach enhances the capture of global contextual information and reduces computational costs, particularly excelling in complex scenes. NLUnet significantly improves the segmentation of small objects and the completeness of edge segmentation.Experimental results on the Aerial Imagery dataset demonstrate the superior performance of NLUnet, achieving an IoU of 93%, an OA of 92%, and an average F1 score of 91%. NLUnet outperforms other methods in small-object segmentation, underscoring its effectiveness in complex remote sensing scenarios.