Performance and Analysis of FCN, U-Net, and SegNet in Remote Sensing Image Segmentation Based on the LoveDA Dataset

Remote sensing image segmentation is a vital method in image analysis that significantly contributes to the extraction of surface information and aids in land use planning. This study utilizes the LoveDA dataset to investigate the segmentation performance of three classic deep learning models: Fully Convolutional Networks(FCN), U-Net, and SegNet, in both urban and rural scenarios. By partitioning the Urban-Rural dataset of LoveDA for training and testing, it was determined that SegNet excels in detail restoration and boundary handling, while U-Net demonstrates robust adaptability across various scenarios. In contrast, FCN, with its simpler architecture, shows lower segmentation accuracy in certain contexts. This paper offers a comprehensive comparison of the strengths and weaknesses of different models in remote sensing image segmentation and proposes enhancements in model structure and data preprocessing optimization. The findings provide valuable insights for improving the performance of semantic segmentation models and are of significant importance for the precise analysis and practical applications of remote sensing images.