Research on remote sensing image semantic segmentation method based on multiscale feature fusion

Semantic segmentation constitutes a fundamental task in the processing and analysis of remote sensing imagery. Deep learning-based semantic segmentation has been extensively utilized in applications such as national land remote sensing monitoring and remote sensing interpretation workflows. To overcome the constrained representational ability resulting from semantic gaps between the encoder and decoder in convolutional neural networks, this study introduces a semantic segmentation framework for remote sensing images that integrates multi-scale feature fusion with contour-aware learning. The designed solution augments the network's feature learning capacity by employing an adaptive channel fusion mechanism and a contour-consistent loss function. Validation experiments performed on the public Vaihingen dataset show that our approach exceeds the performance of existing methods on several evaluation metrics, reaching a mean Intersection over Union (mIoU) of 71.77%. The experimental findings confirm the practicality and robustness of the presented method.