MFGS2Net: A Morphological Features Guided Scale Separation Network for Remote Sensing Panoptic Segmentation

Remote sensing panoptic segmentation (RSPS) plays a crucial role in the intelligent interpretation of remote sensing data. Specifically, panoptic segmentation of airport scenes presents significant challenges due to the diverse shapes of terminal buildings, along with variations in their appearance, size, scale, materials, and color, all of which make the segmentation task highly complex. The intricate geometric shapes and varying scale parameters lead to the entangled object scales in the feature space, posing challenges for accurate segmentation. These factors collectively contribute to the complexity of panoptic segmentation for irregularly shaped buildings. Current mainstream mask-based panoptic segmentation methods do not fully consider the importance of enhancing and integrating low-level and high-level semantic features. They lack the ability to decouple features, leading to misclassification and false negatives. To solve the above-mentioned problem of panoptic segmentation of irregularly shaped buildings, this study designs a morphology-guided feature aggregation (MGFA) module to establish the correlation between low and high level semantic features at different stages. At the same time, a multiorder semantic decoupling (MOSD) module is designed to capture structural information at different scales and improve the network’s understanding of complex structural information. In addition, due to the lack of datasets for transportation hub scenarios in the RSPS task, we have collected the airport remote sensing panoptic segmentation (AP-RSPS) dataset to promote the practical application of RSPS tasks. Extensive experiments on the AP-RSPS dataset and BSB aerial dataset show that the proposed morphological features guided scale separation network (MFGS2Net) outperforms several SOTA methods.