Answers are grounded in this paper and cite their sources.
Irregular mare patches (IMPs) are important lunar features for investigating late-stage volcanism on the Moon. IMPs exhibit considerable variation in size, with many occurring at relatively small scales, while their indistinct boundaries and textures highly similar to the surrounding mare terrain further complicate accurate automatic segmentation. To address these challenges, we propose a dual-branch adaptive fusion network (DBAF-Net) for IMP segmentation. The network employs a full-image global branch and a fixed-center local branch to jointly model the surrounding geomorphic context and localized texture patterns. A Resolution-Matched Selective Fusion (RSF) module matches feature-map dimensions and adaptively integrates the global and local features, while a Content-Matched Global-Local Cross-Attention (CMGCA) module further promotes content-aware interaction between the two branches. The decoder further incorporates spatial and channel squeeze-and-excitation (scSE) and residual hierarchical decoder aggregation (RHDA) to enhance foreground responses and exploit multilevel decoder features. Experiments on an IMP dataset containing 11,136 high-resolution samples show that DBAF-Net achieves a Precision of 85.06%, Recall of 86.74%, F1-score of 85.89%, mIoU of 87.32%, and IoU of 75.27%, outperforming representative segmentation models including ConvNeXt, DeepLabv3, MobileNet, Swinformer, and SegFormer. Quantitative and visual comparisons further show that DBAF-Net reduces missed and false detections in complex mare backgrounds while producing more complete IMP delineation and more accurate boundaries.
No takes yet. Share an insight, caveat, or question.
Lei et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: