WangLibo1995/GeoSeg

# Technical Summary Implements a unified PyTorch Lightning and timm-based framework supporting hybrid architectures including state-space models (PyramidMamba), vision transformers (UNetFormer, DC-Swin), and CNNs for remote sensing segmentation across ISPRS, UAVid, LoveDA, and OpenEarthMap datasets. Provides multi-scale training/testing pipelines and inference optimization for large-scale geospatial imagery through patch-based processing with configurable stride and tile sizes. Architecture choices prioritize efficiency through encoder-decoder designs with dense connection modules (MANet, ABCNet) and feature pyramid mechanisms (A2FPN) to balance computational cost against segmentation accuracy on high-resolution aerial and satellite data.

1,046 stars. No commits in the last 6 months.