// the find
sstary/SSRS
Semantic Segmentation for Remote Sensing
A grab-bag of PyTorch implementations for remote sensing semantic segmentation papers out of one research group — single-modal (SAM_RS, DeepKANSeg, EVSR), unsupervised domain adaptation (GLGAN, MBATA_GAN), and multimodal fusion (FTransUNet, CMFNet, ASMFNet, MSFNet), plus a Mamba-based model (RS3Mamba). It's for someone who wants a working reference implementation of one of these specific TGRS/GRSL/JSTARS papers, not a general-purpose segmentation library.
Every method ties back to a peer-reviewed paper with a citation, several flagged ESI Highly Cited, so you're getting actual published results rather than an unvetted experiment. It covers a real spread of problem settings (unimodal, UDA, multimodal fusion, state-space models) in one place, useful for baseline comparisons. The code leans on solid upstream work (Swin-Transformer, VMamba, SAM, TransUNet) and references those repos directly instead of silently reimplementing them.
There's no root-level setup: no top-level requirements.txt or install instructions, and each method folder has its own ad hoc train.py/utils.py with no shared interface, so switching between methods means reading a different paper and a different folder each time. DeepKANSeg/model/MedSAM vendors an entire copy of the MedSAM project — including its own .github workflows and LICENSE — rather than pinning it as a dependency, which bloats the repo and muddies licensing. Multiple folders have committed __pycache__ directories with .pyc files for Python 3.7/3.8/3.10/3.11, which shouldn't be tracked at all and signals this was never meant to be a maintained package. No tests, no CI, and no indication of which checkpoints/configs actually reproduce the paper numbers.