sstary/SSRS

What it solves

This project provides PyTorch implementations of various deep learning models for semantic segmentation of remote sensing imagery. It addresses the challenge of accurately identifying and classifying pixels in satellite or aerial images, supporting tasks like land-cover mapping and environmental monitoring.

How it works

The repository organizes its methods into three primary technical approaches:

• Single-modal Segmentation: Uses specialized architectures like DeepKANSeg and RS^3Mamba (a visual state space model) to process single-source imagery.
• Unsupervised Domain Adaptation (UDA): Implements GAN-based models (GLGAN, MBATA_GAN) to help models generalize across different geographic regions or sensors without needing labeled data for every new area.
• Multimodal Fusion: Combines data from different sensors or modalities (e.g., MFNet, FTransUNet) to improve segmentation accuracy through cross-modal integration.

Who it’s for

This is primarily for researchers and engineers working in remote sensing, geospatial AI, and computer vision, specifically those focusing on semantic segmentation and domain adaptation.

Highlights

• Diverse Model Zoo: Includes a wide range of recent research papers (IEEE TGRS, JSTARS, GRSL) implemented in PyTorch.
• SOTA Architectures: Implements cutting-edge techniques including Mamba-based visual state space models and Kolmogorov-Arnold Networks (KAN).
• Domain Adaptation: Provides tools for unsupervised domain adaptation to handle data shift between different remote sensing datasets.
• Utility Scripts: Includes helper scripts for image splitting, merging, and loss curve visualization.