Swin Transformer-Based UperNet for Glacier Semantic Segmentation

This repository provides the official PyTorch implementation of the paper: Wang Lecheng, Wang Maozhi, Xia Xueqin. Integrating Spectral, Topographic, and NDSI with Swin-UperNet for Glacier Extraction Under Complex Terrain Conditions, 2025,×(×):×-×.
Schematic diagram of the Swin-UperNet architecture

Introduction

Swin-UperNet is a state-of-the-art semantic segmentation framework that combines the powerful Swin Transformer backbone with the UperNet head architecture. This repository provides an implementation of the Swin-UperNet model for high-performance semantic segmentation tasks.

Key features:

• 🚀 Swin Transformer backbone for hierarchical feature extraction
• 🔄 UperNet decoder for multi-scale feature fusion
• ⚡️ High efficiency with linear computational complexity
• 🏆 State-of-the-art performance on segmentation benchmarks
• ❄️ Specialized support for glacier segmentation tasks

Installation

Prerequisites

• Python 3.12.6
• PyTorch 2.4.1
• TorchSummary 1.5.1
• TorchInfo 1.8.0
• Thop 0.1.1
• CUDA 11.8
• Timm 1.0.15
• numpy 1.26.4
• GDAL 3.8.4
• Linux environment recommended

Training Command

To train a model on the glacier segmentation task, use the following command structure:

python train.py \
    --MODEL_TYPE upernet \
    --BACKBONE_TYPE swin_t \
    --BANDS 10 \
    --NUM_CLASS 3 \
    --DATASET_PATH ./datasets/glacier \
    --BATCH_SIZE 16 \
    --EPOCHS 100 \
    --OPTIMIZER_TYPE sgd \
    --LOSS_TYPE ce \
    --LR_SCHEDULER poly \
    --INIT_LR 0.0005 \
    --GPU_ID 0

Key Parameters:

• MODEL_TYPE: Model architecture (e.g., upernet, deeplab, segnext, etc.)
• BACKBONE_TYPE: Backbone network (for models that support backbones, e.g., swin_t, resnet50, etc.)
• BANDS: Number of input channels (10 for glacier data)
• NUM_CLASS: Number of classes (including background)
• DATASET_PATH: Path to the dataset directory
• BATCH_SIZE: Batch size (adjust based on GPU memory)
• EPOCHS: Total training epochs
• OPTIMIZER_TYPE: Optimizer (sgd or adam)
• LOSS_TYPE: Loss function (ce for cross-entropy or focal for focal loss)
• LR_SCHEDULER: Learning rate scheduler (poly, step, cos, or exp)
• INIT_LR: Initial learning rate
• GPU_ID: ID of the GPU to use Note: The dataset should be organized in the following structure:

DATASET_PATH/
├── annotations/
│   ├── train.txt
│   └── val.txt
├── images/
│   ├── 1.tif
│   ├── 2.tif
│   └── ...
└── labels/
    ├── 1.tif
    ├── 2.tif
    └── ...

Comparative Models links

All models were trained from scratch under identical conditions for fair comparison:

• DeepLabv3+
  Encoder-Decoder with Atrous Separable Convolution (Liang-Chieh Chen et al., 2018, ECCV).
  Paper: arXiv:1802.02611 (PDF).
• U-Net
  U-Net: Convolutional Networks for Biomedical Image Segmentation (Olaf Ronneberger et al., 2015, MICCAI).
  Paper: arXiv:1505.04597 (PDF).
• SegFormer
  SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers (Enze Xie et al., 2021, NIPS).
  Paper: arXiv:2105.15203 (PDF).
• SETR
  Rethinking Semantic Segmentation from a Sequence-to-Sequence Perspective with Transformers (Sixiao Zheng et al., 2021, CVPR).
  Paper: arXiv:2012.15840 (PDF).
• Swin-UperNet (ours)
  Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (Ze Liu et al., 2021, ICCV).
  Paper: arXiv:2103.14030 (PDF).
  Unified Perceptual Parsing for Scene Understanding (UPerNet) (Tete Xiao et al., 2018, ECCV).
  Paper: arXiv:1807.10221 (PDF).