IceSegNet: A stage-aware dynamic kernel network for river ice segmentation in remote sensing imagery

Kaijun Wu, Dingju Zhou*, Juanjuan Du, Yuelian Wu, Lidong Zhang

Applied Soft Computing, Vol. 186, 2026 | 📄 Paper

• 🎯 Stage-Aware Kernel Update Module — Three structurally distinct stages (detail preservation → transition stabilization → semantic purification) with progressively reduced feedforward widths (2048 → 1024 → 512), cutting FFN parameters by 41.7% while improving mIoU by 0.96%.
• 🧠 UPerSCA-MTL Decode Head — Unified Perceptual Parsing enhanced with Spatial Cross-Attention (SCA) and Multi-Task Learning (MTL) for joint semantic segmentation and edge detection.
• 🏆 State-of-the-Art Results — 93.81% mIoU on NWPU_YRCC2 (+1.19% over K-Net, +1.31% over Mask2Former); 93.56% mIoU on NWPU_YRCC_EX (+0.43% over K-Net).

Accurate segmentation of river ice in remote sensing imagery is critical for quantifying ice coverage—a key variable in early warning and risk assessment of ice-jam disasters. IceSegNet addresses the challenges of large-scale variation, spectral similarity between ice and water, and ambiguous boundaries through two core innovations:

1. A stage-aware kernel update module that refines features through three structurally distinct stages with progressively reduced hidden widths.
2. UPerSCA-MTL, a multi-task decoding head that fuses spatial cross-attention and edge detection to enhance boundary accuracy.

Evaluated on two Yellow River ice datasets — NWPU_YRCC2 (1525 images, 4 classes) and NWPU_YRCC_EX (887 images, 3 classes) — IceSegNet achieves state-of-the-art performance among 18 competing segmentation models.

• PPM: Multi-scale context aggregation via pyramid pooling
• SCA (Spatial Cross-Attention): Horizontal + vertical global pooling to capture axis-specific directional dependencies, enhancing ambiguous ice–water boundary delineation
• Depthwise Separable Convolutions: Reduce computational complexity without sacrificing feature quality
• Multi-Task Head: Parallel segmentation + edge detection branches; edge supervision derived automatically from GT mask gradients (no extra annotation needed)

• Python ≥ 3.8
• PyTorch ≥ 1.12 with CUDA
• MMEngine, MMCV ≥ 2.0, MMSegmentation

1. Create and activate conda environment

conda create -n icesegnet python=3.8 -y
conda activate icesegnet

2. Install PyTorch (example: CUDA 11.6)

pip install torch==1.12.1+cu116 torchvision==0.13.1+cu116 \
    --extra-index-url https://download.pytorch.org/whl/cu116

3. Install MMEngine and MMCV

pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.0"

4. Install MMSegmentation

git clone https://github.com/open-mmlab/mmsegmentation.git
cd mmsegmentation
pip install -v -e .
cd ..

5. Clone this repository

git clone https://github.com/fox-4869/IceSegNet.git
cd IceSegNet
pip install -r requirements.txt

Download the datasets:

• NWPU_YRCC2: https://github.com/nwpulab113/NWPUYRCC2
• NWPU_YRCC_EX: https://github.com/nwpulab113/NWPUYRCCEX

Organize the directory as follows:

data/
├── NWPU_YRCC2_JPG1/
│   ├── train/             # Training images (.jpg)
│   ├── train_labels/      # Training annotations (.png)
│   ├── val/               # Validation images
│   └── val_labels/        # Validation annotations
└── NWPU_YRCC_EX/
    ├── train/
    ├── train_labels/
    ├── val/
    └── val_labels/

1. Copy dataset definitions

cp datasets/NWPU_YRCC2_JPG1.py mmsegmentation/mmseg/datasets/
cp datasets/NWPU_YRCC.py       mmsegmentation/mmseg/datasets/

Add to mmseg/datasets/__init__.py:

from .NWPU_YRCC2_JPG1 import NWPU_YRCC2_JPG1
from .NWPU_YRCC import NWPU_YRCC

2. Copy model components

cp models/sefpn.py             mmsegmentation/mmseg/models/necks/
cp models/uper_att_plus_head.py mmsegmentation/mmseg/models/decode_heads/

Register in the corresponding __init__.py files under mmseg/models/necks/ and mmseg/models/decode_heads/.

Single GPU

python tools/train.py configs/icesegnet-config.py

Multi-GPU (recommended — paper uses 2× RTX 3090)

bash tools/dist_train.sh configs/icesegnet-config.py 2

Key training settings:

Standard evaluation

python tools/test.py configs/icesegnet-config.py /path/to/checkpoint.pth

With Test-Time Augmentation (multi-scale + flip)

python tools/test.py configs/icesegnet-config.py /path/to/checkpoint.pth --tta

Reported metrics: mIoU, mDice, mFscore, PA, BFscore

IceSegNet/
├── configs/
│   └── icesegnet-config.py        # Full training & evaluation config
├── datasets/
│   ├── NWPU_YRCC2_JPG1.py         # 4-class dataset (Land/Water/Shore Ice/Drift Ice)
│   └── NWPU_YRCC.py               # 3-class dataset (Others/Water/Shore Ice)
├── models/
│   ├── sefpn.py                   # SEFPN neck with BN+ReLU normalization
│   └── uper_att_plus_head.py      # UPerSCA-MTL decode head
├── tools/                         # Training & testing scripts (MMSeg)
├── README.md                      # English README
└── README_CN.md                   # Chinese README

If IceSegNet is helpful for your research, please cite:

@article{wu2026icesegnet,
  title     = {IceSegNet: A stage-aware dynamic kernel network for river ice
               segmentation in remote sensing imagery},
  author    = {Wu, Kaijun and Zhou, Dingju and Du, Juanjuan and
               Wu, Yuelian and Zhang, Lidong},
  journal   = {Applied Soft Computing},
  volume    = {186},
  pages     = {114120},
  year      = {2026},
  publisher = {Elsevier},
  doi       = {10.1016/j.asoc.2025.114120}
}

This work was supported by the Natural Science Foundation Key Project of Gansu Province (23JRRA860), the Inner Mongolia Key R&D and Achievement Transformation Project (2023YFSH0043, 2023YFDZ0043, 2023YFDZ0054), the Key Research and Development Project of Lanzhou Jiaotong University (ZDYF2304), and the Excellent Graduate Student "Innovation Star" Project of Gansu Province (2025CXZX-682).

This codebase is built on MMSegmentation. We also thank the authors of K-Net for the foundational dynamic kernel framework.

Corresponding Author: Dingju Zhou — dingjuzhou@163.com

Lanzhou Jiaotong University, Lanzhou 730070, China