This repository contains a state-of-the-art glacier semantic segmentation implementation based on DeepLab V3+ with Convolutional Block Attention Module (CBAM) and Test-Time Augmentation (TTA) for the GlacierHack 2025 competition.

The model performs binary semantic segmentation to distinguish glacier pixels from non-glacier pixels in multispectral satellite imagery using 5 spectral bands: B2 (Blue), B3 (Green), B4 (Red), B6 (SWIR), and B10 (TIR1).

• Backbone: ResNet-34 encoder for feature extraction
• ASPP with CBAM: Atrous Spatial Pyramid Pooling enhanced with Convolutional Block Attention Module
• Decoder: Depthwise separable convolutions for efficient upsampling
• Skip Connections: Low-level features fusion for better boundary preservation

• Channel Attention: Focuses on important feature channels
• Spatial Attention: Emphasizes relevant spatial locations
• Integration: Applied to each ASPP branch for enhanced feature representation

• Transformations: Original, horizontal flip, vertical flip, both flips
• Ensemble: Averages predictions from all augmented versions
• Robustness: Improves model consistency and performance

Google Colab notebook for model training

Features:

• Complete training pipeline from data loading to model evaluation
• Data augmentation with albumentations
• Combined BCE + Dice loss for handling class imbalance
• Comprehensive metrics: MCC, IoU, F1-score, R²
• Training visualization and model checkpointing
• Model compression to stay under 200MB limit
• TTA implementation and evaluation

Key Sections:

• Model architecture definition (Attention DeepLab V3+ with CBAM)
• Custom dataset class for .npy file handling
• Training loop with early stopping
• Validation and testing with/without TTA
• Results visualization and model saving

Competition inference script

Features:

• Follows exact competition requirements
• Loads pre-trained model weights (model.pth)
• Processes multi-band TIFF images
• Implements TTA for robust predictions
• Saves binary masks in required format

Key Functions:

• AttentionDeepLabV3Plus: Complete model architecture
• TTAWrapper: Test-time augmentation implementation
• maskgeration: Main inference function
• load_and_preprocess_image: Image preprocessing pipeline

• Input: slice_x_image_y.npy (512×512×15) → Extract bands [1,2,3,5,9] → (512×512×5)
• Target: slice_x_mask_y.npy (512×512×3) → Compress to (512×512×1)
  • Channel 0: Clean-ice glacier
  • Channel 1: Debris-covered glacier
  • Channel 2: HKH region mask
  • Final target: (clean-ice OR debris-covered) AND hkh_region

• Structure:

  dataset/
    Band1/
      img001.tif
      img002.tif
      ...
    Band2/
      ...
    Band5/
      ...
  

1. Setup Environment:

   !pip install torch torchvision tqdm opencv-python pillow scikit-learn numpy tifffile segmentation-models-pytorch
   !pip install albumentations matplotlib seaborn

2. Mount Google Drive:

   from google.colab import drive
   drive.mount('/content/drive')

3. Update Data Path:

   DATA_DIR = '/content/drive/MyDrive/glacier_data'  # Update this path

4. Run Notebook: Execute all cells in glacier_segmentation_training.ipynb

1. Install Dependencies:

   pip install torch torchvision tqdm opencv-python pillow scikit-learn numpy tifffile albumentations

2. Run Inference:

   python solution.py --data /path/to/test/dataset --masks /path/to/masks --out /path/to/output

• Primary: Matthews Correlation Coefficient (MCC) - optimized for balanced performance
• Secondary: IoU, F1-score, R² for comprehensive evaluation

• MCC: >0.85 on validation set
• IoU: >0.80 for glacier segmentation
• F1-score: >0.90 for binary classification
• Model Size: <200MB (competition requirement)

Combined Loss = α × BCE Loss + (1-α) × Dice Loss

• Handles class imbalance effectively
• BCE for pixel-wise classification
• Dice for overlap optimization

• Horizontal/Vertical flips
• 90° rotations
• Brightness/Contrast adjustments
• Gaussian noise injection

Selected 5 most informative bands for glacier detection:

• Band 1 (B2): Blue - Snow/ice detection
• Band 2 (B3): Green - Vegetation contrast
• Band 3 (B4): Red - Rock/soil discrimination
• Band 4 (B6): SWIR - Ice/snow separation
• Band 5 (B10): TIR - Thermal signature

• State dict only (no optimizer)
• Post-training quantization if needed
• Weight pruning for size optimization

✅ Script Name: solution.py
✅ Function Name: maskgeration(imagepath, out_dir)
✅ Model Weights: model.pth (<200MB)
✅ Dependencies: Listed in script header
✅ Output Format: Binary TIFF masks
✅ File Naming: Matches input filenames
✅ Arguments: --data, --masks, --out

Input (5 bands) → ResNet-34 Encoder → ASPP + CBAM → Decoder → Binary Mask
                      ↓                    ↓             ↓
                 Low-level feat.    High-level feat.   Skip Connection
                      ↓                    ↓             ↓
                  Projection         Attention       Upsampling
                      ↓                    ↓             ↓
                      └─────────→ Concatenate �────────┘

1. DeepLab V3+: Chen, L. C., et al. "Encoder-decoder with atrous separable convolution for semantic image segmentation." ECCV 2018.

2. CBAM: Woo, S., et al. "CBAM: Convolutional block attention module." ECCV 2018.

3. Glacier Extraction: Chu, X., et al. "Glacier extraction based on high spatial resolution remote sensing images using a deep learning approach with attention mechanism." The Cryosphere 2022.

@article{glacier_segmentation_2025,
  title={Glacier Semantic Segmentation with Attention DeepLab V3+},
  author={Your Name},
  journal={GlacierHack 2025},
  year={2025}
}

For questions or issues, please contact: [[email protected]]

Note: This implementation is designed specifically for the GlacierHack 2025 competition requirements and follows all submission guidelines.

torch>=1.9.0 torchvision>=0.10.0

numpy>=1.21.0 pillow>=8.3.0 opencv-python>=4.5.0 scikit-learn>=1.0.0 tifffile>=2021.7.2

albumentations>=1.1.0 matplotlib>=3.4.0 seaborn>=0.11.0

tqdm>=4.62.0