Glacier Movement Prediction System

Bharatiya Antariksh Hackathon 2025

A deep learning system for predicting glacier lake formation and GLOF (Glacial Lake Outburst Flood) risks using multi-modal satellite data and velocity measurements.

Features

• Multi-Region Support: 13 RGI regions including Alaska, Karakoram, Iceland, and more
• Multi-Modal Data: ITS_LIVE velocity, DEM (SRTM/ASTER/CartoDEM), Sentinel-1/2 imagery
• Two Model Architectures:
  • Enhanced TimeSformer (video transformer with divided space-time attention)
  • Simple 3D CNN (better for small datasets)
• Memory Optimized: Runs on 4GB GPU with mixed precision training
• Temporal Analysis: GLOF risk detection based on area changes
• Satellite Evaluation: Direct evaluation on Sentinel-2 imagery

Dataset Structure

GlacierMovementPrediction/ ├── data/ │ ├── velocity/ITS_LIVE/ # Velocity data (13 regions) │ ├── dem/ # SRTM, ASTER_GDEM, CartoDEM │ ├── outlines/ # RGI glacier outlines │ ├── centerlines/ # Glacier centerlines │ ├── satellite/ │ │ ├── Sentinel1/ # C-band SAR │ │ ├── Sentinel2/ # Optical (2020-2025, R1-R5) │ │ └── Landsat/ # For future expansion │ └── mass_balance/ # WGMS mass balance data ├── features/ # Extracted features ├── models/checkpoints/ # Trained models └── evaluation/ # Results and visualizations

text

Installation

Clone repository git clone Anaconda Env Create

conda create -n glacier_pred python=3.10 -y && conda activate glacier_pred

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia -y && conda install -c conda-forge gdal=3.4 rasterio geopandas shapely fiona pyproj xarray netcdf4 h5py zarr -y && conda install numpy scipy pandas scikit-learn scikit-image opencv matplotlib seaborn plotly -y pip install albumentations imgaug pillow fastapi "uvicorn[standard]" streamlit pydantic python-multipart aiofiles tqdm pyyaml wandb tensorboard pytest pytest-cov black flake8 mypy accelerate onnx onnxruntime requests boto3 google-cloud-storage

text

Usage

1. Feature Extraction

Extract features from raw data:

python main.py --mode extract

text

Extract specific regions:

python main.py --mode extract --regions RGI01_Alaska RGI14_Karakoram

text

2. Training

Train with TimeSformer:

python main.py --mode train --model_type timesformer --epochs 80

text

Train with Simple 3D CNN (recommended for small datasets):

python main.py --mode train --model_type simple_cnn --epochs 50

text

Custom hyperparameters:

python main.py --mode train --model_type timesformer --epochs 100 --batch_size 2 --learning_rate 1e-4

text

3. Evaluation

Evaluate on test regions:

python main.py --mode evaluate

text

4. Satellite Evaluation

Evaluate on Sentinel-2 imagery:

python main.py --mode satellite_eval

text

5. Full Pipeline

Run complete pipeline:

python main.py --mode all

text

Model Architecture

Enhanced TimeSformer

• Parameters: 3.7M (optimized for 4GB GPU)
• Input: (B, T=6, C=8, H=128, W=128)
• Features: Divided space-time attention, DEM fusion, change detection
• Best for: Larger datasets (>100 samples)

Simple 3D CNN

• Parameters: 2.1M
• Architecture: 3D encoder-decoder with residual blocks
• Best for: Small datasets (<50 samples)

Configuration

Edit scripts/utils/config.py to customize:

Model settings IMAGE_SIZE = 128 # Image size NUM_FRAMES = 6 # Temporal frames BATCH_SIZE = 1 # Batch size MODEL_TYPE = 'timesformer' # or 'simple_cnn'

Training settings LEARNING_RATE = 5e-5 NUM_EPOCHS = 80 USE_AMP = True # Mixed precision

Loss weights DICE_WEIGHT = 0.5 BCE_WEIGHT = 0.3 FOCAL_WEIGHT = 0.2

GLOF detection GLOF_AREA_THRESHOLD = 0.20 # 20% area change

text

Results

Expected performance metrics:

• IoU: 0.30 - 0.50
• Dice: 0.40 - 0.65
• F1 Score: 0.40 - 0.65
• Precision: 0.50 - 0.75

Satellite Regions

Currently supports:

• R1, R2: Sentinel-2 data (2020-2025)
• Expandable to R3, R4, R5: Add data to respective folders

GLOF Risk Detection

Detects GLOF risk based on:

• Rapid area increase (>20% threshold)
• Temporal velocity trends
• Multi-year lake evolution

Troubleshooting

Out of Memory

• Reduce IMAGE_SIZE (128 → 96)
• Increase GRADIENT_ACCUMULATION_STEPS
• Use simple_cnn model

Low Metrics

• Check target generation mode (balanced recommended)
• Ensure proper checkpoint loading
• Verify data quality and alignment

Missing Data

• Run feature extraction first
• Check data folder structure
• Verify file formats (.tif for DEM, .nc/.tif for velocity)

Citation

If you use this code, please cite:

@software{glacier_movement_prediction_2025, title={Glacier Movement Prediction System}, author={[Subhash R]}, year={2025}, }

text

Contact

For questions or issues, please contact [subhashravichandran7432@gmail.com]