GlacierCastAI

GlacierCastAI predicts glacier retreat acceleration before it becomes visible in satellite imagery. By fusing Landsat time series, ERA5 climate signals, and Copernicus DEM terrain features in a multi-modal deep learning model, it forecasts glacier boundary changes at 1, 3, and 5-year horizons - and uses SHAP attribution to identify which climate drivers are responsible, making predictions interpretable and trustworthy.

Research Question

Can climate signals predict glacier retreat acceleration before it becomes detectable in satellite imagery?

Study Glaciers

Glacier	Region	Climate Regime
Aletsch	Swiss Alps	Alpine
Gangotri	Himalayas	Monsoon
Grey	Patagonia	Maritime
Columbia	Alaska	Maritime/Subarctic
Athabasca	Canadian Rockies	Continental

Model Architecture

• Backbone: ResNet50 (pretrained, ImageNet)
• Temporal: ConvLSTM (3 layers, hidden_dim=256, T=4 timesteps)
• Climate encoder: Cross-attention transformer (16-dim ERA5 features × 4 seasons)
• Decoder: UNet-style with skip connections
• Heads: Boundary mask, retreat rate, risk score
• Parameters: 56.1M total

Data

Source	Variables	Resolution
Landsat 5/7/8/9 Collection 2	Green, SWIR1, NDSI	30m
Copernicus DEM GLO-30	Elevation, slope, aspect	30m
ERA5 Monthly Means	T2m, precipitation, snowfall, solar radiation	~31km

• 64 Landsat scenes across 5 glaciers (2000–2023)
• 29,810 patches (256×256px, 64px overlap)
• 40,476 sequences (T=4 input timesteps, horizons 1/2/3yr)
• Train/Val/Test split: 27,725 / 5,861 / 6,890

Experiment Results

Model	Params	test/IoU	test/BF1	Notes
B1: Persistence	—	0.160	0.128	Traditional baseline
B2: Linear trend	—	0.169	0.147	Traditional baseline
exp001: Image only	56.1M	0.326	0.158	Deep learning baseline
exp002: Image + Climate	56.1M	0.337	0.145	Best overall
exp003: Image + Climate + DEM	56.1M	0.331	0.123	Full multimodal
exp005: Climate only (MLP)	0.66M	0.320	0.135	Lightweight baseline

Key findings:

• All deep learning models outperform traditional baselines by ~89–99% relative IoU improvement
• ERA5 climate signals alone (exp005, 661K params) achieve 98% of image-only performance
• Adding climate to imagery (exp002) yields the best overall result

Per-glacier IoU (exp002):

Glacier	IoU	n patches
Columbia	0.500 ± 0.372	200
Grey	0.250 ± 0.201	200
Aletsch	0.122 ± 0.187	200
Athabasca	0.045 ± 0.110	200
Gangotri	—	0 (no test sequences)

Full results: results/baselines.json, results/per_glacier_results.json

Setup

conda create -n glaciercastai python=3.12
conda activate glaciercastai
pip install poetry
poetry install

Required API keys (copy .env.example to .env):

• WANDB_API_KEY - Weights & Biases
• CDS API key in ~/.cdsapirc - ERA5 download
• NASA Earthdata credentials - Landsat download

Training

# Train multimodal model (exp001-003)
python scripts/train.py --config configs/model.yaml

# Train climate-only MLP baseline (exp005)
python scripts/train.py --config configs/exp005.yaml

# Resume from checkpoint
python scripts/train.py --config configs/model.yaml --resume experiments/checkpoints/exp002/last.ckpt

# Debug run (2 batches, no W&B logging)
python scripts/train.py --config configs/model.yaml --debug

Analysis Scripts

# Generate figures (ablation bar chart, study area map, qualitative predictions)
python scripts/generate_figures.py --checkpoint experiments/checkpoints/exp002/10-valiou0.2475.ckpt

# Compute persistence and linear trend baselines + per-glacier breakdown
python scripts/compute_baselines.py --checkpoint experiments/checkpoints/exp002/10-valiou0.2475.ckpt

# SHAP climate attribution analysis
python scripts/shap_analysis.py --checkpoint experiments/checkpoints/exp002/10-valiou0.2475.ckpt

Repository Structure

GlacierCastAI/
├── configs/
│   ├── model.yaml        # Multimodal model config (exp001-003)
│   └── exp005.yaml       # Climate-only MLP config (exp005)
├── scripts/              # Training, preprocessing, analysis scripts
├── src/
│   ├── models/
│   │   ├── glaciercastai.py   # Full multimodal model
│   │   ├── climate_mlp.py     # Climate-only MLP baseline
│   │   ├── backbones/         # ResNet50 encoder
│   │   ├── temporal/          # ConvLSTM, Transformer
│   │   ├── heads/             # Regression, segmentation heads
│   │   └── losses/            # Combined loss function
│   └── training/
│       ├── datamodule.py      # Lightning DataModule
│       └── trainer.py         # Lightning training module
├── experiments/
│   └── checkpoints/           # Per-experiment results (gitignored except JSON)
│       ├── exp001/exp001_test_results.json
│       ├── exp002/exp002_test_results.json
│       ├── exp003/exp003_test_results.json
│       └── exp005/exp005_test_results.json
├── results/                   # Baseline and analysis results
│   ├── baselines.json         # Persistence and linear trend IoU/BF1
│   ├── per_glacier_results.json
│   └── bootstrap_ci.json
└── data/                      # Raw and processed data (gitignored)