LouayN/swiss-snow

Swiss Snow Quality Predictor

Predicts snow quality and good ski day probability for 75 Swiss alpine stations, using MeteoSwiss official open data and XGBoost models — served via an interactive map.

Demo

[image]

Interactive PyDeck map — each station bubble is colored by predicted snow quality (0–100).

Architecture

MeteoSwiss OGD ──┐
                  ├──► Feature Pipeline ──► Hopsworks Feature Store
OpenMeteo ERA5 ──┘          │
                             ▼
                      XGBoost Training ──► Hopsworks Model Registry
                                                    │
                                                    ▼
                                           Streamlit Map (app.py)

If HOPSWORKS_API_KEY is not set, everything runs locally — features write to data/processed/, models save to models/. Hopsworks is purely additive.

Data Sources

No API keys required.

Models

• —Target: next-day (t+1) to prevent label leakage
• —Train: 2009–2022 (49K rows) | Val: 2023–2025 (13K rows)
• —Snow quality score (0–100): composite of snow depth, recent snowfall delta, temperature, cold+sunny bonus, warm-rain penalty
• —Good ski day: binary label using warm-rain conditioning — rain_48h × (temp > 1°C) < 5mm → 36.4% positive rate

MLOps

Feature Store + Model Registry — Hopsworks

Set HOPSWORKS_API_KEY to activate. All pipelines degrade gracefully without it.

Models are versioned in the Hopsworks Model Registry. app.py loads from the registry on startup and falls back to local JSON files if unavailable.

Experiment Tracking — MLflow

All training runs logged: hyperparameters, metrics, and model artifacts.

mlflow ui --port 5000

Automated Retraining — Prefect 3

Weekly pipeline every Monday 06:00 Zurich time:

fetch fresh data → rebuild features → push to Hopsworks → drift check → retrain → log to MLflow

python pipelines/weekly_retrain.py           # run on schedule
python pipelines/weekly_retrain.py --run-now # trigger immediately

Data & Model Drift Monitoring — Evidently

Generates HTML + JSON drift reports using DataDriftPreset + DataSummaryPreset. Reports saved to reports/ and logged to the drift_monitoring MLflow experiment.

python src/monitoring/drift_report.py

CI — GitHub Actions

ci.yml runs on every push and PR: lint (ruff) + smoke-test imports.

Running Locally

# Install dependencies
pip install -r requirements.txt

# Fetch latest data
python src/data/fetch_meteoswiss.py recent

# Rebuild features  (also pushes to Hopsworks if HOPSWORKS_API_KEY is set)
python src/features/build_features.py

# Retrain models    (also pushes to Hopsworks Model Registry if key is set)
python src/models/train.py

# Start map frontend
streamlit run app.py

To activate Hopsworks (optional):

export HOPSWORKS_API_KEY=your_key_here   # from app.hopsworks.ai → Settings → API Keys

With Docker (local only)

docker compose build
docker compose up        # API on :8000, Streamlit on :8501

Deployment (free)

One-time setup:

1. 1.Create a Hugging Face Space (Streamlit SDK)
2. 2.Add HOPSWORKS_API_KEY as a Space Secret (Settings tab)
3. 3.git remote add space https://huggingface.co/spaces/<username>/swiss-snow
4. 4.git push space main

GitHub Actions secrets required (for weekly retrain):

• —HOPSWORKS_API_KEY — pushes retrained models to Hopsworks registry
• —HF_TOKEN — restarts the Space so it loads the new model
• —HF_SPACE — Space ID in username/space-name format (Actions variable, not secret)

Project Structure

swiss_snow/
├── src/
│   ├── data/
│   │   ├── fetch_meteoswiss.py     # MeteoSwiss OGD fetcher
│   │   └── fetch_openmeteo.py      # OpenMeteo forecast + ERA5 fetcher
│   ├── features/
│   │   └── build_features.py       # Feature pipeline + target computation
│   ├── models/
│   │   └── train.py                # XGBoost training + MLflow logging
│   ├── api/
│   │   └── app.py                  # FastAPI serving endpoint
│   └── monitoring/
│       └── drift_report.py         # Evidently drift reports
├── pipelines/
│   └── weekly_retrain.py           # Prefect 3 weekly retrain flow
├── models/
│   ├── snow_quality_regressor.json
│   └── good_ski_day_classifier.json
├── data/
│   ├── raw/                        # Parquet files from MeteoSwiss
│   └── processed/                  # Engineered feature sets
├── reports/                        # Drift report outputs (HTML + JSON)
├── app.py                          # Streamlit + PyDeck map frontend
├── hopsworks_utils.py              # Hopsworks Feature Store + Model Registry helpers
├── Dockerfile
├── docker-compose.yml
└── requirements.txt

Tech Stack

• —Data: MeteoSwiss OGD, OpenMeteo ERA5
• —Features: pandas, numpy
• —Models: XGBoost
• —Feature store / model registry: Hopsworks
• —Experiment tracking: MLflow
• —Orchestration: Prefect 3
• —Monitoring: Evidently
• —API: FastAPI
• —Frontend: Streamlit, PyDeck
• —Containerisation: Docker, Docker Compose
• —CI: GitHub Actions