ModelHub

Overview¶

The ModelHub is a comprehensive collection of machine learning models and algorithms designed for hazard monitoring, prediction, and analysis. We aim to provide pre-trained models, training pipelines, and tools for developing custom hazard assessment models.

Models¶

Hazard Prediction Models¶

• Weather Forecasting, coupling ACE2 with Amtmospheric River Index (Richard Zhuang and Brandon Kerns), and Clima-X (Aditya Grover). Examples of AR tracking and forecast here

• Heatwave Forecasting (Greg Hakim)

• Storm Detector, detecting thunderstorm from seismic and weather data using deep learning (Akash Kharita, Alexandra Anderson-Frey, Marine Denolle)

• Earthquake Wavefields: deep learning models for data-driven full wavefield and peak ground motions reconstruction (nowcasting) and forecasting (lead Yiyu Ni)

• Ground Failure Modeling: Surrogate model to predict liquefaction and ground failure potential index (lead Morgan Sanger), and its coupling to wavefields (Yiyu Ni & Morgan Sanger)

• Large Earthquake Detection: deep learning model to detect earthquakes are strong ground motions in seisbench ecosystem.

Flood Hazard Models¶

• Flood Forecasting: Flood forecasting using stream gauge data, SWE, and meteorological products.

• Flash Flood Detection

Landslide Susceptibility Models (TBD)¶

• Data-Driven Reduced Order Modeling of Landlab modeling for landscape evolution/erosion and debris flow modeling.

• Susceptibility Mapping: Random forests and gradient boosting for landslide susceptibility

• Triggering Models: Models linking precipitation and seismic triggers to landslide occurrence

• Run-out Prediction: Neural networks for estimating landslide travel distance

• Landslides Detection using high temporal resolution of multi-sensor geophysical networks (seismic, infrasound, tilt, DAS) and high spatial resolution remote sensing imagery (sentinel+SAR) with Akash Kharita, Scott Henderson, and collab with AI2.

• Hydromechanical Model data-driven approach to turn multi-sensor raw data (non co-located seismic, geodetic, meteorological sensors) into real time hydromechanical properties (Manuela Köpfli)

Model Architecture¶

This is the draft of a common framework.

Deep Learning Models¶

# Example: Using a pre-trained flood detection model
from gaia_hazlab.models import FloodDetectionModel

# Load pre-trained model
model = FloodDetectionModel.from_pretrained('flood-v1.0')

# Make predictions
predictions = model.predict(satellite_image)

Traditional ML Models¶

# Example: Landslide susceptibility mapping
from gaia_hazlab.models import LandslideSusceptibility

# Initialize and train model
model = LandslideSusceptibility(algorithm='random_forest')
model.fit(X_train, y_train)

# Generate susceptibility map
susceptibility_map = model.predict_map(terrain_features)

Model Development¶

Training Pipeline¶

Our standardized training pipeline includes:

1. Data Preparation: Automated data loading and preprocessing

2. Feature Engineering: Domain-specific feature extraction

3. Model Training: Optimized training with hyperparameter tuning

4. Validation: Cross-validation and performance evaluation

5. Deployment: Model packaging and versioning

Custom Model Development¶

Create custom models using our framework:

from gaia_hazlab.models import BaseHazardModel

class CustomHazardModel(BaseHazardModel):
    def __init__(self, config):
        super().__init__(config)
        # Custom initialization

    def train(self, data):
        # Custom training logic
        pass

    def predict(self, inputs):
        # Custom prediction logic
        pass

Model Performance¶

Benchmarking¶

All models in the ModelHub are benchmarked against:

• Baseline models (traditional statistical approaches)

• State-of-the-art published methods

• Domain-specific metrics and standards

Performance Metrics¶

We evaluate models using:

• Classification: Accuracy, F1-score, ROC-AUC, Precision-Recall

• Regression: RMSE, MAE, R², Bias

• Spatial: Intersection over Union (IoU), Dice coefficient

• Probabilistic: Calibration, Brier score, Log-likelihood

Model Registry¶

Model Versions¶

• All models are versioned using semantic versioning

• Each version includes model weights, configuration, and metadata

• Reproducible training environments documented

Model Cards¶

Each model includes a comprehensive model card with:

• Model description and intended use

• Training data and preprocessing steps

• Performance metrics and limitations

• Ethical considerations and biases

• Citation and attribution

Deployment¶

Model Serving¶

Deploy models for inference:

# Deploy model as API endpoint
from gaia_hazlab.serving import ModelServer

server = ModelServer(model='flood-detection-v1.0')
server.start(port=8000)

Batch Processing¶

Process large datasets efficiently:

# Batch prediction
from gaia_hazlab.batch import BatchPredictor

predictor = BatchPredictor(model='landslide-susceptibility-v2.1')
results = predictor.process_dataset(input_data, batch_size=32)

Contributing Models¶

To contribute a model to the ModelHub:

1. Follow our [model contribution guidelines]({{ github_org_url }}/{{ book_repo }}/blob/main/CONTRIBUTING.md)

2. Ensure model meets quality standards

3. Provide complete documentation and model card

4. Submit a pull request for review

Model Citation¶

When using models from the ModelHub, please cite:

GAIA HazLab Team. (2024). [Model Name]. GAIA HazLab ModelHub.
https://gaia-hazlab.github.io

Future Developments¶

Upcoming additions to ModelHub:

• Foundation models for multi-hazard assessment

• Transfer learning capabilities across hazard types

• Federated learning for privacy-preserving model training

• Automated machine learning (AutoML) for rapid model development