Skip to content

Frequently Asked Questions (FAQ)

Common questions about IBM Geospatial Studio and this workshop.

🎯 General Questions

What is IBM Geospatial Studio?

IBM Geospatial Studio is an open-source platform for training, fine-tuning, and deploying geospatial AI models. It provides:

  • Pre-trained foundation models (Prithvi, Clay, etc.) as starting points
  • Tools for dataset onboarding and management
  • Model fine-tuning capabilities
  • Inference pipeline for running models at scale
  • Web UI and Python SDK for easy interaction

Who is this workshop for?

This workshop is designed for:

  • Data scientists interested in geospatial AI
  • Remote sensing professionals wanting to apply ML
  • Developers building geospatial applications
  • Researchers working with satellite imagery
  • Anyone curious about geospatial foundation models

No prior geospatial experience required, but basic Python knowledge is helpful.

What will I learn in this workshop?

By completing this workshop, you will:

  • Deploy Geospatial Studio locally or on a cluster
  • Understand geospatial AI concepts and workflows
  • Use the Python SDK to interact with the platform
  • Onboard and manage geospatial datasets
  • Run inference with fine-tuned models
  • Fine-tune foundation models for custom tasks
  • Visualize and analyze results

How long does the workshop take?

Total time: 3-4 hours

  • Pre-work (deployment): 1-1.5 hours
  • Introduction: 15 minutes
  • Lab 1 (Getting Started with IBM Geospatial Studio): 10 minutes
  • Lab 2 (Onboarding Pre-computed Examples): 20 minutes
  • Lab 3 (Upload Model Checkpoints and Run Inference): 30 minutes
  • Lab 4 (Training a Custom Model for Wildfire Burn Scar Detection): 60-90 minutes

You can complete labs at your own pace and take breaks as needed.

💻 Technical Questions

What are the system requirements?

Minimum requirements: - CPU: 8 cores - RAM: 16GB - Storage: 100GB free space - GPU: Optional but recommended (8GB+ VRAM) - OS: Linux, macOS, or Windows with WSL2

Recommended for production: - CPU: 16+ cores - RAM: 32GB+ - Storage: 100GB+ SSD - GPU: NVIDIA GPU with 16GB+ VRAM - OS: Linux (Ubuntu 20.04+)

Do I need a GPU?

For this workshop: No, GPU is optional. You can complete all labs using CPU-only mode.

For production use: Yes, GPU is highly recommended for: - Model training (fine-tuning) - Large-scale inference - Faster processing times

CPU-only mode works but is significantly slower for training and inference.

What programming languages are supported?

Primary: Python 3.9+

The Geospatial Studio SDK is Python-based. You can also interact with the platform via: - REST API (any language) - Web UI (no coding required)

Can I use my own data?

Yes! You can onboard your own datasets in several ways:

  1. Upload ZIP files containing imagery and labels
  2. Provide URLs to cloud-stored data
  3. Use Sentinel Hub for on-demand data access
  4. Connect to S3 or other object storage

Supported formats: - GeoTIFF (.tif, .tiff) - Cloud Optimized GeoTIFF (COG) - HDF5 (.h5, .hdf5)

What satellite data sources are supported?

Built-in support: - NASA Harmonized Landsat Sentinel-2 (HLS) - Sentinel-2 (via Sentinel Hub) - Landsat 8/9 (via Sentinel Hub) - Custom data sources via URL

Data connectors: - Sentinel Hub API - Direct file URLs - S3-compatible storage - Local file system

🤖 Model Questions

What foundation models are available?

Pre-loaded models: - Prithvi EO V1 (100M) - NASA/IBM geospatial foundation model - Prithvi EO V2 (300M) - Larger, more capable version - Clay V1 - Self-supervised geospatial model - Custom models - You can add your own

Model capabilities: - Segmentation (pixel-level classification) - Regression (continuous value prediction) - Classification (image-level labels)

Can I use my own models?

Yes! You can:

  1. Upload custom base models to use as foundation models
  2. Import fine-tuned checkpoints from external training
  3. Export trained models for use elsewhere
  4. Register models with the inference service

Models must be compatible with the Terratorch framework (PyTorch-based).

How long does model training take?

Training time depends on: - Dataset size (number of samples) - Model size (100M vs 300M parameters) - Hardware (GPU vs CPU) - Training parameters (epochs, batch size)

Typical times: - Small dataset (100 samples): 10-30 minutes (GPU) - Medium dataset (1000 samples): 1-3 hours (GPU) - Large dataset (10000 samples): 6-12 hours (GPU)

CPU training is 10-50x slower than GPU.

What tasks can I train models for?

Supported tasks: - Segmentation: Flood mapping, burn scar detection, land cover classification - Regression: Biomass estimation, crop yield prediction, temperature mapping - Classification: Land use type, cloud detection, change detection

Custom tasks: You can create custom task templates for specialized applications.

📊 Data Questions

What format should my training data be in?

Required structure: 

dataset.zip
├── image_001_merged.tif    # Input imagery (multi-band)
├── image_001_mask.tif      # Labels/ground truth
├── image_002_merged.tif
├── image_002_mask.tif
└── ...

Requirements: - Matching pairs of data and labels - Same spatial extent and resolution - Consistent band configuration - GeoTIFF format with proper georeferencing

How many training samples do I need?

Minimum: 50-100 samples for basic fine-tuning

Recommended: - Simple tasks: 200-500 samples - Complex tasks: 1000+ samples - Production models: 5000+ samples

Foundation models require less data than training from scratch due to transfer learning.

What image resolution is supported?

Flexible resolution support: - Models work with various resolutions (10m, 30m, etc.) - Input images are automatically tiled if too large - Recommended: 224x224 to 512x512 pixel tiles

HLS data: 30m resolution (Landsat) or 10m (Sentinel-2)

Can I use multi-temporal data?

Yes! The platform supports: - Single-date imagery - Multi-date time series - Before/after comparisons - Seasonal analysis

Specify multiple dates in the temporal domain: 

"temporal_domain": [
    "2024-01-01_2024-01-15",
    "2024-06-01_2024-06-15"
]

🔄 Workflow Questions

What's the typical workflow?

Standard workflow:

  1. Deploy Studio - Set up the platform
  2. Onboard data - Upload or connect to datasets
  3. Explore data - Visualize in UI or SDK
  4. Select model - Choose foundation model
  5. Configure training - Set hyperparameters
  6. Fine-tune - Train model on your data
  7. Validate - Check metrics and visualizations
  8. Deploy - Register model for inference
  9. Run inference - Process new imagery
  10. Analyze results - Visualize and export outputs

Can I skip model training?

Yes! You can use existing fine-tuned models directly:

  • Use models already in the catalog
  • Run inference without additional training
  • Test models on your area of interest
  • Evaluate if further fine-tuning is needed

Fine-tuned models in the catalog work well for common tasks (floods, fires, etc.) but may need additional fine-tuning for specialized applications.

How do I choose hyperparameters?

Start with defaults: - The platform provides sensible defaults - Works well for most use cases

Tune if needed: - Learning rate: 1e-5 to 1e-4 - Batch size: 2-8 (depends on GPU memory) - Epochs: 10-50 (monitor validation loss)

Use HPO (Hyperparameter Optimization): - Automated tuning with Ray Tune - Finds optimal parameters - Requires more compute time

Can I pause and resume training?

Yes! Training can be: - Paused and resumed from checkpoints - Stopped early if validation loss plateaus - Restarted with different parameters

Checkpoints are saved automatically during training.

🌐 Deployment Questions

Can I deploy in the cloud?

Yes! Deployment options:

  1. Local development - Docker Compose on laptop
  2. Single server - VM or bare metal
  3. Kubernetes cluster - Scalable cloud deployment
  4. IBM Cloud - Managed deployment (coming soon)

See Deployment Options for details.

Is it production-ready?

Current status: Beta/Research preview

Production considerations: - Core functionality is stable - Active development and improvements - Community support available - Enterprise support coming soon

Recommended for: - Research projects - Proof of concepts - Development environments - Small-scale production (with testing)

How do I scale for production?

Scaling strategies:

  1. Horizontal scaling - Add more inference workers
  2. GPU acceleration - Use multiple GPUs
  3. Distributed training - Multi-node training
  4. Caching - Cache frequently accessed data
  5. Load balancing - Distribute requests

Kubernetes deployment supports auto-scaling.

What about data privacy?

Data handling: - All data stays in your environment - No data sent to external services (except Sentinel Hub if used) - You control storage and access - Can run fully air-gapped

Security features: - OAuth2 authentication - API key management - Role-based access control (RBAC) - SSL/TLS encryption

🔧 Integration Questions

Can I integrate with existing tools?

Yes! Integration options:

APIs: - REST API for all operations - Python SDK for programmatic access - OpenAPI/Swagger documentation

Data formats: - GeoTIFF output (standard format) - GeoJSON for vector data - COG (Cloud Optimized GeoTIFF)

Visualization: - GeoServer for map services - WMS/WFS standards - Compatible with QGIS, ArcGIS, etc.

Does it work with Jupyter notebooks?

Yes! Full Jupyter support:

  • SDK designed for notebooks
  • Interactive widgets for visualization
  • Example notebooks provided
  • Works with JupyterLab and Jupyter Notebook

Can I export trained models?

Yes! Export options:

  • PyTorch checkpoints (.ckpt)
  • ONNX format (for deployment)
  • TorchScript (for production)
  • MLflow model registry

Models can be used outside Geospatial Studio.

Is there a REST API?

Yes! Full REST API available:

  • All UI features accessible via API
  • OpenAPI/Swagger documentation
  • Authentication via API keys
  • Rate limiting and quotas

API docs: https://your-studio-url/api/docs

💰 Cost Questions

Is Geospatial Studio free?

Yes! Geospatial Studio is: - Open source (Apache 2.0 license) - Free to use - Free to modify - Free to distribute

What about infrastructure costs?

You pay for: - Compute resources (VMs, GPUs) - Storage (disk space, object storage) - Network bandwidth - Satellite data access (if using commercial sources)

Cost optimization: - Use spot instances for training - Scale down when not in use - Use CPU for development, GPU for production - Cache frequently accessed data

Are there any usage limits?

No built-in limits in the open-source version.

Practical limits: - Hardware capacity - Storage space - Network bandwidth - Satellite data quotas (Sentinel Hub)

You control all resources and limits.

📚 Learning Questions

I'm new to geospatial AI. Where do I start?

Recommended path:

  1. Complete this workshop - Hands-on introduction
  2. Read Key Concepts - Understand terminology
  3. Explore examples - Pre-computed datasets
  4. Try simple tasks - Flood or fire detection
  5. Read documentation - Deep dive into features

Additional resources: - Geospatial Studio Docs - Terratorch Tutorials - Additional Resources

What if I get stuck?

Help resources:

  1. Troubleshooting Guide - Common issues and solutions
  2. Documentation - Comprehensive guides
  3. GitHub Issues - Search existing problems
  4. Community - Ask questions, share experiences

See Troubleshooting for detailed help.

Can I contribute to the project?

Yes! Contributions welcome:

  • Report bugs and issues
  • Suggest new features
  • Submit pull requests
  • Improve documentation
  • Share use cases and examples

See Contributing Guidelines

🎓 Next Steps

What should I do after the workshop?

Immediate next steps: 1. Try with your own data 2. Experiment with different models 3. Fine-tune for your use case 4. Share results with community

Long-term: 1. Deploy in production 2. Integrate with existing workflows 3. Contribute improvements 4. Build custom applications

See Next Steps for detailed guidance.

Where can I find more examples?

Example sources: - Workshop notebooks (this repository) - Geospatial Studio Examples - Terratorch Examples - Hugging Face Models

How do I stay updated?

Stay informed: - Watch GitHub repositories - Follow IBM Research blog - Join community discussions - Subscribe to release notes

Key repositories: - Geospatial Studio - Geospatial Studio Toolkit - Terratorch

Still have questions?

← Back: Troubleshooting Next: Next Steps →