This section shows how to train and apply machine learning models in openEO workflows. The focus is on feature engineering, backend-side inference, and reusable model deployment patterns using UDFs and UDPs.
- training ML models on EO data without moving everything locally
- building task-specific feature cubes from Sentinel imagery and ancillary data
- deploying models with ONNX or UDP-based inference patterns
- reusing trained models for large-scale prediction jobs
| Notebook | Data source | Key openEO features | Description |
|---|---|---|---|
| dynamic-land-cover-mapping/dynamic-land-cover-mapping.ipynb | Sentinel-2 + SAR | UDFs for features, MlModel API |
Train a Random Forest for dynamic land cover and run inference. |
| extracting-training-data/ML_ready_data_extraction.ipynb | Sentinel-2 L2A + LUCAS 2022 | feature extraction, ground-truth sampling | Prepare ML-ready training data from EO imagery and reference samples. |
| parcel-delineation/parcel-delineation.ipynb | Sentinel-2 | UDF, pretrained U-Net | Deploy a U-Net to delineate agricultural parcel boundaries. |
| onnx-inference/onnx-ml-inference.ipynb | Sentinel-2 | UDF with ONNX runtime | Run a pretrained CNN in ONNX format at datacube scale. |
| random-forest-forest-fire/random-forest-training.ipynb | Sentinel-2 + SAR | GLCM UDFs, RF training, model persistence | Full training workflow for a forest-fire Random Forest model. |
| random-forest-forest-fire/random-forest-inference-udp.ipynb | Sentinel-2 + SAR | UDP creation, model reuse | Wrap the trained model as a shareable UDP for scalable inference. |
- build or select a feature cube from EO data
- train a model, or load a pretrained one
- run inference with a UDF, ONNX session, or UDP
- deploy the model for reuse across multiple areas or time periods
Additional background is available in:
- random-forest-forest-fire/README.md
- dynamic-land-cover-mapping/README.md
- parcel-delineation/README.md
This section is a good fit when you want to move from exploratory processing to actual predictive workflows in openEO.
Once you are working with ML-oriented workflows, a useful progression is:
- 60 · Geospatial Embeddings if you want learned feature spaces and embedding-based representations.
- 70 · Platform and Large Scale if you need to scale up training, inference, or job orchestration.
- 50 · Thematic Notebooks if you want domain-specific case studies using the same openEO patterns.
The best next step depends on whether your priority is richer feature representation, operational scale, or a thematic application.