WO8 - Integrate Delineate Anything algorithm

[JanssenBrm]

The scope of this ticket is to integrate the Delineate Anything algorithm into the APEx Algorithm Catalogue. Our proposal to ESA is to dedicate an 80 man-hour effort to port the algorithm using the following approach:

• Given the importance of cloud-free input data, we suggest reusing the existing Best-Available Pixel APEx service to generate input for the model. This approach would not only optimise the results but also demonstrate the UDP-in-UDP concept.
• Spend a best-effort approach on the post-processing step, which is necessary to (re)combine the resulting geometries after parallelisation in openEO. This would allow us to achieve a nice result in the short term, with a new Delineate Anything service, and, if needed, analyse the additional effort required to finalise post-processing.

Resources

• GitHub repo: https://github.com/Lavreniuk/Delineate-Anything

Notes

These are a couple of notes from previous discussions with Mykola (maintainer):

• Mykola explained that the delineate anything algorithm is designed for field boundary delineation, aiming to provide vector data for agricultural applications at a global or European scale.
• Mykola clarified that the delineate anything model is based on YOLO V11, not Segment Anything, and is trained on diverse satellite data sources, including Maxar, Planet, Sentinel 2, and Google satellite imagery.
• Mykola described that the model supports both single-date and composite runs, with automatic normalisation to handle varying data ranges and cloud coverage, allowing users to select their preferred time period and input data.
• Mykola shared that the delineate anything model and its weights have been uploaded to GitHub and Hugging Face, enabling users to run the model on their own data via Colab
• Mykola explained that users should select their own data for running the model, as results depend on the chosen input and there are no restrictions on data selection.
• Mykola confirmed that the model is PyTorch-based, with support for conversion to other formats such as TensorRT and ONNX.
• Mykola described the post-processing approach, including patch-based processing (512x512 pixels) and the use of the DSU algorithm to recombine patches and avoid edge artefacts.
• Mykola clarified that the model and its weights are publicly available, hosted on Hugging Face, and governed by the AGPL licence, which allows research use but not commercial use.

Acceptance Criteria

• The service is integrated in an openEO UDP
• The service is integrated in the APEx Algorithm Catalogue
• The service is benchmarked