Hao Li was presenting aspects of our AtlasHDF infrastructure (Werner & Li, 2022) which was explored for surface water segmentation in the context of NFDI4Earth.

In a nutshell, we propose a shift in geospatial big data from Geo to NoGeo quite in the same way as the big data community sacrificed the SQL language in big data towards NoSQL infrastructures. OGC standards and their libraries are valuable assets for data management, interpretation, and preservation. But they have never been designed for computation. While our approach is less powerful in spatial operations at the moment, we have zero dependencies beyond HDF5. But this is mature, stable, and - most importantly - an integral part of both tensorflow and pytorch. The philosophy of this approach is simplicity: our data can be used by every deep learning scientist out of the box. And the best: supercomputers read it in parallel…

Ressources

• Github Repository
• Github Mirror at End of Project

1. Werner, M., & Li, H. (2022). AtlasHDF: An Efficient Big Data Framework for GeoAI. 1–7. https://doi.org/10.1145/3557917.3567615 [PDF] [Online] [BibTeX]

   @inproceedings{2022_atlashdf_Werner,
     author = {Werner, Martin and Li, Hao},
     title = {AtlasHDF: An Efficient Big Data Framework for GeoAI},
     year = {2022},
     isbn = {9781450395311},
     publisher = {Association for Computing Machinery},
     address = {New York, NY, USA},
     url = {https://doi.org/10.1145/3557917.3567615},
     doi = {10.1145/3557917.3567615},
     pages = {1–7},
     numpages = {7},
     keywords = {GeoAI, OpenStreetMap, big data, hierarchical data format, immediate mapping},
     location = {Seattle, Washington},
     series = {BigSpatial '22}
   }