Difference between revisions of "Sprint bdfs25"
| Line 24: | Line 24: | ||
3. '''Machine learning approaches''' | 3. '''Machine learning approaches''' | ||
| − | + | PANGAEA benchmark shows that specialized, not-CNN and not-ViT , machine learning models can perform better than current (trendy) "Foundation Models" for remote sensing data | |
| + | supervised learning, XGBoost / Catboost or various Random Forest learning | ||
| + | perform better than ... ViT variations, CNN for landuse / landcover and other specific classifications | ||
| − | these (https://huggingface.co/ibm-esa-geospatial/TerraMind-1.0-base) mixed models are wild, | + | these (https://huggingface.co/ibm-esa-geospatial/TerraMind-1.0-base) mixed models are wild, |
like this one https://github.com/swiss-territorial-data-lab/proj-vit [DATA\_LINK](https://huggingface.co/datasets/heig-vd-geo/M3DRS) | like this one https://github.com/swiss-territorial-data-lab/proj-vit [DATA\_LINK](https://huggingface.co/datasets/heig-vd-geo/M3DRS) | ||
Revision as of 12:44, 26 September 2025
TorchGeo an OSGeo Project CODE DOCS
Additional Topics:
1. Workshops
Berkeley Climate AI Day LINK
ml4earth.de workshop featured a keynote by Prof. Dr. Xiaoxiang Zhu from TUM Germany, also home to the creator of TorchGeo Adam Stewart PhD POSTDOC. Applications of artificial intelligence (AI) in Earth observation, with a focus on machine learning (ML) approaches for remote sensing.
2. PANGAEA project
a standardized evaluation protocol that covers a diverse set of datasets, tasks, resolutions, sensor modalities, and temporalities. It establishes a robust and widely applicable benchmark for Geospatial Foundation Models (GFMs).
https://bpa.st/BTQAE ## code updates
https://eotdl.com/blog/pangaea
[PANGAEA benchmark](https://github.com/VMarsocci/pangaea-bench) shows that specialized, not-CNN and not-ViT , machine learning models can perform better than current (trendy) "Foundation Models" for remote sensing data
3. Machine learning approaches
PANGAEA benchmark shows that specialized, not-CNN and not-ViT , machine learning models can perform better than current (trendy) "Foundation Models" for remote sensing data supervised learning, XGBoost / Catboost or various Random Forest learning perform better than ... ViT variations, CNN for landuse / landcover and other specific classifications
these (https://huggingface.co/ibm-esa-geospatial/TerraMind-1.0-base) mixed models are wild,
like this one https://github.com/swiss-territorial-data-lab/proj-vit [DATA\_LINK](https://huggingface.co/datasets/heig-vd-geo/M3DRS)
https://iclr-blogposts.github.io/2025/blog/sparse-autodiff/
4. Data availability
The conversation likely emphasized the importance of public datasets for Earth observation research. For instance, projects like TerraMesh, funded by ESA (European Space Agency), provide valuable data sources for researchers and developers.
hm - on the US West Coast side - team MSFT, Planet Labs and The Nature Conservancy .. [D2](https://reatlas42216storage.blob.core.windows.net/public/wind_all_2024q2_3_11_2025.gpkg) [LINK](https://www.microsoft.com/en-us/research/wp-content/uploads/2025/03/Global-Renewables-Watch_Caleb-Robinson_2025.pdf) [DATA0]
project 2020 "[TerraMesh](https://openaccess.thecvf.com/content/CVPR2025W/EarthVision/html/Blumenstiel_TerraMesh_A_Planetary_Mosaic_of_Multimodal_Earth_Observation_Data_CVPRW_2025_paper.html) is part of the FAST‑EO project funded by the European Space Agency Φ‑Lab (contract #4000143501/23/I‑DT)."
hey - public training [data](https://x-ytong.github.io/project/Five-Billion-Pixels.html) for China interior, by Xiou xiang Zhu .. who is very much in Germany and is the boss of the department with the `torchGeo` guy
gdal containers LINK
opendatacube containers LINK
http://dataspace.copernicus.eu/ https://eotdl.com
5. Geospatial data infrastructure
References to geospatial data infrastructure initiatives such as STAC (SpatioTemporal Asset Catalog) and OSGeoLive were likely discussed, highlighting the need for standardized data formats and efficient data access mechanisms.
Key points and takeaways:
Specialized ML models: Traditional ML methods can outperform trendy AI approaches in remote sensing tasks.Public data availability: Public datasets are essential for Earth observation research, enabling collaboration and innovation. Collaboration opportunities: The conversation likely touched upon the potential for international collaborations across different countries and regions. Geospatial data infrastructure: Standardized data formats and efficient data access mechanisms are crucial for geospatial research.
Preparation for Big Data from Space #osgeo code sprint
The participants' discussion is a preparation activity for an upcoming code sprint, where they will work together to develop innovative solutions using remote sensing and Earth observation datasets. A wiki page summarizing these ideas could be a valuable output from this collaboration.
Overall, the conversation highlights the intersection of geospatial technology, machine learning, and open science initiatives in Earth observation, emphasizing the importance of data availability, specialized ML models, and geospatial data infrastructure for advancing research and innovation in this field.