11-19, 16:00–16:25 (Pacific/Auckland), WA220
Groups have wrangled sufficient number of GPU's to scan through a lot of satellite data to create a good set of weights. This talk scans through a few GeoFM's (Geospatial Foundation Models) and how to put data through them while loading it via OpenDataCube.
This talk presents technical implementation and intercomparison of various geospatial foundation models using TerraTorch , the twist is loading satellite data collections via OpenDataCube tooling rather than built in sample datasets. A bit of pre-processing is necessary to make the foundation models comply with the dynamics of the data as-is from cloud native sources, however they stand up quite well and are generalizable for problems they are trained for. In addition, they can also generate embeddings which can be thought of as dimension reduced versions of source arrays that can then be used to fine tune and perform additional tasks the initial foundation model was not trained for.
This approach is used for demonstrating application of multiple foundation models - Prithvi v2.0, Clay and DOFA for burn scar mapping (Segmentation) and water quality inference (Regression) tasks.
Tisham is a Senior Engineer at CSIRO Space and Astronomy division. Previously he held similar roles in various government agencies, private sector companies and startups. He has been pushing pixels since 2004 when he got started using medical imaging devices in Singapore. Since then he has had the priviledge of working with various space agencies such as NASA, DLR, JAXA, CNES etc. and lived through the explosion of publicly available satellite imagery.