Day 4

Detailed paper information

Back to list

Paper title AI-Cube: Combining Datacube Scalability with AI Intelligence
Authors
  1. Peter Baumann Jacobs University Bremen gGmbH Speaker
  2. Begüm Demir TU Berlin
Form of presentation Poster
Topics
  • C1. AI and Data Analytics
    • C1.07 ML4Earth: Machine Learning for Earth Sciences
Abstract text In the AI-Cube project datacube fusion and AI-based analytics will be integrated, demonstrated in several real-life application scenarios, and evaluated on a federation of DIASs and further high-volume EO / geo data offerings.
Starting point is the observation that both Machine Learning (ML) and datacube query languages share the same basis, Tensor Algebra or – more generally – Linear Algebra. This seems to provide a good basis for combining both methods in a way that datacubes can be leveraged by ML better than scene-based methods. The expected benefits include simplification of ML code, enhanced scalability, and novel ways of evaluating spatio-temporal data.
AI-Cube approaches this from both sides: adjusting ML to datacubes and enhancing datacubes with specific operational support for ML model training and application. As to the first part, the project will develop multi-cross-modal AI methods that:
• effectively learn the common representations for the heterogeneous EO data by preserving the semantic discrimination and modality invariance simultaneously in an end-to-end manner.
• consist of intermodality similarity-preserving learning and semantic label-preserving learning modules based on different types of loss functions simultaneously.
• include an inter-modal invariance triplet loss and inter-modal pairwise loss functions in the framework of the cross-modal retrieval problems.
The Big Data aspect is underlined by tapping into the BigEarth.Net collection of 590,000 labelled Sentinel-1 / Sentinel-2 patch pairs for versatile model training. These models will then be used on the 30+ PB of Sentinel datacubes offered by rasdaman on Mundi, Creodias, and further members of the EarthServer datacube federation.
From the database perspective, novel operators will be added to the query language to embed AI into datacube query languages like SQL/MDA and OGC WCPS. Also the models themselves will be stored and handled as datacubes.
Goal is to support scenarios like the following: User selects a topic (such as specific crop types, specific forest types, burnt forest areas). System determines, through a combined analysis of various large-scale data sources, a list of regions showing the criterion selected. User gets this visualized directly or continues analysing, possibly combining with further data sources. Real-life application scenarios will be exercised in the DIASs of the EarthServer federation, doing both single datacube analytics and distributed datacube fusion.
The consortium consists of Jacobs University as coordinator, TU Berlin, and rasdaman GmbH. AI-Cube has commenced in Fall 2021, and first results will be presented at the symposium.

Acknowledgement
This work is supported by the German Ministry of Economics and Energy.