During the second reporting period, the remaining use-cases have been finalized. The use-cases have been employed to validate the GEM framework, used to drive the additional development, provided feedback, and served as a playground for various optimisations of the framework. While we have defined the AOIs and prepared majority of ground truth (GT) data already during the first reporting period, the second reporting period focused on demonstrating and validating the capabilities of GEM framework through the use-cases. Various approaches to change detection were also explored and used within use-cases. The use-cases demonstrated that it is possible to continually run models at low resolution at global scale in a repeated manner and at low costs, making cost-effective continuous monitoring a reality. We reflected on what can be achieved within the cost-effectiveness at scale constraint and showed drill-down and continuous monitoring approaches for the purpose of monitoring large areas (at continental scales), demonstrating clear (cost) benefits of drill-down mechanism.
Service for running large-scale, cost-effective, and continuous monitoring has also been developed with GEM framework during the second reporting period and presented in D5.7 - Demonstration deliverable, on a illustrative example of monitoring NDWI anomalies through time. The service was run for each Sentinel-2 observations since January 1, 2023, on a daily basis, and anomalies were detected with respect to historical data.
Novel ML approaches, methodologies and tools have been explored, and proof-of-concepts were investigated. Discussions about their usability within various use-cases pinpointed where (in use-cases) the novel approaches would work best. These studies were added to eo-learn-examples repository, extending eo-learn functionalities, particularly by bringing eo-learn and PyTorch closer together in a form that fosters easy and fast integration into research or inference pipelines.
We have built GEM processing framework (both eo-learn and eo-grow) to be cloud infrastructure agnostic. Particularly eo-learn is platform independent, meaning that the workflows can run on any infrastructure. During the second review period we have explored various cloud infrastructures for running the workflows, and openly shared several detailed examples and instructions on how to make use of GEM framework not only on AWS, but also on Copernicus Data Space Ecosystem (CDSE). Although commercially not as viable, and with (significantly) reduced capabilities, we understand the importance of CDSE and have spent significant effort preparing instructions how GEM framework can be leveraged there, despite CDSE not yet providing a commercial offering.
The results from use-cases have been made into data cubes, and are available through demonstration applications for general public, and can be accessed through GEM framework and Sentinel Hub services. A number of examples have been published on eo-learn-examples repository to facilitate uptake of the results.