Summary
QueryPlanet aims at democratizing the access of Artificial Intelligence to the Earth Observation community by developing open-tools for the creation of AI-ready EO dataset and use-cases that leverage such datasets to build global insights applications. Material and datasets develop under QueryPlanet are open-access.
The project aims at providing open-source tools for the exploitation of Earth Observation data, in particular of Sentinel-2 imagery.
The target audiences of the project are the EO and AI communities, fostering their engagement in exploiting EO dataset, in particular Sentinel-2, to build applications that tackle relevant topics.In the first phase of the project, tools to annotate Sentinel-2 imagery are developed, allowing any user to set-up and share with the community their labeling campaign. Following the creation of the label datasets, the creation of the EO processing workflow is facilitated by eo-learn, the open-source Python package developed within the project. eo-learn provides common processing tasks to scale the analysis of satellite imagery to global scale through seamless parallelization.
To further promote the upake of the tools created, AI-ready dataset and use-cases capitalising on such datasets are created and published. Currently developed use-cases include:
- super-resolution of Sentinel-2 bands beyond the 10 metre spatial resolution, using the HighResNet multi-frame super-resolution algorithm and VHR imagery as target reference. The training dataset for such algorithm is globally sampled and includes humanitarian targets;
- a hierarchical object detection scheme which uses several data sources with increased spatial resolution to detect buildings in an efficient and scalable way. The hierarchical scheme in this case uses Sentinel-2, Airbus SPOT and Airbus Pleiades imagery to perform object detection using rotated bounding boxes;
- a pan-European map of forest and forest types using Sentinel-2 time-series. The developed algorithm is based on the latest deep learning architectures for the analysis of spatio-temporal datasets, and uses tens of thousands of samples collected over the EEA countries area.
- Open-source tools to ease labelling of EO imagery
- Open-source tools to process EO imagery for the creation of AI-ready datasets
- Publishing of AI-ready datasets covering a wide range of thematics
- Publishing of material to create use-case applications based on AI to extract insights from the EO datasets
- Facilitate entry to the field to non EO experts
