The adoption of deep learning based algorithms on satellite imagery needs high quality and reliability. Metrics needs to be published and audited by various parties to build trust in the technology. Errors and edge-cases need to be analyzed and shared with potential users in order to find corrective actions. In turn, this will support explanability and ethical practices that are also needed for larger adoption.
Building a object detector on satellite imagery cannot be achieved without placing data at the center of the development process. Tagging and data-mining are key elements of the final product. A data-centric workflow enables to label exactly the data which needs to be labelled in order to make the model better at every cycle. This means that a MLOps framework is used and that it includes the human annotator in the loop. This enables to continuously improve the model.
Some organizations can use pre-defined algorithms but some others will need to adapt the algorithms to their specific use-case. It could be a specific classification, a specific type of imagery (bands, resolution, geometry). One size does not fits all in Earth Observation. Consulting, mentoring and training are very important to increase knowledge and know-how in all industries using satellite imagery. That is also why open standards, open source, sharing knowledge and building community are key elements that DL4EO wants to foster.
The ship detection algorithm detects ships and barges on Sentinel-2 images at 10 m. resolution. It has been trained on a 10,000 ships dataset based on Sentinel-2 images from all over the world including harbors, rivers, military sites and high seas.
Vessels detection on Sentinel-2
The infrastructure identification on Sentinel-2 is based on the extraction of feature vectors from the imagery and classification in one of the following classes - airfields, harbors, POL storage, industrial sites, urbanization, residential housing, agriculture, forests.
Infrastructure identification on Sentinel-2
The oil storage algorithm dectects POL (Petroleum, Oil and Lubricants) tanks on visible satellite imagery (typically SPOT imagery but any other type of medium resolution imagery). It has a minimum threshold of 5 pixels for the diameter of tanks.
Oil storage tanks detection
Offering custom analytics development through deep learning on satellite imagery is the core of the activity of DL4EO. Depending on the size of the project, we sub-contract a small or large part of the project while keeping the overall responsibility, project management as well quality management. Typical steps are scoping, data annotation, data-centric model creation, containerization, qualification and delivery.
Custom DL4EO model development
This is typically a 3 to 6 months project which start by the definition of the objectives with the stakeholders and finishes with the delivery of a production-ready container (Docker) which detects objects in satellite imagery. All the anotations work and data-centric model building is created internally and provided to the client.
Applying Deep Learning to the specifics of satellite imagery is a real passion! From taking avantage of hyperspectral bands, to leveraging synthetic images for rare objects detection to the automatic extraction of height and other object characteristics, this is my core expertise at the convergence of physics, computer science and customer-oriented problem-solving skills.
Hackathons and challenges are sometimes a very good way of getting new idea for solving a specifically difficult problem; it also enables to build an ecosystem of partners and potential subcontractors. I can support such an initiative during the various stages of the project from inception, data preparation, mentoring, selection of the winning teams and integration.