Objectives of the Product

Developing and deploying neural network-based computer vision models for geospatial data is challenging for most organisations. Typically, AI experts customise a model architecture and train it using specific datasets provided by the customer, aiming to detect predefined targets. After the model is trained, it is deployed as a “black box,” limiting the user’s ability to improve detections or add new targets without extensive re-engineering and re-training, which incurs significant time and costs. This process is not only inefficient but also poses security risks, as training data must be accessed by external engineering teams, complicating data traceability and increasing the risk of unauthorised data access and leakage. Furthermore, the lack of thorough traceability can lead to data poisoning, facilitating various attacks like adversarial attacks, label flipping, and model inversion, which can undermine the model’s integrity and potentially expose original training data.

Building on the precursor de-risk activity “SatHound: Multi-object detection solution based on artificial intelligence for non-expert EO users – eo science for society”, DVSTAI offers a new Earth Observation (EO) tool to ease the daily work of EO analysts and non AI experts, an object detection solution that end-users can configure autonomously by creating, training and deploying their own AI models to perform vision tasks over satellite imagery.

---

Customers and their Needs

Describe in <200 words the targeted customers (and/or users), how they are involved in the activity, or using the product that will be developed. Define the identified users’/customers needs and the challenges trying to meet these needs.

Target customers are:

• Governmental agencies
• EO Satellite operators
• EO Service providers
• EO Analysts

DVSTAI targets niche markets of governmental and private entities operating Earth Observation satellites that require user-level analysis tools to obtain fast results on specific tasks. It is expected that such customers are owners of geospatial data and expect to use DVSTAI within their own infrastructure, potentially integrated with other processing tools in their data management process.

DVSTAI also targets to address mass markets via EO service providers, that can offer both data provisioning and analysis tools. DVSTAI can be offered bundled with existing EO services to enable their customers to create specific vision tasks over available data sets in the platform to extract additional value while keeping full flexibility of a Software as a Service model.

---

Targeted customer/users countries

DVSTAI targets the customer segments described above globally.

---

Product description

The product is designed to enable end-users to perform object detection on various types of satellite images across different spatial and temporal selections from the catalogue. It features a graphical interface that allows users to label objects of interest, train new or refine existing specialised detection models, manage trained models, administer catalogues, and conduct searches within the catalogue using one or more specific models.

The main differentiation objectives are of DVSTAI are:

• Allowing end-users to create and manage object-detection and other vision neural networks trained to detect their target of interest in satellite imagery
• Allowing end-users to perform target detection for multiple objects in the same area of interest and on multiple search areas simultaneously
• Allow end-users to monitor an area of interest over time, by scheduling the processing of new images available in the catalogue and easily comparing results with previous searches.
• The tool is usable by non-expects (in Artificial Intelligence, ML pipelines or remote sensing) with a good user experience
• The user can work with public data or their own optical, multispectral, SAR; and manage different band combinations
• The tool supports different AI model architectures and is extendable for new ones
• Allow Software as a Service pay-per-use model by data segregation, customer on-boarding and multi-tenancy

---

Added Value

The idea behind DVSTAI is to ease the daily work of EO analysts, but also to offer the general public an object detection solution that end-users can quickly set-up and use as a service to analyse public geospatial or their own data.

The solution offers an end-to-end architecture for data management, labelling, training and inference, that allows users with no or limited knowledge of AI and EO business to implement their own models on top of the platform. DVSTAI models can be tailored to fit any specific use cases where vision is involved such as object detection, change detection, semantic segmentation, etc.

DVSTAI provides added value to its customers by:

• Reducing the knowledge gap with a tool where domain knowledge is the only thing needed to create/configure AI vision tasks.
• Facilitating trainings for new targets after annotating a few tens of representative samples. Re-trained models are automatically ready for production use.
• Providing a unified set of performance metrics, automatically generated on each new version of the model after training. Its model version management allows to roll-back to previous versions if required.
• Managing data governance by secure access control to the solution HMI and APIs. Removing the need for manual data extraction and processing outside the platform that could lead to leakage.
• Native integration with service providers in cloud-based platforms that offer collections of data catalogues from open and commercial sources

---

Current Status

The DVSTAI InCubed project started in June 2025 and the first project milestone (MVP definition) is scheduled for September 2025.