Github

Summary

There is a growing need for accurate and scalable techniques to understand, search, and retrieve satellite Earth Observation (EO) images from massive archives, such as the Copernicus archives. In the era of big data, the semantic content of satellite data is much more relevant than traditional keywords or tags. To meet the increasing demand for automation, image search engines that extract and utilize the content of satellite images are necessary, leveraging cutting-edge technologies and advancements in Natural Language Processing (NLP), Machine Learning (ML), and Computer Vision (CV) applied to Earth Observation challenges (NLP4EO). In other words, there is an emerging need to go beyond traditional queries of EO data catalogues, which are based on technical image metadata (such as location, time of acquisition, and technical parameters), and to enrich the semantic content of image catalogues. This will enable a brand new class of query possibilities powered by the combination of NLP (to understand the query and describe the content of the data) and CV (to massively annotate data and implement multi-modal text-to-image and image-to-image searches). Currently, search engines with 'query by content' functionalities do not exist within DIAS platforms or other satellite EO data platforms. Moreover, a Digital Assistant capable of understanding complex requests related to geospatial data searches could significantly expand the dimensions used to query EO data archives. It could also include advanced capabilities to understand and process user requests, select the most suitable workflow to satisfy these requests, autonomously execute processing on EO and non-EO data, and ultimately answer the user's initial question. In this scenario, the development of a precursor demonstrator of a Digital Assistant will adhere to the following high-level objectives:

  • [OBJ-1] Explore innovative CV and NLP techniques for Content-Based Image Retrieval (CBIR), considering both their level of maturity and their applicability to real EO use cases.
  • [OBJ-2] Develop a prototype Digital Assistant that leverages current capabilities for massive processing of EO data. This will be used both for training the content-based query engines and for implementing the prototype Digital Assistant, enabling users to interact with EO data by asking questions in natural language and starting conversations with the Digital Assistant.
  • [OBJ-3] Demonstrate the value of the Digital Assistant in real-life EO use cases to ensure that the demonstrator Digital Assistant can positively impact the user community.
  • [OBJ-4] Engage with the ML, NLP, and CV communities, as cross-disciplinary collaboration is crucial for accelerating the development of innovative solutions.

Architecture

Platform Architecture schema

The digital assistant's high-level architecture is structured into three main components:

  • API. This serves as the entry point for both user queries and the user interface. It is responsible for activating the back-end components that work to return meaningful results to natural language queries focused on the use cases.
  • Content-Based Engines. These are the functionalities directly used by the API for query and search, comprising the heart of the digital assistant solution.
  • Catalogue. This component stores metadata about the different resources that will be used by the API. EO data include SpatioTemporal Asset Catalogue (STAC) metadata, which contain references to the data stored by the engines. Auxiliary data include datasets used as references for semantic purposes, extracted from input images.

Partners