Back to search

BIA-Brukerstyrt innovasjonsarena

ATELIER-EO: Automated machine learning framework tailored to Earth Observation

Alternative title: ATELIER-EO: Automatisert maskinlærings-rammeverk skreddersydd til jordobservasjon

Awarded: NOK 1.8 mill.

Earth Observation (EO) covers everything from drones to satellite imagery, and its applications have a huge impact on everyday life: predicting weather, tracking container ships, monitoring agriculture and so much more. These applications require the processing of an enormous amount of data, which can only be processed with automated methods. Deep learning (DL), the most successful branch of artificial intelligence (AI), relies on both a high amount of carefully curated data and specialised models. This leads to a problem: experts are required both for data collection and for designing appropriate DL models. Therefore, if a small company would like to utilise AI for customised mapping, that would require significant expertise and cost. The goal of the ATELIER-EO (Automated machine learning framework tailored to Earth Observation) project was to develop an automated artificial intelligence solution for EO, with more effective learning schemes able to reduce the amount of required data, and automated machine learning reducing the necessary expertise. As an example, the automatic mapping of forest roads is currently very challenging: roads are small, and while a human knows they are connected, the tree cover often hides parts of the road. ATELIER-EO was foreseen to be able to automatically handle this connectivity, as well as incorporate terrain models and mitigate other issues like clouds in the images, requiring training images where example roads were present as the only input. In conclusion, ATELIER-EO's goal was to automatically create customized maps from satellite or drone images with limited data and minimal human interaction, broadening the domains where AI can be applied, and reducing the cost of customised products.

With the R&D performed within this project, ST has paved the way to build a solid foundation for a significant step forward in its technical roadmap and a consequent expansion of its portfolio of AI-based EO products, both for research and commercial customers. During the project, the gained capabilities (all internally tested) were the following: • Rapid prototyping for EO-related segmentation problems. It is possible to set up a model in less than 1 working day, and provide a trained model within 24 hours (Example: combined building detection and forest segmentation). • Capability for making a national forest map (Norway, 1m resolution) from LIDAR data within 2 days of processing time, if the above-mentioned model is available. • Tree-based hyper-parameter search for complex configuration spaces. • Template processing system for AutoML framework generation. • Noise robust training with robust loss functions and with mislabel detection. • LIDAR-based pretraining for building detection. The foundation built within ATELIER-EO will allow S&T to access new markets, reduce costs and increase the performance of existing products. The overall goals of the project were in line with Norway’s National Strategy for AI. Furthermore, the inheritance of this project will contribute to several UN development goals (eg. 13, 15) by further developing non-invasive monitoring methods, which were (and will be) applied to hot topics such as land use (i.e. building detection) and deforestation. Thanks to the developments achieved within ATELIER-EO, S&T expects the impacts in a stimulation of the public and private sectors to EO-applied AI, economic viability by offering cheaper, more reliable, faster-to-market, customised DL solutions, and integration into the customer’s existing business processes.

Science and Technology AS (ST) is a young company specialised in the development of services using Artificial Intelligence (AI) applied to EO and onboard processing systems for satellites. Most notably, we focus on the efficient use of satellite data for sustainable EO services. The share of AI is continuously growing in ST’s portfolio. The company has ample experience in the execution of large scientific data processing and intelligent software application projects for both ESA, EU, and national clients within, but not limited to, the domains of forestry (e.g. disease detection, deforestation, tree species) and ice sheet monitoring (e.g. calving front lines, mapping of supraglacial lakes, snowpack properties). DL has already shown great success in the areas mentioned above, despite two major bottlenecks: limited EO-specific domain knowledge within the DL field (and vice versa, limited DL knowledge in the EO field), and a long time to market. Together with the large amount of data often required, these bottlenecks currently limit ST’s commercial opportunities and the adoption of AI within the EO field both in Norway and internationally. This R&D project aims at extending and streamlining ST’s current DL technology stack, transforming it into an EO-specific, AutoML solution. Within this context, AutoML automates the tasks of applying ML to real-world problems, covering a complete pipeline from the raw dataset to the deployable machine learning model. Critical milestones of this project are related to the research required to reach the goals described above and incorporated in the progress plan.

Funding scheme:

BIA-Brukerstyrt innovasjonsarena