Area Statistics Deep Learning (ADELE)
Summary
The FSO’s land use statistics are an invaluable tool for long-term land observation. This project involved learning and mastering the use of artificial intelligence (AI) technologies to eventually automate (even partially) the visual interpretation of aerial images in order for change detection and classification. The project ran from 2017 to 2021 and allowed, from 2022, the automatic classification of about a quarter of sample points with an accuracy satisfying the very high requirements of public statistics.
Description
Among the learning methods that have emerged from AI for image recognition, Deep Learning is particularly adapted to statistics on land use and cover for which learning data are available in very large quantities.
The project was carried out in two phases. The first phase consisted of developing a computer prototype and the second one consisted of implementing it in the land use statistics production system. The resulting observations and evaluation criteria are detailed in a production transfer report.
Thanks to the knowledge and experience accumulated during this project, new developments are planned in the medium term in order to improve the performance, as well as the integration of AI tools in the production environment.
Objectives
The following objectives were achieved:
- Develop collaborators' skills in remote sensing and AI technologies.
- Characterize future principal data sources and integrate those determined to be adequate.
- Identify and utilize the potential for reducing the interpretation workload through AI technologies and artificial intelligence.
