CoSIA and FLAIR-HUB: multi-source AI models for land cover mapping Bookjans E. , Garioud A., Marchand G., Vo Quang A., Dekeyne F. and Masse A.
Institut national de l^information geographique et forestiere (IGN), France
IGNFI, France
Abstract
The French Mapping Agency (IGN) has invested in deep learning methods and technologies to accelerate the production of high-resolution land cover maps to better monitor the evolution of the French territory, i.e. to map the Anthopocene. This initiative, named CoSIA (Land Cover by Artificial Intelligence), is crucial for monitoring the evolution of the French territory and improving land management. More than 2,800 km2 of aerial imagery have been annotated by image analysts, enabling IGN to train a robust semantic segmentation AI model applicable at the national scale. In parallel, the IGN launched the FLAIR (French Land cover Aerospace ImagRy) challenges to help improve its AI land cover models. To support continued innovation, experimentation and encourage collaboration, the IGN has not only made this dataset and its AI models publicly available but has recently augmented it with multi-modal data: FLAIR-HUB also contains satellite images, historical aerial images (1950s) and crop type annotations (23 agricultural classes). The resulting land cover maps CoSIA find a wide range of applications in environmental monitoring, land use planning and resource management (e.g., change detection, forestry, hydrology, agriculture). The IGN already exploits CoSIA for several different applications, including quantification of artificial surfaces, identification of green spaces in cities, and monitoring agricultural features like hedges.. In summary, CoSIA and FLAIR-HUB provide geospatial communities with valuable resources and tools allowing for a better overall land management in the face of the current technological, socio-economic and environmental challenges.
In complement, IGNFI, the international subsidiary of IGN, is providing for close to 40 years expertise and services to help decision-makers around the world, including Indonesia, to fully own and take advantage of the potential of geographical information, including Land Cover Mapping. Results will be presented during
Keywords: Land Cover Mapping, Deep Learning, Aerial Imagery, Satellite imagery, CoSIA
