SNF Ambizione project granted to Dr Devis Tuia (ENAC/IIE/LASIG)

To provide detailed and up-to-date information about the surface of the Earth, data from satellite images need to be processed and modeled before being transformed into intelligible maps. To this end, scientists have used artificial intelligence and machine learning algorithms with great success. However, while they can provide accurate, computationally efficient models for any given image, they are not able to do so for the thousands of high resolution images that are now available. The increase in image resolutions (spatial, spectral and temporal) and the sheer number of sensors acquiring images make it impossible to develop a specific model for each image.

It is thus necessary to develop transferable models. This consists in extracting the intrinsic structure of images (their "manifold"), and in describing the transformation between the manifolds. Here, the shifts in the spectra captured by the sensor, caused by the variations of weather conditions, illumination or geometry, must be identified and compensated for. Finally, using the knowledge acquired, it will be possible to develop models that adapt themselves automatically to new images of a similar nature. It is planned to achieve these tasks by means of manifold and transfer learning methods, two recent machine learning fields.

Dr Tuia's developments will take place in the context of research works conducted at LASIG and IIE (urban planning, land survey, landscape genetics, etc.), and partly relying on remote sensing imagery. The use of satellite imagery in such domains is so far impeded by technical limitations this project proposes to push back.