Explainable Deep Learning: Paradigms for Earth Observation

Abstract

The volume and variety of valuable Earth Observation (EO) images as well as non-EO related data is rapidly growing. The open free data access becomes widespread and has an enormous scientific and socio-economic relevance. EO images are acquired by sensors on satellite, suborbital or airborne platforms. They extend the observation beyond the visual information, gathering physical parameters of the observed scenes in a broad electromagnetic spectrum. The sensed information depends largely on the imaging geometry, orbit, illumination and other specific parameters of the space instruments. Typical EO systems can be classified into optical or radar instruments. During the last years, both types of sensors deliver widely different images, and both have made considerable progress in spatial and radiometric resolution, image acquisition strategies, and data rates. Generally imaging sensors generate an isomorphic representation of the observed scene. This is not the case for EO, the observations are a doppelgänger of the scattered field, an indirect signature of the imaged object. This positions the load of EO image understanding, and the utmost challenge of Big EO Data Science, as new and particular challenge of Machine and Deep Learning and Artificial Intelligence (AI). The presentation reviews and analyses the new approaches of EO imaging leveraging the recent advances in physical process based AI methods and signal processing, and leading to explainable paradigms where intelligence is the analytical component of the end-to-end sensor and Data Science chain design. A particular focus is on the semantic aspects as a key component in the explainable learning paradigms.