Probabilistic graphical models and deep learning for remote sensing image analysis

Dr. Martina Pastorino, Assistant Professor at the Department of Electrical Electronic Telecommunications Engineering and Naval Architecture, University of Genoa gives a lecture titled: Probabilistic graphical models and deep learning for remote sensing image analysis.

Abstract:

Deep learning has become the dominant paradigm for image processing tasks thanks to its non-parametric formulation and powerful feature extraction capabilities. At the same time, probabilistic graphical models such as Markov random fields (MRFs) and conditional random fields (CRFs) provide principled tools for structured prediction. Depending on the underlying graph topology, these models can effectively capture spatial and multiresolution dependencies.

This presentation explores the connections between these two major modeling paradigms – ranging from simple integration strategies to end-to-end formulations with theoretical equivalence results, as well as intrinsically integrated approaches – and discusses how these connections can be leveraged for remote sensing image analysis.

Experimental results on multimodal satellite and aerial imagery demonstrate the effectiveness of these methods for structured prediction, multimodal image fusion and segmentation, and image generation tasks, highlighting the potential of combining probabilistic modeling with neural architectures for remote sensing applications.

The lecture is organized by the Faculty of Electrical and Computer Engineering, University of Iceland and IEEE in Iceland.