• Gabriele Moser, Full Professor, RIOS Lab Head, University of Genoa. Website: https://www.rioslab.it/

Title: Fully unsupervised heterogeneous change detection with multitemporal remote sensing imagery

Abstract: Change detection has been a fundamental remote sensing image analysis task for long, with prominent applications in environmental monitoring, damage assessment after natural disasters, and climate change mitigation. Indeed, traditional change detection techniques typically rely on the assumption that the input multitemporal imagery have been taken using the same (or almost the same) acquisition modality – same sensor, same acquisition geometry, same spectral bands, same radar frequency and polarization, etc. From an application-oriented viewpoint, this could be highly restrictive, especially in the framework of rapid response after a natural disaster. The opportunity of formalizing change detection with input multitemporal data characterized by heterogeneous modalities has been studied more recently, and especially interesting results have been obtained lately using methods based on deep image-to-image translation. In this talk, the problem of heterogeneous change detection will be discussed, with special focus on the particularly challenging case of fully unsupervised detection. First, the main methodological approaches that have been presented in the literature, ranging from earlier semi-parametric regression to current deep generative models, will be recalled. Then, recent advanced fully unsupervised methods for heterogenous change detection, which are based on the combination of deep image-to-image translation, spectral clustering, and dimensionality reduction, will be presented. Examples of experimental results will be shown in applications involving heterogeneous change detection from multispectral-SAR and hyperspectral-SAR multitemporal pairs.

• Jonas Elm, Assistant Professor, Department of Chemistry, Aarhus University. Website: https://chem.au.dk/forskning/forskningsomraader-og-grupper/teoretiskkemi/computational-atmospheric-chemistry

Title: Machine Learning Approaches for Studying Atmospheric Molecular Cluster Formation

Abstract: The formation of molecular clusters is an important step in atmospheric new particle formation. Despite much research within the field, the exact contribution of different vapors remains highly uncertain. Quantum chemical (QC) calculations and kinetics modelling allow the detailed mapping of cluster formation pathways, but such calculations are extremely computationally demanding. In addition, there is a significant gap in the size between the currently modeled clusters using highly accurate QC methods (around 1 nm) and aerosol particles routinely measured in the field and the laboratory (around 2 nm). Hence, the transition from small clusters to freshly nucleated particles (FNPs) remain completely unexplored. 

In this presentation I will outline our recent work on the comprehensive screening of atmospherically relevant cluster systems. This allows us to gain direct insight into the cluster formation potential of each system and hence their ability to grow into larger sizes. Furthermore, I will present how to apply machine learning models to accelerate the exploration of the configurational space in cluster formation studies. I will present our current work involving the studying of FNP compositions. The presented work will bridge the current theoretical and experimental work in the field allowing for molecular level cluster formation to be directly implemented into atmospheric aerosol process models.