A Mathematical Framework for Uncertainty Quantification in Multimodal Image Analysis via Probabilistic Clustering Models

• Activity Number: 630 - Uncertainty Quantification, Reliability and Robust Inference
• Type: Contributed
• Date/Time: Thursday, August 2, 2018 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Defense and National Security
• Author(s): Maximillian Chen* and David John Stracuzzi and Michael Christopher Darling
• Companies: Sandia National Laboratories and Sandia National Laboratories and Sandia National Laboratories
• Keywords: probabilistic clustering models; nonparametric mixture model; Gaussian mixture model; uncertainty quantification; multimodal data; data integration

Abstract

Recent years have witnessed an explosion in the availability and application of remote sensor data to detecting and classifying objects and events. With the increase in data availability comes new opportunities to improve analytic results and reduce uncertainty by combining the data from multiple sensors that observe the same region. The sensors can provide complementary, but differing information, and the data oftentimes cannot be represented by well-defined and well-known parametric distributions. This potentially negates the validity of using the Gaussian Mixture Model (GMM) to detect features of interest in the region. Therefore, we utilize a nonparametric probabilistic clustering method to estimate the model-form uncertainty associated with analyzing each data source individually. We introduce a novel mathematical framework that combines the estimated uncertainties and determines each dataset's value of information for understanding the region. We run the uncertainty analysis and visualize its results to assess the performance of the nonparametric clustering method compared to the GMM. Our methods are implemented on simulated data, as well as real optical and LiDAR imagery.