Title: Bayesian Inference Driven Active Learning with Unknown Class Discovery for Geospatial Image Analysis
Presenting Author: Saurabh Prasad
Organization: University of Houston

Abstract:
Incorporating disparate features from multiple sources can provide valuable diverse information for remote sensing data analysis. However, multi-source remote sensing data require large quantities of labeled data to train robust supervised classifiers, which are often difficult and expensive to acquire. Likewise, with rapid gains in hardware technologies, sensing modalities such as hyperspectral imagery and full-waveform LiDAR provide rich information about ground cover and topography, but result in very high dimensional feature spaces, posing challenges with regards to robust machine learning. Additionally, missing labels (e.g. unknown classes) are commonly encountered (and typically not accounted for) in remote sensing workflows --- this often contributes to sub-optimal performance with traditional machine learning algorithms. In this presentation, we will cover two synergistic research efforts that attempt to alleviate these problems with the design of new machine learning paradigms. Results presented here cover two NASA projects - an AIST project on multi-source active learning, and a NIP project on Bayesian inference approaches for geospatial image analysis. We will present a multi-view multi-kernel active learning approach that was developed to facilitate robust analyst-in-the-loop machine learning for multi-source remote sensing. We will also present our Bayesian inference approach to active learning that is capable of learning statistics of a geospatial scene effectively, providing unique capability to account for missing label information, and providing an effective uncertainty measure of the image analysis outcomes in single and multi-sensor scenarios. We will also discuss synergy between these efforts and our planned future work to connect these efforts, i.e, to develop an active Bayesian inference paradigm for multi sensor remote sensing.