RAPID: Collaborative Research: Machine Learning for Dehazing Unmanned Aerial System Imagery from Volcanic Eruptions

点击翻译按钮获取中文摘要

The ongoing eruption of the Kilauea volcano in Hawaii is the first reported time that small unmanned aerial systems (UAS) have been used for the emergency response to a volcanic eruption. The Center for Robot-Assisted Search and Rescue (CRASAR) flew 44 small UAS flights for the Hilo Fire Department and Hawaii County Civil Defense. The eruption imagery was partially occluded by plumes of steam carrying toxic gases, something that had not been encountered before. The plumes interfere with responders comprehending the tactical situation because it obscures the ground below and often prevents software from generating useful surface maps. While volcanic eruptions are fairly rare, the same plume problem is likely to occur in other hazardous material events. Machine learning techniques for dehazing were only partially successful because plumes present a very different set of challenges than removing urban haze or smog. This project conducts rapid research to remove or reduce plumes, from stills and video, in near real-time in order to support responses to the ongoing disaster. It will make the datasets available so that they can be used for training and evaluating new machine learning algorithms. The project will host a follow up workshop at the 2019 AAAI Conference on Artificial Intelligence in Hawaii.This project creates a UAS open-source imagery dataset from the ongoing Leilani, Hawaii, volcanic eruption event. It uses the dataset to expand and refine dehazing algorithms that will help Hawaii public safety agencies and volcanologists see through the plumes of steam and gas that is interfering with mapping the extent and volume of the lava. Plumes of steam mingled with sulfur dioxide interfered with interpreting the boundaries of the lava field and introduced errors into stitching images together or caused details to be averaged out. Smog is a homogeneous, thin visual phenomenon while plumes are heterogeneous and thick, limiting the utility of current techniques and requiring focused research. The dataset offers an opportunity for a corpus of real imagery that can serve as machine learning training data and enable comparison of before and after results. The intellectual merit of the project is twofold. It provides a unique opportunity to explore a new area of machine learning for heterogeneous, thick plumes in images. The comprehensive dataset will enable foundational work in computer vision, machine learning, and emergency informatics. The research will immediately improve emergency management of the Leilani eruption event and emergency management in general.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.