NASA EPSCoR Rapid Response Research: BDNN Hackathon: Uncertainty Aware Few Shot Learning from Citizen Science Data and Bayesian Deep Neural Network for Land Cover Image Classification

The goal of this proposal is to apply Machine Learning (ML) and Artificial Intelligence (AI) paradigms for quantifying uncertainty in the Citizen Science image data, and train a Bayesian Deep Neural Network (BDNN) with few shot learning for prediction of classes and labeling of the images in the GLOBE database. The GLOBE database of color (RGB) images are taken by people worldwide, with different camera settings, lighting, acquisition parameters leading to uncertainty in predicting land cover classes and thereby labeling the images. A robust method that can account for the uncertainty in the mixed pixels found in the edges and corners of neighboring classes in the image is necessary for accurate prediction of land cover classes and labeling of the Citizen Science database of images. The other goal of this project is to organize a year-long multievent hackathon for capability building in ML and AI in underrepresented minorities. The objectives for this NASA Rapid Response Project are:

1) Quantify uncertainty due to edge and boundary pixels in Citizen Science image data using uncertainty aware few shot learning method. 2) Develop a Bayesian Deep Neural Network (BDNN) with regularization of uncertainty for land cover classification of Citizen Science images. 3) Organize a year-long multi-event BDNN Hackathon for education and training in developing ML and AI approaches for land use classification of GLOBE images.

The edge and boundary data points will be identified by checking for optimality of the estimated class boundary known as Bayes boundary-ness, which will be quantized using Shannon entropy. We will implement self-trained few shot learning that selects instances from the unlabeled pool of these data points for uncertainty awareness. The uncertainty estimates will be added to the objective function of the BDNN which will be solved to output predictions of land cover classes and generating labels for the images. We will characterize images from homogeneous regions to more complex heterogeneous regions using region, shape and texture based descriptors, to improve land cover class predictions. The final labeling of the image will be done by majority voting. We will apply the BDNN tool for assessing the impact of hurricanes in the country. An interdisciplinary year-long multi-event virtual BDNN Hackathon will be organized and conducted by the PI and Co-I. Collaborators of the PI and CO-I from UPR and other educational institutions will contribute to the ML and AI training and learning modules. Dr. Roberto Rivera and Dr. Fernando Vega leading faculty in ML and AI in UPRM will be providing education modules for the hackathon. It is expected that atleast one hundred students from UPR campuses and high school students will participate in the hackathon. The hackathon will provide a hands-on learning experience for students with the tools developed in this project. Dr. Peder Nelson, Science lead for the land cover tool in GLOBE will be collaborating on this project. He will provide guidance, and feedback during the Land Cover Classification (LCC). The year-long virtual BDNN hackathon will be directed towards education and training of underrepresented minority students to gain expertise in learning ML and AI methods and applying them to NASA datasets. The hackathon will be organized in UPR-Mayaguez, a Hispanic minority serving institution with 99% Hispanic students who will gain knowledge in the fields of satellite remote sensing, airborne hyperspectral imaging and Python ML and AI tool development through the Hackathon. The BDNN hackathon will also enable Hispanic minority students to take up internships with NASA, and add to the workforce development for future NASA missions.

This is a mature technology (TRL 7+) — the realistic path in is usually NASA's Technology Transfer Program: licensing an existing NASA patent, or a Space Act Agreement to use NASA facilities/expertise directly. NASA also runs a startup licensing program with no upfront fee for companies formed to commercialize a specific NASA technology.

None of these are guaranteed paths for this specific project — TechPort itself doesn't have an "apply" button. Reaching out to the contact(s) above with a specific question is usually the fastest way to find out what's actually open.