Trillion Pixel 2021

April 21-22, 2021

2nd SERIES ON THE WAY FORWARD IN SCALING GeoAI

UPDATE – 2021 Workshop report is available! 

The increasing availability of geospatial data has converged with advancements in artificial intelligence, cloud infrastructure, and high-performance computing to enable mapping and analysis of the earth’s surface in unprecedented detail. Rapid innovations in sensing technologies will soon collect this data in even higher resolution and throughput. These developments offer the potential for breakthroughs in science, policy, and national security via end-to-end GeoAI systems that can provide fresh insights into how humans occupy and alter their environment over time.

This workshop seeks to imagine and shape how the scientific community can approach this challenge without losing sight of likely societal questions and impacts. Join us at this virtual event for an interdisciplinary gathering of experts from the fields of image science, computer vision, high-performance computing, architecture, machine learning, advanced workflows, and societal AI challenges to discuss the Trillion-Pixel GeoAI Challenge.

Day 1: April 21, 2021

#colspan##colspan#

Welcome

Dr. Budhu Bhaduri, Director, Geospatial Science and Human Security Division,  Oak Ridge National Laboratory
11:00 am - 11:05 am EDT 

#colspan##colspan#


Introductory Remarks

Dr. Deborah Frincke, Associate Laboratory Director, National Security Sciences, Oak Ridge National Laboratory
11:05am - 11:20 am EDT 

#colspan##colspan#

Session 1- Trillion Pixels - Grand Challenges

Moderators - Dr. Budhu Bhaduri, Oak Ridge National Laboratory; Ms. Jordan Lieberman, National Geospatial-Intelligence Agency
11:20 am - 12:45 pm EDT

#colspan##colspan#
With a growing number of global challenges, it is imperative to understand and fully explore the societal impacts of AI in the context of geographic knowledge. Coupling AI with geospatial data to address global challenges has had early successful use cases, though the envisioned societal impacts are yet to be fully appreciated. The grand challenges will require engaging the frontlines and visionary societal perspectives for guidance, as well as the true understanding of how GeoAI can play an essential role in addressing the greatest challenges. Join us in this session for a forward framing of the GeoAI initiative as a bridge toward uncovering unlimited possibilities for impacting global sustainable development goals and challenges for society’s benefit.#colspan##colspan#

Key Questions

#colspan#
  • Revisiting Global challenges with GeoAI: What are the emerging GeoAI successes and Impacts, can we do more? 
  • What are the gaps, limitations of GeoAI addressing end user application challenges in 2021? 
  • The next GeoAI : Toward Interdisciplinary GeoAI. Given these emerging successes and recognition of those limitations, what do we expect the next GeoAI would look like for solving end use grand challenges, engaging with end users across disciplinary and developing certain standards in data privacy, declaring GeoAI assurance? 

Panelists

  • Mr. Robert Shields,  Technical Executive,  Source Operations and Management, NGA
  • Dr. Nadine Alameh, CEO, Open Geospatial Consortium (OGC)
  • Dr. Jack Cooper, Program Manager, Space-based Machine Automated Recognition Technique (SMART), IARPA
  • Dr. Gerald Geernaert, Director, Climate and Environmental Sciences Division, DOE
  • Ms. Sylvia Wilson, Project Manager, National Land Imaging Program, USGS
  • Dr. Manil Maskey, Senior Research Scientist, Earth Science Data Systems, NASA

Session 2 - Generalization and Transferability

Moderators - Professor Hannah Kerner, University of Maryland and Dr. Timothy Doster, Pacific Northwest National Laboratory
1:00 pm - 2:15 pm EDT

#colspan##colspan#
The volume, velocity and variety of geospatial data are constantly growing at an unprecedented pace. Generalizable and transferable models will be critical to enable transformational and disruptive machine learning capabilities for AI-driven monitoring of the entire planet, every day, with unprecedented clarity and fidelity. In this session, we will discuss challenges and opportunities related to generalization and transferability of AI methods motivated by geospatial applications. Topics in this session will include generalization across space and time, domain adaptation, transfer learning, few-shot learning, model insights specific to geospatial data, and more.#colspan##colspan#

Key Questions

#colspan#

  • More labeled training data is not the only answer to improve generalization across space and time. Are we there yet on the challenge of collecting high quality training data? Unlabeled geospatial data is in abundance. How do we leverage it to address the generalization and scalability needs? How do we learn from both labeled and unlabeled data?

  • What are current promising directions to promote model generalization? How should we approach model transferability of multi-modal GeoAI data? 

  • New tasks and new datasets are emerging in GeoAI. Fine-tuning with typical pre-trained models has moved from generic benchmark datasets like ImageNet to tailored remote sensing benchmark datasets such as BigEarthNet/SpaceNet/xView. However these benchmark datasets may not be sufficient for the broad range of GeoAI applications and challenges. How do we train generalizable models and transfer between tasks/domains beyond fine-tuning? Are there lessons learned in scaling and generalizing models from these dedicated efforts on remote sensing datasets and competitions? (synthetic data, domain-specific benchmark datasets, domain transfer, few-shot learning, meta-learning, etc)

Panelists

#colspan#
  • Professor Xiao Xiang Zhu, Data Science in Earth Observation, Technischen Universität München
  • Mr. Andre Coleman, Senior Research Scientist, Pacific Northwest National Laboratory
  • Mr. Jesse Lew, National Geospatial-Intelligence Agency
  • Dr. Jacob Hinkle, Research Scientist, Oak Ridge National Laboratory
  • Mr. Christopher Brown, Senior Software Engineer, Google 

Session 3 - Scalable Geospatial Processing Architectures

Moderators - Dr. May Casterline, NVIDIA and Professor Eric Shook, University of Minnesota
2:30 pm - 3:45 pm EDT

#colspan##colspan#
The majority of “big data” analytical systems are designed to operate on light-weight data types that scale by quantity of record, not density of record. The complexities encountered with geospatial imagery does not inherently fit this model and the architecture to feed any type of processing engine becomes challenging at large scales. Join us in this session for a discussion about current HW/SW processing architectures for these data loads and where these solutions are meeting the needs or not covering the gaps.#colspan##colspan#

Key Questions

#colspan#
  • How would you define “scale” in this context?  Is it the size of the collection? The average size of queries?
  • Given what architectures are available today for large scale processing, are they meeting the demands of today?  Are they well designed to manage the demands of tomorrow?
  • What are the gaps that current architectures are not addressing?

Panelists

  • Dr. Hendrik F. Hamann, Chief Scientist for Geoinformatics and Solutions and Distinguished Researcher, IBM
  • Mr. Simeon Fitch, Astraea, VP of R&D at Astraea
  • Dr. Mallikarjun (Arjun) Shankar, Section Head and CADES Director, Oak Ridge National Laboratory
  • Professor Shashi Shekhar, McKnight Distinguished Professor, University of Minnesota 
  • Dr. Arvind Sujeeth, Senior Director, Sambanova

Day 2, April 22, 2021

#colspan##colspan#

Session 4 - Trustworthiness in GeoAI Systems

Moderators - Dr. Dalton Lunga and Dr. Edmon Begoli, Oak Ridge National Laboratory
11:30 am - 12:45 pm EDT

#colspan##colspan#
GeoAI systems, including techniques and tools based on machine learning and deep learning, are emerging as fundamental and integral for advancing breakthroughs in science, policy, and national security. However, the current progress of GeoAI systems is lagging in key areas that are limiting its effectiveness in a variety of societal challenges. This includes the key decision making challenges for high-consequence international agendas and safety critical missions where trust in GeoAI is critical and yet to be established. Join us in this session for a discussion on interrelated challenges and trade-offs in designing responsible and accountable GeoAI systems. We will cover trustworthiness attributes ranging from understanding the vulnerabilities of infrastructure, data, models, to interpretability, to explainability, to transparency, to robustness, to geoprivacy and data bias, and to responsible GeoAI designs.#colspan##colspan#

Key Questions

  • What should a GeoAI Trust framework look like – is it application dependent? What are the suitable requirements and components to such a framework? If data, model, infrastructure are key, is there an integral approach to infuse trust toward responsible systems?
  • Where are we on governance and ethics in AI for Earth observation?
  • For Trillion pixel challenges what tools should system designers, regulators, policymakers and other end user design to handle trust related challenges?
  • What is our reference to trustworthiness in GeoAI? Who is considered a trustworthy source given EO data gets mirrored all over the place?
  • How to explain real-world shifts and explain spatial and temporal consistency of GeoAI models under shifts?

Panelists

  • Dr. Benjamin Tuttle,  Chief Technology Officer, Arturo.ai
  • Ms. Vandy Tombs, Applied Mathematician, Oak Ridge National Laboratory
  • Professor John Monaco, Naval Postgraduate School
  • Mr. Rob Jasper, Manager, Pacific Northwest National Laboratory
  • Professor Claudio Persello, University of Twente

Session 5 - GeoAI Beyond Pixels

Moderators - Dr Rahul Ramachandran, National Aeronautical and Space Agency and Dr. Jitendra Kumar, Oak Ridge National Laboratory
1:00 pm - 2:15 pm EDT

#colspan##colspan#
Environmental observations collected in fields and laboratories, across networks of environmental observatories, environmental sensors and Internet of Things (IoT) devices provide rich sources of information. These observations can be quantitative or qualitative measurements, single snapshot or continuous time series, sparse and heterogeneous in nature. Integration of these unstructured non-geospatial data with increasingly available high resolution geospatial data offers opportunities to address important Earth and Environmental science problems. We will discuss GeoAI algorithms, computational methods and frameworks needed to go beyond pixels and integrate structured and unstructured data, and explore existing and potential applications.#colspan##colspan#

Key Questions

  • Is unstructured data an untapped resource for GeoAI? What are the types of unstructured data that could augment GeoAI? What are some existing applications? What are potential applications?
  • Where is the current state of the art? What approaches have been used to extract information and knowledge from unstructured data? How have they been combined with traditional data sources?
  • What are the best practices for collecting high quality unstructured data? What are the challenges in using unstructured data in GeoAI? What is the next frontier in this area? 
  • What are the early successes, challenges of GeoAI for multi-modal/cross-modal Earth observation analysis?

Panelists

#colspan#
  • Professor Laura Duncanson, University of Maryland College Park
  • Dr. Xiaoyuan Yang, Principal Applied Science Manager, Microsoft
  • Dr. Forrest Hoffman, Distinguished Scientist, Oak Ridge National Laboratory
  • Dr. Ryan Keisler, Chief Scientist, Descartes Labs

Session 6 - Collaborations and Community Engagements

Moderators - Professor Shawn Newsam, University of California Merced and Dr. Fabio Pacifici, Maxar
2:30 pm - 4:00 pm EDT

#colspan##colspan#
On January 15, 2021, then-President-Elect Biden sent a letter to Dr. Eric S. Lander, his appointee as the President’s Science Advisor and nominee as Director of the Office of Science and Technology Policy, tasking him to refresh and reinvigorate our national science and technology strategy[1].  (This letter is reminiscent of one sent by President Franklin D. Roosevelt to his science advisor after the Second World War in 1944.)  The letter poses five questions with the first two being:
  • What can we learn from the pandemic about what is possible—or what ought to be possible—to address the widest range of needs related to our public health?
  • How can breakthroughs in science and technology create powerful new solutions to address climate change—propelling market-driven change, jumpstarting economic growth, improving health, and growing jobs, especially in communities that have been left behind?
Collaborations and community engagements will be key to GeoAI playing a role in addressing these and other societal grand challenges.		#colspan##colspan#

Key Questions

#colspan#
  • What kinds of collaborations and community engagements should be fostered so that GeoAI can maximize its relevance to society particularly in the context of grand challenges?
  • These grand challenges are global in scale and so point to engagements with international entities.  But the problems and their solutions vary spatially—what works in one country might not be best for another especially with respect to different levels of development.  How do we ensure that local context is taken into account in order to avoid a “one size fits all” solution?
  • Who is missing from the list of attendees of this and similar meetings?  Are there entities, especially non-technical ones, that GeoAI needs to engage with?  How do we engage these entities (other than inviting them to meetings)?
  • The questions in the President’s letter seek to set a strong course for the next 75 years (similar to the 1944 letter).  Indeed, the grand challenges will likely take decades to fully address and so educating the next generation is essential.  Focusing on our field, how can we better attract and train future GeoAI talent?  Is GeoAI as attractive to students as other application areas of AI?  If not, how can we make it?  Is there a lack of qualified talent?  What specific expertise is lacking? How do we ensure geographic and other dimensions of diversity?

Panelists

#colspan#
  • Professor Danielle Wood, Media Arts and Sciences, Massachusetts Institute of Technology 
  • Dr. Giuseppe Borghi,  Head of the Φ-lab Division, European Space Agency
  • Dr. Ashley (Holt) Antonides, Chief AI Officer and Head of Data Science, Anno.Ai
  • Mr. Joe Flasher, Open Data Lead, Amazon Web Services (AWS)

Closing Remarks

Dr Lexie Yang, Dr Steven Ward, Oak Ridge National Laboratory
4:00 pm - 4:10 pm EDT

#colspan##colspan#