Virtually every environmental and social challenge being faced globally can be characterized by digital geospatial data. Artificial intelligence (AI) is increasingly embedded into the workflows for managing and processing those data, but much more is possible. By enabling machines to interpret and process the spatial relationships captured by those data sets, especially as they exist in the physical 3D world, spatial AI adds tremendous capacity to make breakthroughs. At first glance the advances may seem minor, such as a tweak to an algorithm that allows features to be identified from a satellite image slightly more accurately, or a new approach to merging data sets of varying spatial resolutions. However, the power of spatial AI is the ability for these small changes to potentially have a massive impact at scale. And yet, the tweaks and new approaches are being released so rapidly that it’s dizzying.
This year, nine teams completed I-GUIDE’s second Spatial AI Challenge, leveraging the I-GUIDE Platform to develop innovative, responsible, and reproducible solutions guided by FAIR data and open science principles. Challenge participants rely on the I-GUIDE Platform, a pioneering AI-powered cyberinfrastructure environment designed to advance geospatial data-intensive convergence science and knowledge discovery across various domains. All teams are required to produce at least one Jupyter Notebook that documents and supports any workflows in a way that makes the research inquiry as widely distributable and reproducible as possible. The Platform itself provides straightforward access to high-performance computing capabilities, and when necessary I-GUIDE facilitated allocations of GPUs, through NSF’s ACCESS program.
Enhancing computer vision was a key goal of several teams this year, including the group whose project won first place in this year’s Challenge. Led by Master’s student Raj Bhattarai and Dr. Fangzheng Lyu (both in the Department of Geography, Virginia Tech University), MURD-ViT: Urban Retrofitting Detection with Vision Transformer, created a deep learning pipeline for detecting urban retrofitting using temporal street view images and demographic data. Urban “retrofitting” means making localized improvements in pre-existing infrastructures that result in positive changes to functionality and sustainability. Bhattarai and Lyu trained a Vision Transformer (ViT) based model that integrates features from before/after temporal images of different locations with demographic data to readily and accurately classify urban retrofitting into useful categories.
Along with computer vision, training models to accommodate factors of uncertainty is another key area for Spatial AI. The 2nd place winner in this year’s Challenge was GeoLocate: Spatial Modeling of Market Entry Variability, led by a team based at the Federal University of Rio Grande do Sul (UFRGS) in Brazil, in collaboration with Harvard University’s Center for Geographic Analysis (CGA). Doctoral student Jaiany Rocha Trindade (Marketing, UFRGS; Visiting Fellow at CGA) and her advisor, Dr. Vinicius Andrade Brei (Marketing, UFRGS; former visiting Fellow at CGA), know that “location, location, location” is the mantra for successful real estate decisions. Working with Devika Jain (CGA), they explored the power of spatial AI. Their project focused on the finer-scale, nuanced distinctions of location differences, or spatial heterogeneity, represented by variations in infrastructure, accessibility, demographics, and economic activity that together influence real market opportunities. By integrating heterogeneous geospatial data within a Bayesian spatial model to estimate the probability of business survival in two large Brazilian cities, they are able to single out market opportunities that are more reliable and less uncertain.
Another collaborative effort led by UFRGS, From Mobility Intensity to Market Infrastructure: A Spatial AI Framework for Recovering Hidden Spatial Regimes and Multiscale Variation, was selected as the 3rd place winner in this year’s Spatial AI Challenge. Doctoral students Rafael Albuquerque (Marketing, UFRGS) and Jessica Miranda (Marketing, UFRGS), along with Dr. Vinicius Brei (Marketing, UFRGS), and Dr. Sisi Wang (Institute of Spatial Sciences, University of Southern California), focused their spatial AI efforts on the misleading characteristics of traffic in its role for economic planning. As they describe it, “Traffic can make weak locations look strong and durable locations look ordinary.” Instead of considering only a single, overly simplistic attribute of the volume of traffic, this group converted traffic to a “structured behavioral signal.” Combining diverse data such as mobility infrastructure, nighttime light intensity, and Census data, their efforts enriched the spatial structure of traffic volume alone as an explanatory variable.
Two other teams contributed projects that received Honorable Mention from the Challenge evaluation committee, and both of these thoughtfully integrated physics knowledge into AI models. Physics-informed Geo-AI for Irrigation Quantification, led by doctoral graduate students Esmaeel Adrah (Geography, Kent State University) and Daniel Dominguez (Computer Science, Colorado State University), made its advances by innovating the hybridization of satellite foundation model embeddings with mechanistic soil–water balance equations. Doctoral student Jibin Joseph and Dr. Venkatesh Merwade, of Purdue University’s Lyle School of Civil Engineering, used deep learning and remote sensing to model the relationships between hydrologic forcing, terrain controls, and observed flood extent to produce high-resolution flood inundation maps in their project Probabilistic Flood Inundation Mapping using Physics-Aware Spatial AI.
There is an important reason why a Spatial AI activity like this is called a Challenge: generating real results from real data, in a manner that is computationally reproducible, scalable and sharable with the rest of the world to subsequently build on these advances, is no trivial feat. I-GUIDE congratulates all teams who participated in this year’s Challenge. The Jupyter Notebooks and datasets for all of the Challenge’s projects can be found on I-GUIDE’s Platform, the enabling digital infrastructure that is supporting today’s and tomorrow’s most promising Spatial AI research. Keep watch on I-GUIDE for announcements for any future Challenges.