Spatial AI Challenge 2024: Accepted Abstracts : I-GUIDE

We are pleased to announce that the following abstracts have been accepted for the Spatial AI Challenge 2024, hosted on the I-GUIDE Platform.

These submissions exemplify innovation in AI-ready spatial data, machine learning models, and applications, while promoting FAIR data principles and open science practices.

Perceiving Multidimensional Disaster Damages from Street-View Images Using Visual-Language Models

Team Name: Rayford

Team Members:

Lei Zou, Texas A&M University
Yifan Yang, Texas A&M University

Summary: This project explores the application of Large Language Models (LLMs) like GPT-4 Mini for disaster perception using street-view imagery. The study evaluates LLMs' ability to provide structured scoring, detailed natural language descriptions, and comparative analysis of disaster impacts. While LLMs enhance interpretability, the integration of traditional image-based classification methods with LLM-generated textual descriptions addresses challenges like information loss and accuracy limitations. To optimize performance, we propose a hybrid framework that combines high-dimensional image data with LLM-generated text for model training, improving both classification accuracy and result visualization. Additionally, a benchmark framework was developed to systematically evaluate pre-trained vision models, text-based models, and LLMs, providing actionable insights for automating disaster perception tasks. This multimodal approach aims to advance disaster management by balancing efficiency and interpretability in impact assessments and decision-making.

A Vision System for Geospatial Data Extraction

Team Name: Purdue RCAC

Team Members:

Jungha Woo, Purdue Research Computing
Elham Jebalbarezi Sarbijan, Purdue Research Computing

Summary: This project aims to develop a vision system capable of understanding map images, extracting bounding box coordinates, and generating captions that describe map content. By fine-tuning models like CLIP and image retrieval techniques, the system will interpret legends, scales, symbols, and coordinates in geospatial maps, facilitating automated extraction of spatial information. This advancement will enhance the capabilities of large language models in geospatial reasoning, aiding researchers in efficiently querying and utilizing legacy datasets.

Creation of a Machine-Learning-Ready High-Resolution RGB+NIR Orthoimagery Dataset with LiDAR-Derived "Ground Truth" Reference Products

Team Name: Individual Submission

Team Members:

Uku-Kaspar Uustalu, Tufts University

Summary: This project addresses the limited use of open-access satellite imagery in deep learning and computer vision due to the lack of reliable "ground truth" reference data and the geospatial expertise required for preprocessing. Estonia’s Estonian Land Board (ELB) provides extensive geospatial data through aerial LiDAR scanning and high-resolution RGB+NIR orthoimages, captured twice every four years. By leveraging this data, the project aims to create a machine-learning-ready dataset that aligns orthoimage tiles with LiDAR-derived reference products (e.g., DSMs, DTMs, building footprints). This dataset will be packaged into a Python tool for seamless integration with deep learning frameworks, lowering the barrier for researchers and practitioners working with aerial imagery.

Extraction of Surface Water Features from High-Resolution Orthoimagery and Elevation Data in Alaskan Watersheds using Novel Computer Vision Techniques

Team Name: Individual Submission

Team Members:

Uku-Kaspar Uustalu, Tufts University

Summary: This project seeks to apply advanced computer vision techniques—such as attention mechanisms, transformer-based models, residual networks, and graph neural networks (GNNs)—to predict hydrographic features like flowlines and surface water boundaries using elevation data and high-resolution orthoimagery. Building on the success of U-shaped convolutional neural networks (CNNs) in similar tasks, the project will explore modifying state-of-the-art segmentation models like SAM (Segment Anything Model) and YOLO (You Only Look Once) to improve surface water extraction while leveraging pretrained models to reduce computational costs. Additionally, traditional edge and contour detection methods (e.g., Canny Edge Detector, Hough Transform) will be investigated for refining vectorized predictions. The project will emphasize reproducibility by documenting the full workflow—from data acquisition to model validation—in a single executable notebook.

Identify Critical Transportation Infrastructure at Risk during Wildfires

Team Name: Team CNA

Team Members:

Steven Habicht, CNA
Matthew Prebble, CNA
Lars Hanson, CNA
Rebekah Yang, CNA
Shaun Williams, CNA
Jeremiah Huggins, CNA
Angie De Groot, CNA

Summary: This study focuses on developing a dataset to identify and assess risks to road and rail infrastructure in California regions susceptible to wildfires. By integrating high-resolution satellite imagery and authoritative data sources, such as CAL FIRE, the team will use segmentation models, including U-Net, to detect and classify transportation features and track evolving wildfire risk. The goal is to produce an AI-ready dataset and map visualizing transportation infrastructure at risk from wildfire disruption, providing a reproducible process documented in Jupyter Notebooks.

Integrating Semantic and Geospatial Contexts: Using Geography-Aware Language Models to Improve Urban Livability

Team Name: Citywalker

Team Members:

Zhaonan Wang, New York University Shanghai
Jing Tang, University of Zurich

Summary: This project focuses on improving urban livability recommendation systems by addressing the limitations of traditional methods, which struggle with personalization and geographical coherence. Current approaches often fail to align user preferences with spatial constraints, resulting in impractical and scattered travel routes. To overcome this, the proposed solution integrates a language model that combines semantic embeddings (capturing user preferences and POI features from text) with position embeddings (encoding spatial relationships). A benchmark dataset of approximately 1 million POIs and 10 million reviews is developed to train the model, ensuring recommendations that are both personalized and geographically coherent. This approach bridges the gap between semantic understanding and spatial reasoning, enabling greater urban livability.

Scalable Roof Material Classification to Inform Disaster Resilience Using OpenStreetMap Data

Team Name: Roof Mappers

Team Members:

Julian Huang, University of Chicago
Alex Monito Nhancololo, University of São Paulo
Yue Lin, University of Chicago

Summary: This project aims to develop a scalable pipeline for roof material classification in Nepal using OpenStreetMap (OSM) roof material labels. The study leverages OSM building outlines to bypass computationally expensive segmentation techniques. A convolutional neural network (CNN) model will be trained to classify roofs into categories such as concrete, reinforced cement concrete (RCC), and tin/metal, with scalability tests in India and Mozambique. The resulting high-resolution roof material maps will support disaster resilience planning and risk assessment models for natural disasters. The project will also explore how roof material classification can serve as a proxy for housing quality, income levels, and climate vulnerability, integrating this information into urban heat mitigation, storm risk assessment, and economic development planning.

The Geography of Human Flourishing

Team Name: Harvard University

Team Members:

Stefano Iacus, Institute for Quantitative Social Sciences (IQSS), Harvard University
Andrea Pio Nasuto, Institute for Quantitative Social Sciences (IQSS), Harvard University
Devika Jain, Center for Geographic Analysis (CGA), Harvard University

Summary: The Human Flourishing Program is a research initiative whose goal is to study and promote human flourishing across a broad spectrum of life domains, integrating interdisciplinary research in social sciences, philosophy, psychology, and other fields. The Global Flourishing Study (GFS), a five-year traditional longitudinal data collection on approximately 200,000 participants from 20+ geographically and culturally diverse countries and territories, measures global human flourishing in six areas: Happiness and life satisfaction; Mental and physical health; Meaning and purpose; Character and virtue; Close social relationships and Material and financial stability. Our research plan is to analyze Harvard’s collection of 10 billion geolocated tweets from 2010 to mid-2023. The project will apply large language models, to extract 46 human flourishing dimensions across the six areas of human flourishing, generate high-resolution spatio-temporal indicators and produce interactive tools to visualize and analyze the result.

TorchSpatial: A Location Encoding Framework and Benchmark for Spatial Representation Learning

Team Name: The Spatially Explicit Artificial Intelligence Lab (SEAI Lab)

Members:

Gengchen Mai, University of Texas at Austin
Jielu Zhang, University of Georgia
Nemin Wu, University of Georgia
Ni Lao, Google LLC
Qian Cao, University of Georgia
Yanlin Qi, University of California Davis
Zhangyu Wang, University of California Santa Barbara
Zeping Liu, University of Texas at Austin

Summary: We have been developing a Python Package called TorchSpatial, a comprehensive framework and benchmark suite designed to advance spatial representation learning (SRL). It will include a unified location encoding framework that supports location encoding model development and the LocBench benchmark tasks that support location encoding model evaluation. Additionally, TorchSpatial introduces the Geo-Bias Score, a novel metric to evaluate model performance and geographic bias, promoting spatial fairness in GeoAI applications.

We congratulate all the authors on their successful submissions and look forward to their contributions to advancing spatial AI and open science. Participants are reminded of the upcoming deadlines:

Preliminary Submission Deadline: March 31, 2025
Final Submission Deadline: April 15, 2025
Announcement of Winners: May 15, 2025

For more information and updates, please visit the Spatial AI Challenge 2024 page.