III: Small: Spatial Deep Learning from Imperfect Volunteered Geographic Information

This project aims to investigate novel spatial machine learning algorithms based on imperfect volunteered geographical information as ground truth for applications at the intersection of machine learning and geographic information science. The rapid growth of geospatial and spatiotemporal data being collected from space, airborne, and terrestrial platforms provides scientists, farmers, and first responders critical information they need about the surface of the Earth.

This emerging area that intersects machine learning, especially deep learning, with geographic information science is called GeoAI. GeoAI can potentially transform society by addressing grand challenges such as rapid disaster response, water resource management, and transportation. One major obstacle, however, is that deep learning heavily relies on a large number of training labels, which are often not easily available for geographic applications due to slow and expensive field surveys.

Existing research on semi-supervised learning could not fully resolve the issues due to the complex nature of geographic data such as spatial heterogeneity. This project will fill the gap by exploiting large scale, low-cost, and near real-time volunteered geographic information. The project will contribute towards the next generation water resource management for the U.S. in the 21st century.

This research can not only improve the situational awareness for disaster response agencies but also enhance the flood forecasting capabilities of the National Water Model. Planned algorithms will be implemented into open source tools that will enhance the research infrastructure for disaster management and hydrology communities. Educational activities include curriculum development, K-12 computer science education at Alabama Computer Science Summer Camps.

The principal investigator has a past record in mentoring undergraduate students from a historically black university and will continue the efforts at the University of Alabama, which has a reputation for producing African American researchers.

The planned framework will bring about several innovations to address significant technical challenges due to data quality issues. First, to address the noisy, biased, and incomplete label semantics, the project will develop novel label enhancement and enrichment algorithms based on physics-aware spatial structural constraints, which advance existing methods (that often assume independent labels) by jointly enhancing labels based on structural dependency.

Second, the project will explore a new location error model that better captures geometric shapes than existing square patch-based models, and design efficient joint learning algorithms to update deep model parameters while inferring true shape locations. Finally, to address location ambiguity, the project will explore location ambiguity models and mitigating location ambiguity by leveraging geographical contexts from input imagery features as well as spatial hierarchical constraints.

Such ideas advance existing location disambiguation methods merely based on the textual semantic context in natural language processing. The project can potentially transform the field of geospatial data science by addressing a major obstacle of limited training labels by a systematic framework of exploiting large-scale, low-cost, and near real-time volunteered geographic information data.

The project can potentially make transformative impacts on interdisciplinary GeoAI applications such as rapid disaster response and national water forecasting. This project is jointly funded by III and the Established Program to Stimulate Competitive Research (EPSCoR).

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.