Peter Flo Grinde-Hollevik L&S Social Sciences
Applying Satellite-Based Geospatial Embeddings for Predictive Tasks
The advent of satellite-based geospatial embedding data has allowed for a new opportunity to explore an unforeseen set of capabilities in environmental predictive tasks. In this project, I first establish the predictive relationship between census block-level in-situ measurements of median NO2 concentrations derived from the Google Street View Car and geospatial embeddings sourced from the Sentinel-2 satellite and the Earth Genome Project.
This hypothesized relationship is then extended into the construction of a generalizable and robust predictive modeling framework. I then hope to deploy this framework on 4 counties of the Bay Area, seeking to accurately predict median concentrations of the set of census block-level NO2, NO, Black Carbon (BC), and Ultra Fine Particle Count (UFP) concentrations.
With the goal of minimizing the final model’s prediction error, this paper also suggests a replicable framework for carefully diagnosing spatially distributed model-induced bias.