Comparison of remote sensing, PSPNet, and green screen scores in place pulse dataset
Author: Andrew Larkin
Affiliation: Oregon State University, College of Public Health and Human Sciences
Principal Investigator: Perry Hystad
Summary
This is a comparison of different methodologies for measuring environmental composition and its relationship to participant perceptions in the MIT Place Pulse 2.0 dataset. Measurement methodologies include:
- Remote sensing measurements - examples include NDVI, impervious surface area, air pollution
- PSPNet labels - labels include trees, roads, buildings, and paths
- Green nature - percent green in an image directly attributable to nature
Repository Structure
Files are divided into two folders, with each folder corresponding to a unique stage of model development. Datasets are too large to upload to the github repository, but are available at (insert link here)
- data preprocessing - calculating built environment composition, downloading NDVI from the Google Earth Engine, and performing QA steps
- statistical analysis - calculating summary statistics, creating regression models
External Links
- MIT Place Pulse (archived) - https://www.media.mit.edu/projects/place-pulse-new/overview/
- PSPNet with ADEK20 Weights - https://arxiv.org/pdf/1608.05442.pdf
- Green Screen Repository - https://github.com/larkinandy/GSV_NDVI_Comparison