It is not always greener on the other side: Greenery perception across demographics and personalities in multiple cities

This repository is the official implementation of It is not always greener on the other side: Greenery perception across demographics and personalities in multiple cities.

Requirements

To install requirements:

python install -r requirements.txt

Dataset

This work uses the SPECS dataset (labels and SVI) and the NUS Global Streetscapes dataset (contextual metadata). Remaining files (depth-related metrics, prediction model, segmentations) are hosted in huggingface.

data
|- depth (depth computations of only greenery mask for all svi, download from huggingface)
|   |- met_depth (absolute depth .npy for each svi)
|   |- rel_depth (relative depth .npy for each svi)
|   |- met_depth_statistics_combined.csv
|   |- rel_depth_statistics_combined.csv
|- global-streetscapes (download contextual data from original repo)
|- labels (download files from original repo)
|   |- final (SPECS survey anwswers)
|   |- processed (trueskill and qscores computed values)
|- segmented_panoptic (segmented and blended outcome for all svi, download from huggingface)
|- svi (raw images within each country's folder and an 'all' folder, download files from original repo)
|   |- metadata.csv (uuid and survey image number mapping for all svi)
|   |- img_paths.csv (path within this repo for all svi)

The notebook 0a_gvi-calculations.ipynb peforms the semantic segmentation for greenery and computes the GVI values. Then, the notebook 0b_preprocessing.ipynb creates different datasets with Q Scores based on all available and different demographic and personality groups.

Methodology

Analyses

The notebook 0a_preprocessing.ipynb creates different datasets with Q Scores based on all available and different demographic and personality groups. Then, the notebook 0b_gvi-calculations.ipynb peforms the semantic segmentation for greenery and computes the GVI values.

Correlation and bias evaluation

The file 1_greenery-gvi-location.R runs a Pearson correlation analysis on images' green perception Q scores and GVI based on the images' and participants' location, i.e., location pairs.

The file 2_green-gvi-scatter-altman.R generates a Bland-Altman analysis and plot, based on the images' locations and then by location pairs. It also runs a Wilcoxon Signed-Rank between the median of score differenecs (Q score - GVI) and a median of 0.

The file 3a_greenery-beautiful-location.R runs a cubic correlation analysis on images' green and beautiful perception Q scores based on location pairs.

The file 3b_greenery-lively-location.R runs a cubic correlation analysis on images' green and lively perception Q scores based on location pairs.

Greenery distribution

Spatial entropy

The Jupyter notebook 4a_spatial_entropy_calculation.ipynb computes the spatial entropy for a range of window sizes and the window size is determnined with a sensitivty plot done by 4b_spatial_entropy_sensitivity.R.

The file 4c_spatial_entropy_analysis.R performs Mann-Whitney U statistical tests on spatial entropy values based on most and least perceived green images (Q1 and Q3 quantiles of Q score values).

Distance

The file 5a_depth_calculations.ipynb computes the mean depth distribution across images grouped by based on most and lest perceived green images (Q1 and Q3 quantiles of Q score values). Then, the file 5b_depth_analysis.R performs Mann-Whitney U statistical tests between the Q1 and Q3 Q scores distributions on images grouped on location pairs. Complementary, the file 5c_qscore_by_depth_analysis.R performs the same tests between the Q1 and Q3 of their vegetation proximity in meters (median of depth values).

Demographic differences

The file 6_demographic_diff.R highlights a comparison of different perception of the same urban scene by participants from different locations.

Perception modeling

The Jupyter notebook 7_perception-modeling.ipynb trains and evaluations a Random Forest model based on different participants' metadata (demographic and personalities) and image-based features, and performs a conditional feature permutation evaluation. The file 7_perception-modeling_plots.R generates the respective plots based on the permutation analysis.

Auxiliary files

The following files were used for complementary computations and analyses

• 9a_green_thresholding.ipynb produces green masks based on green pixel values and 9b_color-thresholding-vs-semantic-gvi.R compares the GVI values based on pixel-based green masks and semantic segmentation masks.
• 9c_trueskill_vs_qscore.R compares the Trueskill scores and Q scores for all images.
• 9d_shanon_entropy_analysis.R computes the Shanon entropy values for all images.

Paper / Attribution / Citation

If you use our dataset or work, please cite the paper:

BibTeX:

@article{TODO
}

Postprint

Besides the published paper, a free version (postprint / author-accepted manuscript) can be downloaded here.