CP8883: Major Assignment 2

Exploring Walkability Through Street View and Computer Vision

Section 0. Packages

library(tidycensus)
library(tidyverse)
library(magrittr)
library(osmdata)
library(sfnetworks)
library(units)
library(sf)
library(tidygraph)
library(tmap)
library(here)
library(progress)
ttm()

Section 1. Choosing Census Tracts.

Use the Census Tract map in the following code chunk to identify the GEOIDs of the tracts you consider walkable and unwalkable.

# TASK ////////////////////////////////////////////////////////////////////////
# Set up your api key here
census_api_key(Sys.getenv('CENSUS_API'))
# //TASK //////////////////////////////////////////////////////////////////////

# =========== NO MODIFICATION ZONE STARTS HERE ===============================
# Download Census Tract polygon for Fulton and DeKalb
tract <- get_acs("tract", 
                 variables = c('pop' = 'B01001_001'),
                 year = 2023,
                 state = "GA", 
                 county = c("Fulton", "DeKalb"), 
                 geometry = TRUE)

tmap_mode('view')
tm_basemap("OpenStreetMap") +
  tm_shape(tract) + 
  tm_polygons(fill_alpha = 0.2)
# =========== NO MODIFY ZONE ENDS HERE ========================================

Once you have the GEOIDs, create two Census Tract objects – one representing your most walkable area and the other your least walkable area.

# TASK ////////////////////////////////////////////////////////////////////////
# 1. Specify the GEOIDs of your walkable and unwalkable Census Tracts. 
#    e.g., tr_id_walkable <- c("13121001205", "13121001206")
# 2. Extract the selected Census Tracts using `tr_id_walkable` and `tr_id_unwalkable`

# For the walkable Census Tract(s)
tr_id_walkable <- c('13121001002')

tract_walkable <- tract %>% filter(GEOID %in% tr_id_walkable) 


# For the unwalkable Census Tract(s)
tr_id_unwalkable <- c('13121001902')

tract_unwalkable <- tract %>% filter(GEOID %in% tr_id_unwalkable) 


# //TASK //////////////////////////////////////////////////////////////////////


# TASK ////////////////////////////////////////////////////////////////////////
tmap_mode("view")

tm_shape(tract_walkable) +
  tm_polygons(col = "green", alpha = 0.6) +
tm_shape(tract_unwalkable) +
  tm_polygons(col = "red", alpha = 0.6) +
tm_layout(
  title = "Walkable & Unwalkable Census Tracts",
  legend.outside = TRUE
)

# //TASK //////////////////////////////////////////////////////////////////////

Provide a brief description of your selected Census Tracts. Why do you consider these tracts walkable or unwalkable? What factors do you think contribute to their walkability?

Section 2. OSM, GSV, and Computer Vision.

Step 1. Get and clean OSM data.

To obtain the OSM network for your selected Census Tracts: (1) Create bounding boxes. (2) Use the bounding boxes to download OSM data. (3) Convert the data into an sfnetwork object and clean it.

# TASK ////////////////////////////////////////////////////////////////////////
# Create one bounding box (`tract_walkable_bb`) for your walkable Census Tract(s) and another (`tract_unwalkable_bb`) for your unwalkable Census Tract(s).

# For the walkable Census Tract(s)
tract_walkable_bb <- tract %>% 
  filter(GEOID == tr_id_walkable) %>%
  st_bbox()

# For the unwalkable Census Tract(s)  
tract_unwalkable_bb <- tract %>% 
  filter(GEOID == tr_id_unwalkable) %>%
  st_bbox()

# //TASK //////////////////////////////////////////////////////////////////////


# =========== NO MODIFICATION ZONE STARTS HERE ===============================
# Get OSM data for the two bounding boxes
osm_walkable <- opq(bbox = tract_walkable_bb) %>%
  add_osm_feature(key = 'highway', 
                  value = c("primary", "secondary", "tertiary", "residential")) %>%
  osmdata_sf() %>% 
  osm_poly2line()

osm_unwalkable <- opq(bbox = tract_unwalkable_bb) %>%
  add_osm_feature(key = 'highway', 
                  value = c("primary", "secondary", "tertiary", "residential")) %>%
  osmdata_sf() %>% 
  osm_poly2line()
# =========== NO MODIFY ZONE ENDS HERE ========================================


# TASK ////////////////////////////////////////////////////////////////////////
# 1. Convert `osm_walkable` and `osm_unwalkable` into sfnetwork objects (as undirected networks),
# 2. Clean the network by (1) deleting parallel lines and loops, (2) creating missing nodes, and (3) removing pseudo nodes (make sure the `summarise_attributes` argument is set to 'first' when doing so).

net_walkable <- osm_walkable$osm_lines %>% 
  # Drop redundant columns 
  select(osm_id, highway) %>% 
  sfnetworks::as_sfnetwork(directed = FALSE) %>% 
  activate("edges") %>% 
  filter(!edge_is_multiple()) %>% 
  filter(!edge_is_loop()) %>%
  convert(., sfnetworks::to_spatial_subdivision) %>%
  convert(., sfnetworks::to_spatial_smooth, summarise_attributes = "first")

net_unwalkable <- osm_unwalkable$osm_lines %>% 
  # Drop redundant columns 
  select(osm_id, highway) %>% 
  sfnetworks::as_sfnetwork(directed = FALSE) %>% 
  activate("edges") %>% 
  filter(!edge_is_multiple()) %>% 
  filter(!edge_is_loop()) %>%
  convert(., sfnetworks::to_spatial_subdivision) %>%
  convert(., sfnetworks::to_spatial_smooth, summarise_attributes = "first")
  
# //TASK //////////////////////////////////////////////////////////////////////
  
  
# TASK //////////////////////////////////////////////////////////////////////
# Using `net_walkable` and`net_unwalkable`,
# 1. Activate the edge component of each network.
# 2. Create a `length` column.
# 3. Filter out short (<300 feet) segments.
# 4. Randomly Sample 100 rows per road type.
# 5. Assign the results to `edges_walkable` and `edges_unwalkable`, respectively.

# OSM for the walkable part
edges_walkable <- net_walkable %>% 
  st_as_sf("edges") %>% 
  select(osm_id, highway) %>% 
  mutate(length = st_length(.) %>% unclass()) %>%       
  filter(length >= 91) %>%  
  group_by(highway) %>% 
  slice_sample(n = 100) %>%                       
  ungroup() %>%
  mutate(edge_id = seq(1,nrow(.)))


# OSM for the unwalkable part
edges_unwalkable <- net_unwalkable %>% 
  st_as_sf("edges") %>% 
  select(osm_id, highway) %>% 
  mutate(length = st_length(.) %>% unclass()) %>%       
  filter(length >= 91) %>%  
  group_by(highway) %>% 
  slice_sample(n = 100) %>%                       
  ungroup() %>%
  mutate(edge_id = seq(1,nrow(.)))

# //TASK //////////////////////////////////////////////////////////////////////
  
# =========== NO MODIFICATION ZONE STARTS HERE ===============================
# Merge the two
edges <- bind_rows(edges_walkable %>% mutate(is_walkable = TRUE), 
                   edges_unwalkable %>% mutate(is_walkable = FALSE)) %>% 
  mutate(edge_id = seq(1,nrow(.)))
# =========== NO MODIFY ZONE ENDS HERE ========================================

Step 2. Define getAzimuth() function.

In this assignment, you will collect two GSV images per road segment, as illustrated in the figure below. To do this, you will define a function that extracts the coordinates of the midpoint and the azimuths in both directions.

If you can’t see this image, try changing the markdown editing mode from ‘Source’ to ‘Visual’ (you can find the buttons in the top-left corner of this source pane).

getAzimuth <- function(line){

  # TASK ////////////////////////////////////////////////////////////////////////
  # 1. Use the `st_line_sample()` function to sample three points at locations 0.48, 0.5, and 0.52 along the line. These points will be used to calculate the azimuth.
  # 2. Use `st_cast()` function to convert the 'MULTIPOINT' object into a 'POINT' object.
  # 3. Extract coordinates using `st_coordinates()`.
  # 4. Assign the coordinates of the midpoint to `mid_p`.
  # 5. Calculate the azimuths from the midpoint in both directions and save them as `mid_azi_1` and `mid_azi_2`, respectively.
  
  # 1-3
  mid_p3 <- line %>% st_line_sample(line, sample = c(0.48, 0.5, 0.52))  %>% 
    st_cast("POINT") %>%        
    st_coordinates() 
  
  # 4
  mid_p <- mid_p3[2, ]  
  
  # 5
  mid_azi_1 <- atan2(mid_p3[1,"X"] - mid_p3[2, "X"],
                   mid_p3[1,"Y"] - mid_p3[2, "Y"])*180/pi
  
  mid_azi_2 <- atan2(mid_p3[3,"X"] - mid_p3[2, "X"],
                   mid_p3[3,"Y"] - mid_p3[2, "Y"])*180/pi
  
  # //TASK //////////////////////////////////////////////////////////////////////
 
  
  # =========== NO MODIFICATION ZONE STARTS HERE ===============================
  return(tribble(
    ~type,    ~X,            ~Y,             ~azi,
    "mid1",    mid_p["X"],   mid_p["Y"],      mid_azi_1,
    "mid2",    mid_p["X"],   mid_p["Y"],      mid_azi_2,))
  # =========== NO MODIFY ZONE ENDS HERE ========================================
}

Step 3. Apply the function to all street segments

Apply the getAzimuth() function to the edges object. Once this step is complete, your data will be ready for downloading GSV images.

# TASK ////////////////////////////////////////////////////////////////////////
# Apply getAzimuth() function to all edges.
# Remember that you need to pass edges object to st_geometry() before you apply getAzimuth()
edges_azi <- edges %>% 
  st_geometry() %>% 
  map_df(getAzimuth, .progress = T)

# //TASK //////////////////////////////////////////////////////////////////////

# =========== NO MODIFICATION ZONE STARTS HERE ===============================
edges_azi <- edges_azi %>% 
  bind_cols(edges %>% 
              st_drop_geometry() %>% 
              slice(rep(1:nrow(edges),each=2))) %>% 
  st_as_sf(coords = c("X", "Y"), crs = 4326, remove=FALSE) %>% 
  mutate(img_id = seq(1, nrow(.)))
# =========== NO MODIFY ZONE ENDS HERE ========================================

Step 4. Define a function that formats request URL and download images.

getImage <- function(iterrow){
  # This function takes one row of `edges_azi` and downloads GSV image using the information from the row.
  
  # TASK ////////////////////////////////////////////////////////////////////////
  # 1. Extract required information from the row of `edges_azi`
  # 2. Format the full URL and store it in `request`. Refer to this page: https://developers.google.com/maps/documentation/streetview/request-streetview
  # 3. Format the full path (including the file name) of the image being downloaded and store it in `fpath`
  type <- iterrow$type
  location <- paste0(iterrow$Y %>% round(5), ",", iterrow$X %>% round(5))
  heading <- iterrow$azi %>% round(1)
  edge_id <- iterrow$edge_id
  img_id <- iterrow$img_id
  key <- Sys.getenv("google_api")
  
  endpoint <- 'https://maps.googleapis.com/maps/api/streetview'
  
  request <- glue::glue("{endpoint}?size=640x640&location={location}&heading={heading}&fov=90&pitch=0&key={key}")
  fname <- glue::glue("GSV-nid_{img_id}-eid_{edge_id}-type_{type}-Location_{location}-heading_{heading}.jpg") # Don't change this code for fname
  fpath <- file.path("C:/GT MS UA/CP 8883/Projects/Major Project 2", fname)
  # //TASK //////////////////////////////////////////////////////////////////////

  
  
  # =========== NO MODIFICATION ZONE STARTS HERE ===============================
  # Download images
  if (!file.exists(fpath)){
    download.file(request, fpath, mode = 'wb') 
  }
  # =========== NO MODIFY ZONE ENDS HERE ========================================
}

Step 5. Download GSV images

Before you download GSV images, make sure the row number in edges_azi is not too large! Each row corresponds to one GSV image, so if the row count exceeds your API quota, consider selecting different Census Tracts.

You do not want to run the following code chunk more than once, so the code chunk option eval=FALSE is set to prevent the API call from executing again when knitting the script.

# =========== NO MODIFICATION ZONE STARTS HERE ===============================
for (i in seq(1,nrow(edges_azi))){
  getImage(edges_azi[i,])
}
# =========== NO MODIFY ZONE ENDS HERE ========================================

ZIP THE DOWNLOADED IMAGES AND NAME IT ‘gsv_images.zip’ FOR STEP 6.

Step 6. Apply computer vision

Use this Google Colab script to apply the pretrained semantic segmentation model to your GSV images.

Step 7. Merging the processed data back to R

Once all of the images are processed and saved in your Colab session as a CSV file, download the CSV file and merge it back to edges_azi.

# TASK ////////////////////////////////////////////////////////////////////////
# Read the downloaded CSV file containing the semantic segmentation results.
seg_output <- read.csv('C:/GT MS UA/CP 8883/Projects/Major Project 2/seg_output.csv')

# //TASK ////////////////////////////////////////////////////////////////////////

# TASK ////////////////////////////////////////////////////////////////////////  
# 1. Join the `seg_output` data to `edges_azi`.
# 2. Calculate the proportion of predicted pixels for the following categories: `building`, `sky`, `road`, and `sidewalk`. If there are other categories you are interested in, feel free to include their proportions as well.
# 3. Calculate the proportion of greenness using the `vegetation` and `terrain` categories.
# 4. Calculate the building-to-street ratio. For the street, use `road` and `sidewalk` pixels; including `car` pixels is optional.

edges_seg_output <- edges_azi %>% 
  left_join(seg_output, by="img_id") %>%
  mutate(
    prop_building = building/(640*640),
    prop_sky      = sky/(640*640),
    prop_road     = road/(640*640),
    prop_sidewalk = sidewalk / (640*640),
    prop_greenness = (vegetation + terrain)/(640*640),
    street_pixels = road + sidewalk + car,
    building_to_street_ratio = ifelse(street_pixels > 0,
                                   building / street_pixels,
                                   NA))
  
# //TASK ////////////////////////////////////////////////////////////////////////

Section 3. Summarize and analyze the results.

At the beginning of this assignment, you specified walkable and unwalkable Census Tracts. The key focus of this section is the comparison between these two types of tracts.

Analysis 1 - Visualize Spatial Distribution

Create interactive maps showing the proportion of sidewalk, greenness, and the building-to-street ratio for both walkable and unwalkable areas. In total, you will produce 6 maps. Provide a brief description of your findings.

# TASK ////////////////////////////////////////////////////////////////////////
# Plot interactive map(s)
# As long as you can deliver the message clearly, you can use any format/package you want.

walkable_edges <- edges_seg_output %>% filter(is_walkable == TRUE)
unwalkable_edges <- edges_seg_output %>% filter(is_walkable == FALSE)

make_map <- function(tract_data, t_col, edges_data, value_col, palette_name, title_) {
  tm_shape(tract_data) +
  tm_borders(col = t_col, lwd = 2) +
  tm_shape(edges_data)+
  tm_dots(col = value_col,
      palette = palette_name,
      size = 0.6)+
  tm_layout(
  title = title_,
  legend.outside = TRUE)
}

tmap_mode("view")

w1<-make_map(tract_walkable, "darkgreen", walkable_edges, 'prop_sidewalk', "PuRd", "Walkable Census Tract: Proportion of Sidewalks")

w2<-make_map(tract_walkable, "darkgreen", walkable_edges, 'prop_greenness', "PuRd", "Walkable Census Tract: Proportion of Greenness")

w3<-make_map(tract_walkable, "darkgreen", walkable_edges, 'building_to_street_ratio', "PuRd", "Walkable Census Tract: Building-to-Street Ratio") 

uw1<-make_map(tract_unwalkable, "red", unwalkable_edges, 'prop_sidewalk', "PuRd", "Unwalkable Census Tract: Proportion of Sidewalks")

uw2<-make_map(tract_unwalkable, "red", unwalkable_edges, 'prop_greenness', "PuRd", "Unwalkable Census Tract: Proportion of Greenness")

uw3<-make_map(tract_unwalkable, "red", unwalkable_edges, 'building_to_street_ratio', "PuRd", "Unwalkable Census Tract: Building-to-Street Ratio") 

tmap_arrange(w1,w2,w3,uw1,uw2,uw3) 

# //TASK //////////////////////////////////////////////////////////////////////

Analysis 2 - Boxplot

Create boxplots for the proportion of each category (building, sky, road, sidewalk, greenness, and any additional categories of interest) and the building-to-street ratio for walkable and unwalkable tracts. Each plot should compare walkable and unwalkable tracts. In total, you will produce 6 or more boxplots. Provide a brief description of your findings.

# TASK ////////////////////////////////////////////////////////////////////////
# Create boxplot(s) using ggplot2 package.
vars_to_plot <- c(
  "prop_building",
  "prop_sky",
  "prop_road",
  "prop_sidewalk",
  "prop_greenness",
  "building_to_street_ratio"
)

edges_seg_output <- edges_seg_output %>%
  mutate(
    tract_type = ifelse(is_walkable == TRUE, "walkable", "unwalkable")
  )

plot_df <- edges_seg_output %>%
  select(tract_type, all_of(vars_to_plot)) %>%
  pivot_longer(
    cols = all_of(vars_to_plot),
    names_to = "variable",
    values_to = "value"
  )

ggplot(plot_df, aes(x = tract_type, y = value, fill = tract_type)) +
  geom_boxplot(alpha = 0.8) +
  facet_wrap(~ variable, scales = "free_y") +
  scale_fill_manual(values = c("red", "lightgreen")) +
  labs(
    x = "Tract Type",
    y = "Proportion / Ratio",
    fill = "Tract Type"
  ) +
  theme_minimal(base_size = 13)

# //TASK //////////////////////////////////////////////////////////////////////

Analysis 3 - Mean Comparison (t-test)

Perform t-tests on the mean proportion of each category (building, sky, road, sidewalk, greenness, and any additional categories of interest) as well as the building-to-street ratio between street segments in the walkable and unwalkable tracts. This will result in 6 or more t-test results. Provide a brief description of your findings.

# TASK ////////////////////////////////////////////////////////////////////////
# Perform t-tests and report both the differences in means and their statistical significance.
# As long as you can deliver the message clearly, you can use any format/package you want.
cols_to_test <- c(
  "prop_building",
  "prop_sky",
  "prop_road",
  "prop_sidewalk",
  "prop_greenness",
  "building_to_street_ratio"
)

test_func <- function(data, vars) {
  results <- lapply(vars, function(v) {
    w <- data %>% filter(tract_type == "walkable") %>% pull(v)
    u <- data %>% filter(tract_type == "unwalkable") %>% pull(v)
    
    t_out <- t.test(w, u)
    
    data.frame(
      variable = v,
      mean_walkable = mean(w, na.rm = TRUE),
      mean_unwalkable = mean(u, na.rm = TRUE),
      mean_diff = mean(w, na.rm = TRUE) - mean(u, na.rm = TRUE),
      p_value = t_out$p.value
    )
  })
  do.call(rbind, results)
}

# Run
t_test_results <- test_func(edges_seg_output, cols_to_test)

# View
print(t_test_results)

##                   variable mean_walkable mean_unwalkable   mean_diff
## 1            prop_building    0.11225617      0.28643016 -0.17417400
## 2                 prop_sky    0.34096407      0.20382502  0.13713905
## 3                prop_road    0.51311640      0.53326304 -0.02014664
## 4            prop_sidewalk    0.05776996      0.06520475 -0.00743479
## 5           prop_greenness    0.35135364      0.28870766  0.06264598
## 6 building_to_street_ratio    0.19081552      0.45973186 -0.26891634
##        p_value
## 1 6.362724e-20
## 2 2.803339e-17
## 3 2.407518e-03
## 4 6.903160e-02
## 5 2.141336e-03
## 6 3.056429e-18

# //TASK //////////////////////////////////////////////////////////////////////

Inferences

Analysing the boxplots shows that the largest differences appear in the building-to-street ratio and the proportion of building, both of which are considerably higher in unwalkable areas. These tracts likely reflect dense, auto-oriented development with busy streets and a lower sense of pedestrian comfort. Walkable areas, by contrast, show more visible sky suggesting lower building heights and higher presence of greenery, likely tied to consistent investment in street trees and vegetation.

The t-tests reinforce these visual patterns. Unwalkable tracts have significantly higher building proportions and building-to-street ratios, while walkable tracts display substantially more visible sky and greenness. Sidewalk proportions do not differ in a meaningful way, and road proportions vary only slightly despite statistical significance.

Together, these spatial, visual, and statistical insights show that perception of walkability emerges not from the mere presence of sidewalks or roads but from a coordinated urban form with human-scale building placement, continuous green infrastructure, well-proportioned streets, and coherent design patterns that support comfortable pedestrian movement.