oregon/03-cluster.md

In cwickham/routes: Generate data-based landscapes from driving routes

Count and cluster colors for route Corvallis, OR to Portland, OR

library(tidyverse)

## ── Attaching packages ────────────────────────────── tidyverse 1.2.1 ──

## ✔ ggplot2 3.2.0     ✔ purrr   0.3.2
## ✔ tibble  2.1.3     ✔ dplyr   0.8.3
## ✔ tidyr   0.8.3     ✔ stringr 1.4.0
## ✔ readr   1.3.1     ✔ forcats 0.4.0

## ── Conflicts ───────────────────────────────── tidyverse_conflicts() ──
## ✖ dplyr::filter() masks stats::filter()
## ✖ dplyr::lag()    masks stats::lag()

library(colorspace)
library(routes)

## Attaching routes version 0.0.2

library(here)

## here() starts at /Users/wickhamc/Documents/Projects/routes/routes

set.seed(21978) # both subsampling and clustering are random, 
                # set a seed to get reproducible image

Read in Street View image manifest

route_images <- read_rds(here("oregon", "images", "manifest.rds"))

Get sample of pixels from images

Sample 50 pixels from each image:

pixel_sample_rgb <- route_images %>% 
  pull(image) %>% 
  map_dfr(sample_pixels, sample_size = 50) %>% 
  with(RGB(R, G, B)) 

Create cluster centers

cluster_space <- "LUV"
n_clusters <- 50

Clustering done in LUV color space, with 50 clusters:

pixel_sample_space <- pixel_sample_rgb %>% 
  as(cluster_space)

clusters <- kmeans(pixel_sample_space@coords, centers = n_clusters)
centers <- match.fun(cluster_space)(clusters$center)

centers %>% hex() %>% pal()

Count pixels in each cluster

color_freq_sample <- count_colors(pixel_sample_rgb, centers,
  colorspace = cluster_space)

Histogram of counts based on sample:

color_freq_sample %>% 
  ggplot(aes(reorder(hex, freq), freq, fill = hex)) +
    geom_col() +
  scale_fill_identity() +
  coord_flip() +
  theme_classic()

Save cluster centers and frequencies in sampled pixels:

color_freq_sample %>% 
  write_rds(here("oregon", "data", "color_freq_sample.rds")) %>% 
  write_csv(here("oregon", "data", "color_freq_sample.csv"))

Count pixels in each image

route_pixels <- route_images %>% 
  mutate(
    pixels = map(image, read_pixels) %>% map(RGB),
    color_count = map(pixels, count_colors, 
      centers = centers, colorspace = "LUV")) 

Save some summaries so plots can be reproduced without images/:

route_pixels %>% 
  mutate(hex = map(pixels, hex)) %>% 
  select(-color_count, -pixels) %>% 
  unnest() %>% 
  group_by(lat, lon, order, image, hex) %>% 
  count() %>% 
  write_csv(here("oregon", "data", "route_pixels_raw_counts.csv.gz"))

route_pixels <- 
  route_pixels %>% 
  unnest(color_count) %>% 
  mutate(H = as(LUV(L, U, V), "polarLUV")@coords[, "H"])

route_pixels %>% 
  write_rds(here("oregon", "data", "route_pixels.rds"))

Initial Plot

route_pixels %>% 
  ggplot(aes(order, freq)) +
    geom_area(aes(fill = hex)) +
  scale_fill_identity() +
  scale_color_identity() +
  equal_margins()

orderby <- quo(V)

You might try re-ordering the colors by one of the color dimensions, e.g. V, other options are H, U or L:

route_pixels %>% 
  ggplot(aes(order, freq)) +
    geom_area(aes(fill = reorder(hex, !!orderby))) +
  scale_fill_identity() +
  scale_color_identity() +
  equal_margins()

(Optional) Smoothing

The landscape is too complex and you want to smooth out the wiggles. This approach uses a loess smoother to average frequencies for each color. Adjust the span closer to 1 for more smoothing, closer to zero for less:

span <- 0.15 
n_points <- 1000 # number of points on x-axis

order_grid <- seq(0, max(route_pixels$order), length.out = n_points)
route_smooth <- 
  route_pixels %>% 
  group_by(hex, L, U, V, H) %>% 
  nest() %>% 
  mutate(
   smooth_fun = map(data, ~ 
      loess(sqrt(freq) ~ order, data = ., span = span)),
   smooth = map(smooth_fun, 
     ~ tibble(
        order = order_grid, 
        freq = predict(., newdata = order_grid)^2)
      )
  ) %>% 
  unnest(smooth)  

There is no guarantee the areas add to a constant, so this often gives a wavy top and bottom:

route_smooth  %>% 
  ggplot(aes(order, freq)) +
    geom_area(aes(fill = reorder(hex, !!orderby), color = hex)) +
  scale_fill_identity() +
  scale_color_identity() +
  equal_margins()

## Warning: Removed 200 rows containing missing values (position_stack).

Scaling to have frequencies add to 1 restores the rectangular boundaries:

route_smooth  %>% 
  group_by(order) %>% 
  mutate(freq = freq/(sum(freq))) %>% 
  ggplot(aes(order, freq)) +
    geom_area(aes(fill = reorder(hex, !!orderby), color = hex)) +
  scale_fill_identity() +
  scale_color_identity() +
  equal_margins() 

## Warning: Removed 200 rows containing missing values (position_stack).

Save image

height and width are in inches, dpi = 300 is good for professional printing:

ggsave(here("oregon", "oregon_route.jpeg"), 
  height = 6, width = 20, dpi = 300)

## Warning: Removed 200 rows containing missing values (position_stack).

cwickham/routes documentation built on Nov. 4, 2019, 9:34 a.m.