Plot test dataset incorrectly classified time series

[lindsayplatt]

Look at the time series for sites incorrectly classified. See if we are still happy with those

[lindsayplatt]

For the latest model, here are those time series for both the train and test data.

Code for finding those sites and time series

library(targets)
library(tidyverse)
library(sf)

source('5_DefineCharacteristics/src/visualize_attribute_distributions.R')
source('5_DefineCharacteristics/src/prep_attr_randomforest.R')

# Load the new model
rf_model <- tar_read(p5_rf_model_optimized)
rf_model_test_pred_results <- tar_read(p5_rf_testpreds) 

tar_load(p5_site_attr_rf_optimal)
tar_load(p5_site_attr)
rows_used <- as.numeric(names(rf_model$predicted))
incorrect <- which(rf_model$predicted != rf_model$y)
sites_wrong_rf <- left_join(p5_site_attr_rf_optimal[rows_used[incorrect],], p5_site_attr) %>% 
  select(site_no, site_category_fact)

rows_used_test <- as.numeric(rownames(rf_model_test_pred_results))
incorrect_test <- rf_model_test_pred_results %>% 
  mutate(row_num = rows_used_test) %>% 
  filter(site_category_fact != site_category_predicted) %>% pull(row_num)
sites_wrong_rf_test <- left_join(p5_site_attr_rf_optimal[incorrect_test,], p5_site_attr) %>% 
  select(site_no, site_category_fact)

tar_load(p3_ts_sc_qualified)
ts_sc_incorrect_sites <- p3_ts_sc_qualified %>% 
  filter(site_no %in% c(sites_wrong_rf$site_no, sites_wrong_rf_test$site_no)) %>% 
  mutate(rf_type = ifelse(site_no %in% sites_wrong_rf_test$site_no, 'Test', 'Train')) %>% 
  left_join(bind_rows(sites_wrong_rf, sites_wrong_rf_test)) %>% 
  mutate(year = year(dateTime))
  
ts_sc_incorrect_sites %>% 
  filter(rf_type == 'Test') %>% 
  ggplot(aes(x = dateTime, y = SpecCond, color = site_category_fact, group=year)) +
  geom_line() + 
  facet_wrap(vars(site_no), scales='free', ncol=2) +
  ggtitle('Incorrectly predicted sites, test data') +
  theme_bw()

ts_sc_incorrect_sites %>% 
  filter(rf_type == 'Train') %>% 
  ggplot(aes(x = dateTime, y = SpecCond, color = site_category_fact, group=year)) +
  geom_line() + 
  facet_wrap(vars(site_no), scales='free', ncol=4) +
  ggtitle('Incorrectly predicted sites, train data') +
  theme_bw()