--- title: "Subsampling recordings" output: rmarkdown::html_vignette: code_folding: s description: > This article covers the workflow to randomly sample recordings. vignette: > %\VignetteIndexEntry{Subsampling recordings} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r, include = FALSE} library(ARUtools) knitr::opts_chunk$set( collapse = TRUE, comment = "#>" ) options(tibble.print_min = 4L, tibble.print_max = 4L) ``` This vignette will walk through the workflow of subsampling recordings. This can be particularily useful if you have many more recordings than you can interpret manually. ## Create data We will generate some data from the sites in `example_sites`. Click on the triangle below to see the details of this method.
Details of simulation To simulate the file names we generate a series of recordings at each site. Here the schedule is every 30 minutes between 5:30 and 8:00 AM every second day between 1 May and 10 July. Normally you would want to schedule based around local sunrise if targeting the dawn chorus, but this will work for our purposes. ```{r, code_folding = 'hide'} library(dplyr) library(stringr) library(lubridate) simple_deploy <- tidyr::expand_grid( site_id = unique(example_sites$Sites), doy = seq(121, 191, by = 2), times = seq(-30, 120, by = 30) ) |> tidyr::separate(site_id, into = c("plot", "site"), sep = "_", remove = F) |> left_join(example_sites, join_by(site_id == Sites)) |> mutate( # aru_id = glue::glue("BARLT-000{as.numeric(as.factor(site_id))}"), date = ymd("2028-01-01") + doy, date_time = ymd_hm(glue::glue("{date} 06:00")) + minutes(times), date_time_chr = str_replace(as.character(date_time), "\\s", "T"), file_name = glue::glue("{plot}/{site_id}/{ARU}_{date_time_chr}.wav") ) simple_deploy ``` Our site info will be that used in `example_sites`. ```{r} site_info <- simple_deploy |> slice_min(order_by = date_time, n = 1, by = site_id) |> dplyr::select(site_id, ARU, lon, lat, date_time) ```
## Clean metadata To clean the metadata we can use the same code found in the [Getting Started](https://arutools.github.io/ARUtools/articles/ARUtools.html) vignette. I have used a pipe to save space here, but you can see the details in the linked article (`vignette("ARUtools")`). ```{r} sites <- clean_site_index(site_info, name_aru_id = "ARU", name_site_id = "site_id", name_date_time = c("date_time"), name_coords = c("lon", "lat") ) metadata <- clean_metadata(project_files = simple_deploy$file_name) |> add_sites(sites) |> calc_sun() |> dplyr::mutate(doy = lubridate::yday(date)) dplyr::glimpse(metadata) ``` ## Parameters Generally for random sampling you may not want to have recordings selected with equal weight across time and dates. For example if you are sampling songbirds in the breeding season, most species will be most active for a couple hours from around sunrise. You will also want to limit the dates to ensure you are picking up breeding birds and not migrants. To deal with this issue we allow the user to specify selection weights based on the time to sunrise (or sunset) as well as the day of year. To visualize the selection parameters use `gen_dens_sel_simulation`. This will show you how the sample weights will change over time and time to sunrise/sunset. ```{r fig.height=6, out.height=8, warning=FALSE} p <- sim_selection_weights( min_range = c(-70, 240), day_range = c(120, lubridate::yday(lubridate::ymd("2021-07-20"))), min_mean = 30, min_sd = 60, day_mean = lubridate::yday(lubridate::ymd("2021-06-10")), day_sd = 20, offset = 0, return_log = TRUE, selection_fun = "norm" ) ``` ## Calculate selection weights Once you have your parameters set up, you can use them to calculate the sampling weights from the metadata. ```{r} full_selection_probs <- metadata |> calc_selection_weights( col_site_id = site_id, col_min = t2sr, col_day = doy, params = p ) ``` Below you can see the selection weights `psel_normalized`. The highest selection weights should match the output from `gen_dens_sel_simulation()`. Not that as our schedule is set to start based on time, the time to sunrise that the recordings occur differ by sites. ```{r, echo=FALSE} library(ggplot2) ggplot(full_selection_probs, aes(doy, t2sr, colour = psel_normalized)) + geom_point() + scale_colour_viridis_c() + facet_wrap(~site_id) ``` ## Assign sample size The next step is to assign sample sizes. This can be done on a per site basis or across the board. Here I use a 2% sampling rate, but your own subsampling intensity will depend on project goals. You can also set an oversample to draw extra samples in case some samples are unusable (e.g. due to wind). ```{r} sample_size <- count(full_selection_probs, site_id) |> transmute(site_id, n = floor(n * .02), n_os = ceiling(n * .3) ) ``` The figure below can provide a coarse rule of thumb on how many minutes of recordings you might need or number of samples. It of course has many assumptions including knowledge of probability of observing per minute, that this probability will change by date, time of day, conspecific behaviour, weather, etc. ```{r, echo=F} p_obs <- tidyr::expand_grid( p_obs = seq(0.01, 0.99, by = 0.01), n_obs = 1:20 ) |> mutate(p_total = 1 - exp(-p_obs * n_obs)) |> ggplot(aes(n_obs, p_obs, fill = p_total), colour = NA) + geom_raster() + scale_fill_viridis_c() + labs( x = "Number of minutes observed", y = "Probably of observing per minute", fill = "Total\nprobability\nof\ndetection" ) p_obs ``` ## Draw subsample ### GRTS There are a few ways to subsample recordings. `ARUtools` has imported the [grts()](https://usepa.github.io/spsurvey/articles/sampling.html) (Generalized Random Tessellation Stratified algorithm) function from [spsurvey](https://usepa.github.io/spsurvey/) package. We developed a wrapper around the function to simplify it's usage for our particular use case. GRTS allows us to sample using a dispersed samples, while maintaining a stochastic element to sampling. In our case we are selecting samples dispersed across dates and time in the dates. ```{r} grts_res <- sample_recordings(full_selection_probs, n = sample_size, col_site_id = site_id, seed = 2024, col_sel_weights = psel_normalized ) dplyr::glimpse(grts_res$sites_base) ``` ### slice_sample If you don't want to sample using `grts()` you can just use `dplyr::slice_sample()`. ```{r} withr::with_seed(2024, { random_sample <- full_selection_probs |> dplyr::slice_sample( n = 4, by = site_id, weight_by = psel_normalized, replace = F ) }) ```
If you want to set site specific sample sizes, you can use some dplyr tricks to handle that: ```{r eval=F} withr::with_seed(2024, { random_sample_stratified <- full_selection_probs |> left_join(sample_size, by = join_by(site_id)) |> nest_by(site_id, n) |> rowwise() |> mutate(sample = list(dplyr::slice_sample( .data = data, n = .data$n, weight_by = psel_normalized, replace = F ))) |> dplyr::select(site_id, sample) |> tidyr::unnest(sample) }) ```
Oversamples with random sampling, just involves dropping recordings you've selected and drawing some more. ```{r} oversample <- filter( full_selection_probs, !path %in% random_sample$path ) |> dplyr::slice_sample( n = 2, by = site_id, weight_by = psel_normalized, replace = F ) ``` ### Compare methods Looking at the draws, you may notice the `grts` results are more spaced out, while the random sample, contains more clumping, which is characteristic of random sampling. ```{r, echo=F} filter(full_selection_probs, path %in% grts_res$sites_base$path) |> ggplot(aes(doy, t2sr)) + geom_point() + xlim(range(full_selection_probs$doy)) + ylim(range(full_selection_probs$t2sr)) + labs(x = "Day of Year", y = "Time to sunrise", title = "GRTS selection") ``` ```{r, echo=F} random_sample |> mutate(z = 1) |> ggplot(aes(doy, t2sr)) + geom_point() + xlim(range(full_selection_probs$doy)) + ylim(range(full_selection_probs$t2sr)) + labs(x = "Day of Year", y = "Time to sunrise", title = "Random selection") ``` ## Assigning recording lengths While some surveys may include all full recordings, others may only interpret a portion of some recordings, allowing for more recordings to be at least partially interpreted. To assign lengths, we can repeat the same process as above sequentially assigning lengths or more simply just assign them all at once. ```{r} withr::with_seed(6546, { random_sample$length <- sample( x = c("5min", "3min", "1min"), size = nrow(random_sample), replace = T ) }) ``` ```{r, echo=F} count(random_sample, length, site_id) |> tidyr::pivot_wider( names_from = site_id, values_from = n, values_fill = list(n = 0), values_fn = as.numeric ) ``` This however creates the issue that some sites will have less and some more of each length. ```{r} withr::with_seed(569, { sample5min <- slice_sample(random_sample, n = 1, by = site_id, weight_by = psel_normalized ) sample3min <- slice_sample( random_sample |> filter(!path %in% sample5min$path), n = 1, by = site_id, weight_by = psel_normalized ) random_sample_with_lengths <- random_sample |> mutate( Length_group = case_when( path %in% sample5min$path ~ "5min", path %in% sample3min$path ~ "3min", TRUE ~ "1min" ), length_clip = as.numeric(str_extract(Length_group, "^\\d")) * 60 ) }) ``` Sampling by site and length creates an equal sample by length and site. ```{r, echo=F} count(random_sample_with_lengths, Length_group, site_id) |> tidyr::pivot_wider( names_from = site_id, values_from = n, values_fill = list(n = 0), values_fn = as.numeric ) ``` # Assign start times ARUtools includes the function `get_wav_length`, which will return the length of a 'wav' file in seconds. ```{r, eval=F} random_sample_with_lengths$length_clip <- purrr::map( 1:nrow(random_sample_with_lengths), ~ get_wav_length( path = random_sample_with_lengths$path[[.x]], return_numeric = T ) ) ``` This can be used to safely add a random start time to your recordings if you don't want to all start at zero. For this document we will assume all recordings are 5 minutes. ```{r} random_sample_with_lengths$length <- 5 * 60 ``` ```{r, eval=T} random_sample_with_lengths <- random_sample_with_lengths |> rowwise() |> mutate(StartTime = case_when( Length_group == "5min" ~ 0, TRUE ~ runif( 1, 0, pmax( 0, length - length_clip ) ) )) |> ungroup() ``` ```{r, echo=F} ggplot(random_sample_with_lengths, aes(StartTime, fill = Length_group)) + geom_histogram(binwidth = 10) ``` ## Format output filename Finally we can format the filenames for [Wildtrax](https://portal.wildtrax.ca/), which processes sites, dates, and times based on file name. ```{r} final_selection <- random_sample_with_lengths |> add_wildtrax() final_selection |> head() |> dplyr::select(path, wildtrax_file_name) ``` ## Clip and copy files for upload The `format_clip_wave` will rename, clip and copy your file to a location that can then be used for uploading to [Wildtrax](https://portal.wildtrax.ca/) Setting up a subset of folders will allow you to copy files to an organized folder structure which can help with checking for errors prior to uploading. ```{r, eval=F} out_directory <- "/path/to/upload/directory/" dir.create(out_directory, recursive = T) ul_tab <- expand_grid( period = c("Dawn"), # Add 'Dusk' if using more than one time period length = unique(selected_recordings$Length_group) ) purrr::map(glue::glue("{out_directory}/{ul_tab$period}/{ul_tab$length}"), dir.create, recursive = T ) ``` If you have a large number of files, you may want to run this separately for files that require clipping and those that do not as it will run much quicker for the files that do not need to be clipped. You can use the option `use_job=T` to allow the use of the `job` package, which will launch a background job if the copying will take a long time. ```{r eval=F} log_output <- format_clip_wave( segment_df = final_selection, in_base_directory = "", out_base_directory = out_directory, length_clip_col = "length_clip", sub_dir_out_col = c("Time_period", "Length_group"), filepath_in_col = "path", out_filename_col = "wildtrax_file_name", use_job = F, filewarn = F ) ``` That's it, your output folder should now have your selected files, properly named for [Wildtrax](https://portal.wildtrax.ca/) and ready to upload.