[NewGraphEnvironment]

Summary

New pkgdown article, When Did the Floodplain Change? Dating Detected Change with Sentinel-2 and Esri Wayback (vignettes/articles/floodplain-change-dating.Rmd). It replaces the first version of this PR (wayback-change.Rmd), which covered Wayback mechanics and was built on the superseded bulk_co_ff04 item. The new article follows the use case on the Bulkley watershed group's floodplain, along Maxan and Crow creeks: date the change drift detects in a floodplain, then look at each change with the best imagery available.

• The data comes from the current STAC item. bulk_all_ff04 on stac-floodplains-bc, read with rstac: the all_ff04 floodplain and the 2017–2023 IO LULC land cover. dft_rast_break_class() and dft_rast_break_category() give each floodplain cell the year of its clean switch. Connected same-year switches of 1 ha or more become the lettered change areas A–J.
• The first map shows where the box is: the Bulkley watershed group's floodplain, the Bulkley River and the towns on a shaded grey relief basemap (Esri Light Gray Base x World Hillshade), with a BC inset. A second map colours floodplain cells by switch year over the 2023 Sentinel-2 composite.
• Each change area gets a verdict (change / class error where Wayback brackets it; likely change / unclear where only Sentinel-2 does), shown above its panels with a Google Maps satellite link for an undated look at the ground today. The table groups the areas by verdict, class errors first. The verdicts are keyed to areas.csv and stop the build if the areas change.
• The article links the stac_floodplains_bc catalogue and the floodplains modelling sites.
• What the sharp images show (2018–2020). For the five areas Wayback brackets (2016-09-11 before, 2020-09-18 after), only one (C, creekside forest opened up) is a change on the ground. Three are parts of one open meadow relabelled from Crops to Rangeland, and one is the same stand of trees. Sentinel-2 brightening alone would have flagged one of the fields as cleared.
• After 2020, Sentinel-2 alone. In the last sharp image, four of the five 2021–2023 areas are forest and one is an open wet meadow. In its year, each brightens more than the box as a whole does. One lies partly in a 2023 cutblock.
• Why the sharp images stop. Over the box, every Wayback release from 2022-05-18 to the latest (2026-09-24) still draws the 2020-09-18 capture.
• The rest: one table of tools, one table of numbers, and short limits. All code folds under a "Code" toggle. The fetch code shown is read from the data-raw script, so what is shown is what ran.
• data-raw/floodplain_change_dating.R makes every network call and writes 31 small files (5.9 MB) to vignettes/articles/floodplain-change-dating/, which is outside the tarball. The article reads only those files.

Closes #108

What the data changed

• The verdicts for A–E came from looking at the Wayback views at full size, not from the land-cover labels. Four of the five dated switches are the map, not the ground. That became the article's main point.
• Each area's Sentinel-2 brightening is measured against the whole box's. The 2019 composite is pale across the box, and 2023 is brighter, so a raw before/after would date haze.
• Code folding: pkgdown 2.2.0 ignores code_folding: hide, so a knitr source hook wraps each chunk in <details>. knitr::read_chunk() of a ../../data-raw path breaks the build, so the article reads the script with readLines() and passes the code through code =.

Measurement

• Data script: 1.4 min on a warm cache. A warm rerun reproduces all 28 files byte for byte (md5). The first cold run took 32.9 min for 25 of 182 Wayback metadata releases, then hit a transient HTTP 502. It was cold only because the box corners had been rounded differently from the old script's, so every cache key missed. With the old full-precision corners, all 182 releases read in 0.5 min.
• Body prose at the redo: 884 words, Rmd 348 lines (previous article: 1,477 words, 414 lines). The verdicts, links and context map since then bring the Rmd to 445 lines.
• No R/ changes. devtools::test(): FAIL 0 | PASS 1709 | SKIP 19. pkgdown::check_pkgdown(): no problems. No footnotes or \@ref() in the built HTML.

Checks

• pkgdown::build_article() renders from the committed files; every dft_ call is in an eval = FALSE chunk
• The code folds (8 <details> blocks), including the chunk that reads the saved files, so the figure code on the page runs as shown
• Every date and count in the prose is computed from the CSVs; the per-area verdicts are keyed to areas.csv and stop the build if it changes

Relates to NewGraphEnvironment/sred-2025-2026#16

🤖 Generated with Claude Code

https://claude.ai/code/session_01PRhUJsuKABLfpBGktPoiBN