# Then vs Now: daily rasters

One cloud-optimized GeoTIFF per (dataset, variable, place, year) with **366 bands, band i =
day-of-year i of a leap-year calendar**: band 1 is 1 Jan, band 60 is 29 Feb, band 61 is 1 Mar in
every year. In a non-leap year band 60 is all NaN. A month-day is therefore the same band in every
year file and in every [climatology](../climatology/README.md) file, with no shifting. Published at
`https://storage.oceanmetrics.io/gazetteer/rasters/`.

Currently: NOAA Coral Reef Watch CoralTemp daily SST (`noaacrwsstDaily` `analysed_sst` on NOAA
CoastWatch ERDDAP, the product PacIOOS re-serves as `dhw_5km` `CRW_SST`, our
[`erddap/dhw_5km`](../erddap/dhw_5km/collection.json)), 1985 to the archive's last day (2026-10-04 at build time), for 13
sanctuaries.

## Files

```
rasters/
  collection.json                                # summaries list dataset, variable, place ids, years
  items/<dataset>_<variable>_<place>_<year>.json  # one STAC Item per file (proj:, file:size, file:checksum)
  dhw_5km/CRW_SST/NMS:FKNMS/1985.tif ... 2026.tif
```

The place id keeps its colon in the object key (`NMS:FKNMS/2020.tif`), as in [`stats/`](../stats/README.md).

## Encoding

| | |
|---|---|
| CRS | EPSG:4326, 0.05° cells on the CRW grid (cell edges on multiples of 0.05°) |
| extent | the sanctuary's bounding box buffered by 20 % of √area (the old climate-dashboard-app box), e.g. FKNMS 78 × 43 px |
| bands | 366, band description = ISO date (`2020-07-18`); `--02-29` for band 60 of a non-leap year |
| dtype / nodata | float32, NaN (no data: land, or a day missing from the archive) |
| layout | COG, `INTERLEAVE=BAND`, 256 × 256 tiles, DEFLATE + floating-point predictor (3) |
| overviews | `OVERVIEWS=AUTO`, which makes none: every grid fits inside one 256 px tile |
| band tags | `STATISTICS_MINIMUM/MAXIMUM/MEAN/STDDEV/VALID_PERCENT` (`nan` and 0 for an empty band) |
| tags | `SOURCE_URL`, `SOURCE_DATASET`, `SOURCE_VARIABLE`, `SOURCE_ARCHIVE`, `PROVENANCE`, `PROCESSING`, `FIRST_DATE`, `LAST_DATE`, `N_DAYS`, `UNITS` |

The current year is partial (it ends at `LAST_DATE`, also the Item's `end_datetime`); the bands
after it are NaN.

## Reading one day in a browser

Because every band is its own single tile, one day for one place is **one HTTP range request of
~6 kB** (FKNMS SST: 410 B to 6.9 kB per band) after the header, which geotiff.js reads once per file:
~150 kB (FKNMS: 147 kB), the IFD with the 366 tile offsets plus the GDAL_METADATA tag holding the
366 band descriptions and per-band `STATISTICS_*`, which the Portolan profile requires to be embedded
in the file. Switching the day re-reads one tile per file; keep the opened `GeoTIFF` objects.

```js
import { fromUrl } from 'geotiff'
const base = 'https://storage.oceanmetrics.io/gazetteer/'
const tiff = await fromUrl(`${base}rasters/dhw_5km/CRW_SST/NMS:FKNMS/2024.tif`)
const img  = await tiff.getImage(0)
const band = 61                                      // 1 Mar (dayofyear of the date in 2000)
const [d]  = await img.readRasters({ samples: [band - 1] })   // Float32Array, row-major, north row first
const bbox = img.getBoundingBox()                    // [west, south, east, north] of the pixel edges
```

maplibre-cog-protocol reads only EPSG:3857 COGs and colours whole tiles, so it cannot address band
*d* of these EPSG:4326 files; read with geotiff.js and draw the array (canvas image source or a
custom layer).

## A custom "Then"

The two prebuilt baselines (1985–2005, 2003–2012) are in [`climatology/`](../climatology/). For any
other range, read the same band from each year file and average per pixel, ignoring NaN:

```js
const years = [1990, 1991, 1992, 1993, 1994]
const bands = await Promise.all(years.map(async (y) => {
  const img = await (await fromUrl(`${base}rasters/dhw_5km/CRW_SST/NMS:FKNMS/${y}.tif`)).getImage(0)
  return (await img.readRasters({ samples: [band - 1] }))[0]
}))
const n = bands[0].length, mean = new Float32Array(n)
for (let i = 0; i < n; i++) {
  let s = 0, k = 0
  for (const b of bands) if (!Number.isNaN(b[i])) { s += b[i]; k++ }
  mean[i] = k ? s / k : NaN
}
```

That is one header + one ~6 kB tile per year: a 20-year custom Then costs ~20 × 155 kB (≈ 3 MB) on
first use and ~20 × 6 kB per further day; a prebuilt climatology costs one header + one tile. For 29 Feb only the leap years contribute.

## Provenance

Converted by [`catalog/build_then_now.py`](../../build_then_now.py) from the archive of the NOAA ONMS
[climate-dashboard-app](https://github.com/noaa-onms/climate-dashboard-app) (`process/erddap.qmd` →
`extractr::ed_extract()`), `mbon:/share/data/noaa-onms/climate-dashboard-app/erddap_sst/<NMS>/<year>.tif`:
float64 → float32 (the source values carry two decimals), bands moved onto the 366-day calendar, and
the transform snapped from its float32 noise (0.04999998°) to the exact 0.05° grid. Archive folder
`CPNMS` (the proposed name) is `NMS:CHNMS`; `MBNMS-david` and `MBNMS-main` (sub-units) have no
gazetteer place and are not published.

## Gaps in the source archive

- **Whole years that are all NaN** in the old pipeline's output (its fetch failed; the CSV is all
  `NA` too): MBNMS 2015; MNMS 1991, 1999, 2015; NMSAS 1991; OCNMS 1999; SBNMS 2005, 2010, 2019. The
  files exist (366 NaN bands) and their Items say `erddap-places:n_days_valid: 0`; the
  climatologies simply have one year fewer there (see `_n.tif`). Re-running the old pipeline for
  those place-years on `mbon` and then `build_then_now.py --codes <NMS> --years <y> <y>` (plus
  `--steps climatology,series,stac` for the place) would fill them.
- **TBNMS (Thunder Bay, Lake Huron) is not published**: CRW CoralTemp has no SST over the Great
  Lakes, so its archive is all NaN (bar a spurious 0 on 2022-12-01).

License: CC-BY-4.0 for these derived files; the CRW data are NOAA, public domain.