tgf_wmo_plugins

TethysDash visualizations for the WMO impact-based forecasting trainings in Guatemala, Haiti and Antigua and Barbuda, in English, Spanish and French.

Every plugin family shares one implementation, and its language variants differ only in the strings they render — all of which live in strings.py, in the three languages — so the languages cannot disagree about a number.

The repository has three parts besides the plugins: docs/ holds the step-by-step guides for building the hands-on dashboards, dashboards/ the finished dashboards those guides build, and notebooks/ the analysis behind them as plain Python. All three are organised by country; start at docs/README.md.

Plugins

entry point	type	what it shows
wmo_impact_summary_en / _es	table	People, buildings and roads in each hazard level
wmo_hazard_layer_en / _es	map_layer	Hazard classification as map polygons
wmo_impact_layer_en / _es	map_layer	Buildings and roads coloured by hazard level
wmo_storm_card_en / _es	card	Magnitude, flooded area and depth for one storm
wmo_storm_impact_summary_en / _es	table	People, buildings and roads by flood depth for one storm
wmo_storm_impact_layer_en / _es	map_layer	Buildings and roads coloured by flood depth for one storm

The three map layers are dynamic_map_layers: they re-fetch whenever a bound variable input changes, which is what makes the thresholds and the storm slider interactive.

Country variants

The hazard family also ships per-country variants under the uffis_ prefix. Each one is a subclass that sets LANG, country and name and nothing else; the country selects the data (PROB_URLS, FEATURES_URLS, PROB_FIELDS, GPKG_LAYERS) and the language selects the strings.

entry point	language	data
uffis_impact_summary_guatemala	es	Guatemala_IBF/impact_features.csv
uffis_impact_layer_guatemala	es	Guatemala_IBF/impact_features.geojson
uffis_impact_summary_haiti / uffis_impact_layer_haiti	fr	Haiti training/outputs/impact_features.gpkg
uffis_impact_summary_antigua_barbuda / uffis_impact_layer_antigua_barbuda	en	antigua_barbuda_IBF/AntiguaBarbuda_Jerry_cycle_10151010_IBF_outputs.gpkg
uffis_hazard_layer_antigua_barbuda	en	antigua_barbuda_IBF/depth_prob/antiguabarbuda_prob_depth_ge_*_overbank.tif

Haiti and Antigua and Barbuda ship geopackages with buildings and roads as separate layers, which are stacked into one frame with a type column; in both, the buildings layer misnames its severe probability probability_30cm, and GPKG_LAYERS renames it back before the gates apply. The Antigua and Barbuda probability rasters are EPSG:4326 at 1 arc-second (2658×921), unlike Guatemala's EPSG:3857 copies, and are served in that CRS since OpenLayers resolves it natively.

Two products, two questions

The plugins split into two families that answer different questions, and the distinction is worth keeping straight when building a dashboard:

• Hazard (impact_summary, hazard_layer, impact_layer) works from four EF5 exceedance-probability rasters. Each hazard level has its own probability gate, and a feature takes the level of the deepest flood threshold it clears. It answers "what are the odds of at least 30 cm here".
• Storm (storm_card, storm_impact_summary, storm_impact_layer) samples depth from one storm of a 200-member RainyDay ensemble. It answers "in storm 150, how deep is the water on this building". Depth is a physical quantity, so the bands need no reference to a threshold.

The two datasets sit on exactly the same grid — EPSG:3857, 424×319, 5 m cells, identical origin — so no reprojection or resampling happens anywhere.

Dashboards

dashboards/ holds the ready-to-import dashboards for the training, one folder per country. Each is the finished solution to one of the three hands-on exercises, whose step-by-step guides live in docs/<country>/.

file	dashboard	guide
Guatemala/Guatemala_Hands_On_1_English.json	Guatemala Hands On 1 (English)	docs/Guatemala/exercise_1_en.rst
Guatemala/Guatemala_Hands_On_1_Espanol.json	Guatemala Práctica 1 (Español)	docs/Guatemala/exercise_1_es.rst
Guatemala/Guatemala_Hands_On_2_English.json	Guatemala Hands On 2 (English)	docs/Guatemala/exercise_2_en.rst
Guatemala/Guatemala_Hands_On_2_Espanol.json	Guatemala Práctica 2 (Español)	docs/Guatemala/exercise_2_es.rst
Guatemala/Guatemala_Hands_On_3_English.json	Guatemala Hands On 3 (English)	docs/Guatemala/exercise_3_en.rst
Guatemala/Guatemala_Hands_On_3_Espanol.json	Guatemala Práctica 3 (Español)	docs/Guatemala/exercise_3_es.rst
Antigua_Barbuda/Antigua_Barbuda_Hands_On_1.json	Antigua and Barbuda Hands On 1	docs/Antigua_Barbuda/exercise_1.rst
Antigua_Barbuda/Antigua_Barbuda_Hands_On_2.json	Antigua and Barbuda Hands On 2	docs/Antigua_Barbuda/exercise_2.rst
Antigua_Barbuda/Antigua_Barbuda_Hands_On_3.json	Antigua and Barbuda Hands On 3	docs/Antigua_Barbuda/exercise_3.rst
Antigua_Barbuda/Antigua_Barbuda_Hands_On_3_Depth.json	Antigua and Barbuda Hands On 3 with Depth	docs/Antigua_Barbuda/exercise_3.rst, step 10

Import them from the landing page once the plugins are installed and the server has restarted. Exercise 1 uses no plugins at all; exercise 2 uses the storm family (Guatemala only); exercise 3 uses the hazard family.

The Antigua and Barbuda set follows the same three exercises on the Tropical Storm Jerry data. Exercise 1 shows the four depth_prob rasters with the GADM parish boundaries, plus flood depth for one storm read straight out of the Saint John's flood-map library (AnB_IBF/AG04_SaintJohnS_v1.zarr, 200 storms) through a Zarr layer with a fixed index; there is no standalone depth GeoTIFF for Antigua and Barbuda. Exercise 2 binds that index to a Storm number input; no storm plugin exists for these libraries yet, so there is no table or card. Exercise 3 binds the three uffis_*_antigua_barbuda plugins to four threshold inputs, preset to the plugin defaults.

Hands On 3 with Depth is exercise 3 with exercise 2's depth layer added: a Storm number input drives a Zarr layer on the Saint John's flood-map library under the parish outlines and the buildings and roads at risk, so one dashboard shows how deep the water gets in one storm next to which features the probability gates classify. The hazard classification polygons are opaque and would cover the depth, so they start hidden; turn them on from the layer control. The map opens on Saint John's because the depth library covers only that parish.

A dashboard is bound to the plugins of its own language, and not only through the source names. A variable input that draws its options from a plugin argument stores the string "<group>: <label> - <Arg>", which embeds the plugin's translated group and label — so the English and Spanish copies of exercise 2 reference different option sources for the same Storm slider. Renaming a plugin's label or group breaks that binding silently; regenerate the dashboards rather than editing the strings by hand.

Installing

pip install -e .

Then restart the Tethys server; TethysDash discovers plugins through the intake.drivers entry-point group at startup.

Adding or changing text

All user-facing text lives in strings.py, keyed by language. Hazard levels and depth bands are keyed by their numeric class value — the value the raster and the feature attributes actually carry — so a translation can never change a classification.

check_parity() runs at import and raises if the two dictionaries have drifted, so a missing translation fails at install time rather than in front of a room of trainees.

Data

Everything is read from a public S3 bucket; nothing is bundled.

dataset	what it is
floodmaps_test/	Zarr store, 200-storm RainyDay ensemble, depth in metres
floodmaps_test/ensemble_stats.csv	Precomputed per-storm summary
Guatemala_IBF/impact_features.geojson	Buildings and roads with sampled probabilities
Guatemala_IBF/impact_features.csv	The same rows without geometry, for tables
PBI_Actividad_2/prob_*.tif	The four exceedance-probability rasters, EPSG:3857
PBI_Actividad_2/depth_m.tif	Flood depth in metres, EPSG:3857, on the same grid

Two caveats that matter when reading the output:

• The probability layer depths are mislabelled upstream. The filenames say 7.62 / 10 / 30 / 76 cm, but three of the four are actually 15.24 / 30.48 / 60.96 cm. The ordering is correct either way, so the classification is unaffected — but the labels are not what they claim.
• The 76 cm layer only ever contains 0 or 0.2. Any gate above 0.2 therefore makes the top hazard level unreachable, not merely rare.

The dashboards deliberately use the PBI_Actividad_2/ copies rather than the Guatemala_IBF/ originals. The originals are EPSG:32615 (UTM 15N), and a UTM raster makes the map adopt that projection through the GeoTIFF auto-fit, which costs a reprojection on every tile. All five 3857 rasters share one grid exactly — 424×319, 5 m cells, identical origin — which is also the grid the RainyDay ensemble uses, so nothing is resampled at render time.

depth_m.tif was resampled from the UTM original with nearest neighbour rather than bilinear: the layer is drawn with mask_below: 0.01, and averaging across the dry/wet boundary inflated the flooded footprint by 16% and clipped the peak depth. Nearest holds the footprint to 7,705 cells against the original's 7,710, the mean to 0.9281 m against 0.9289, and the maximum exactly.

Site

Santa Inés Petapa, Guatemala. The model domain is 3.38 km² and covers 5,029 of the 5,147 features in the geopackage; the remaining 118 fall outside it and never appear as affected, whatever the storm.