Skip to content

Dashboard

The planning dashboard answers one question for a place, a health outcome and a month: how much of this outcome may be attributed to heat, and where.

It is composed of three cards, all reading the same selection.

The selection

The URL records the geography, optional admin_unit, outcome, month, and view (observed, forecast or projection). The forecast keeps its own month in forecast_month. Projected choices also retain scenario, period, selector, and reference; the observed group uses observed_selector.

The filter bar labels every dropdown (place levels, Climate hazard, Health domain, Health outcome) and ends with Configure plan, which opens the plan wizard (below) for the same place and outcome. Domains no model covers yet are listed as "coming soon" and cannot be picked.

/dashboard/geo-in-madhya-pradesh?admin_unit=geo-in-madhya-pradesh-division-bhopal&outcome=lbw&month=2026-08

They live in the URL rather than in component state so a reload, the back button and a shared link all land on the same view. Changing any of them re-reads every card.

Each geography dropdown names its own level: All divisions, All districts, All counties, or All constituencies. Choosing this first option clears the child selection and returns to the current parent's map and available results, retaining the other filters. The vulnerability page uses the same place selectors. Routing-only roots such as India do not offer an aggregate option when they lead directly to their supported state.

Which pregnancy windows or age groups, scenarios and data views the cards offer is set per country and outcome under Settings → Models → Model options. A pregnancy window or age group dropdown appears only when more than one option is switched on; with just the default, the cards use it without asking. The API applies those settings, so the dashboard only lists what it is given. See Model settings.

The breadcrumb under the selectors is built by walking the geography ancestry of the selected area, broadest first — Kenya › Garissa, or India › Madhya Pradesh › Bhopal Division. It follows the sub-area picker, and consecutive repeats collapse, so a country-level view shows Kenya once rather than Kenya › Kenya.

Clicking an area with children opens its child map: Madhya Pradesh → division → district, or Kenya → county → constituency. Back to … returns to the parent map. Map navigation retains the outcome, view and compatible filters, as do reload and browser history. Both view tabs remain selectable at finer levels. Where an observed model is unavailable, the dashboard explains this and keeps the geographic outlines visible with no health values; going back restores the parent's available observed view.

The risk overview tabs

The climate-attributable risk card has three tabs that share one layout: switching changes the month or period controls, the figures, the icon array and the map's shading, not the place, question or columns.

  • Observed — a year and a month with observed data (ERA5); months without data stay in the grid, greyed.
  • Forecasted — the C3S multi-system seasonal forecast, the same one Explore shows: up to six months ahead, shown as one row of month chips with the year written on the first month and on January (no year dropdown). It is loaded only when the tab opens, and a place without a saved forecast says so. Forecasts cannot be saved to a plan yet, so Save plan is disabled on this tab.
  • Projected — three rows: Time horizon (the window lengths, e.g. 10 or 20 years, when the release publishes several — picking one opens its first future period, named in the sentence; otherwise the periods themselves), Emission scenario (cards with the plain-language name over the SSP code) and Reference period (annual average; peak heat season, the share over the place's heat season as set in Settings › Heat seasons, named without its months for now; or highest-risk months, the three months the model ranks highest).

The panel shows no caveat card: a release's model caveat (for example the national LBW model's climate-zone note) stays in the API's selector_note and in the release documentation, and is used only as the message when no estimate is reported. A See more link beside the panel label opens the published methodology; there are no seasonal or all-periods tables.

Switching place, tab, month or scenario never blanks the card. The overview is drawn by the dashboard layout (web/src/app/(workspace)/dashboard/layout.tsx) rather than the [geo] page, so a new place updates the mounted dashboard instead of rebuilding it. Reads go through one client-side query cache (TanStack Query, web/src/lib/queryClient.tsx): the place list, model catalogs, capabilities and each outlook view are kept for the session, so a place, month or setting seen before answers from memory. Published views are not re-read during a session; saving model settings in Settings clears them (and the kept maps) so the dashboard reads them afresh.

All three tabs (Observed, Forecasted, Projected) load together, and once the open tab is drawn the other tabs' maps are fetched too, so switching tab is instant. The icon array and the map stay mounted throughout: while a new selection loads, the icon array keeps its last share and the map its shapes and last colours, then the figures shade over and the map's fills fade to the new values. Each place's outlines are kept from its first map; every later map of that place (another tab, month or setting) is requested with geometry=false and carries only each area's value, reason and odds ratio, which are laid on the kept outlines. Published reads and outlines are also kept in the tab's sessionStorage, so a reload draws them at once and refreshes them behind the page; the stored copy names its user and is dropped for anyone else. While the session is restored the dashboard placeholder already shows the science video. Signing out clears the cache. Settings keeps a tab you have opened mounted (hidden) so going back to it does not reload it.

The sentence under the icon array reads "At a [forecasted / projected average annual] maximum temperature of X °C in , N% of cases in may be attributable to heat exposure." The temperature is the month's ERA5 mean daily maximum (observed), the central forecast (forecasted) or the mean of the period's typical months (projected); without one the sentence starts at "In ". Beneath it: the share's 95% CI (or the forecast-system / climate-model range), the odds or risk ratio with its 95% CI and the precision badge (High, Moderate or Low, from the ratio's interval).

The cards

Risk vs protection

Illustrates what the selected model is about. The figures are not fixed: each model release declares them in its manifest (presentation.visualization), so low birth weight draws a newborn and a pregnant woman while under-five mortality draws a child. A new outcome brings its own illustration rather than inheriting another's.

Heat and the health outcome

The number, for the selected month. Choosing a month is choosing a temperature: the month's observed exposure is what the model scores.

Two things on this card are easy to misread and are therefore stated explicitly:

  • The model reads a window, not one month. A low-birth-weight model scores the selected month and the two before it, and the card lists all three temperatures beneath the headline. Two months can show the same temperature and return different answers because their earlier months differ.
  • A zero is explained, over the same window. If none of the scored months reached the model's reference temperature, nothing is attributed to heat. If they did but the combined odds ratio is 1 or less, the card says the model finds no extra risk, gives the odds ratio and its range, and says when that range is too wide to tell either way. It never prints a bare "0%".

Spatial risk map

Every administrative area beneath the selected geography, shaded by attributable fraction. Areas are administrative shapes carrying an area-level value — not a modelled grid, and they must not be drawn as cells, which would imply a spatial resolution the model does not have.

Geometry is simplified for drawing only (the response discloses the tolerance); every climate extraction uses the unsimplified boundary.

Moving around it stays smooth, as in Explore: resting the pointer on an area for a moment fetches the map that clicking it opens, and opening a place whose map is not loaded yet keeps the last map on screen, dimmed, until it arrives, never a blank. Up to 60 maps are kept for the session (cleared on sign-out). Custom maps with their own loader (outlook, vulnerability) keep the last-map behaviour but are not fetched ahead.

A click never empties the risk card either. Which views a place and outcome publish (its capabilities) are kept for the session, and while the next place's arrive the current card stays drawn; only the first load of the dashboard shows its placeholder.

For what the map queues by itself and what its four states mean, see Data pipeline.

Where a number comes from

Each stored prediction is a row in prediction_result, keyed by admin unit, outcome, model release, month, scenario, horizon and pregnancy window. Results are shared: a month computed by one planner is visible to everyone with access to that place, and is never recomputed for the next person.

Each row carries the model release and artifact checksum that produced it, the exposure vector it was scored on, and the reference it was measured against, so a figure on screen can be traced back to the fitted model and the observations behind it.

Saved plans and recommendations

Save plan on the outlook card saves the outcome, place (with any sub-area) and the observed or projected selection under a name. The API resolves the selection and captures the estimate itself through the same outlook service the card reads: the attributable share with its 95% CI, the odds or risk ratio, the model release and the country's official heat season. A plan therefore keeps the modelling it was made from after a newer release is published. Observed views from India's live monthly model are saved without a figure, with the reason recorded. Plans are private to their owner, at most 100 per user, and hidden once their place leaves the user's geography scope (/saved-plans, see the API reference).

Saved plans are listed on the Planning center, where each reopens its dashboard view or its recommendations. The action repository page opened with ?plan=<id> ranks the repository's actions for the plan:

  • The API reads the figure as a risk level: clearly raised when the ratio's 95% CI lies above 1, possibly raised when only the central estimate does (or a share is reported without a ratio), not detected otherwise, and unknown without a figure.
  • Timing: a projection plans ahead (before the heat season); an observed month inside the official season is during, up to two months after it after, otherwise before.
  • The API ranks the actions (rules by default, or a self-hosted classifier) and shows the top three with the reasons they rank.

The insight beside them is a template unless an administrator chose a language model; see Recommendations and AI. The figure and the ranking never depend on it.

Configure plan

Configure plan opens a five-step wizard at /dashboard/<geo>/configure?outcome=…&admin_unit=…&plan=<id>&step=<step>: Timeline, Hazards & outcomes, Districts, Actions, and Review & export. The first four are built; Review & export shows in the step tabs but cannot be opened yet.

  1. Timeline. The plan's name and the dates it runs between. Quick select picks the next planning cycle (today to the end of the review months after the next heat season), the heat season only, or the next 6 or 12 months, from the place's calendar in Settings > Heat seasons (the official season until one is set). A bar colours each month by its phase. A plan starts today or later and ends within 24 months of its start; the API checks the same rules and allows a day's difference for the browser's clock. Next saves the plan as a draft: a saved plan with status: draft, the dashboard's projection captured as for any saved view.
  2. Hazards & outcomes. The hazard the models cover for the outcome (extreme heat) and its health domain, with the roadmap's hazards and domains listed as coming soon, and the health outcomes to address (at least one).
  3. Districts. The areas beneath the plan's place (districts, divisions or counties) beneath the place the dashboard framed, ranked by the attributable share for the plan's first outcome, from the same map the action repository's priority areas read: the published outlook's annual average where it covers the place, else the observed risk map's latest results. Areas at 5% or more are pre-ticked as Suggested to prioritise, with the area the dashboard was showing when Configure plan was pressed; areas without an estimate are listed last and never suggested. Choose on the list or the map (search, Suggested, Select all, Clear; at least one). Save draft stores the choice with each area's rank, share and source (plan_district, migration 038_plan_district), so the report shows the figures the plan was made with.
  4. Actions. The action repository's actions for the plan's hazard in its country, ranked by the recommendation rules read against the plan: an action whose season falls inside the plan's dates (on the place's heat-season calendar) scores 4 and is pre-ticked as Recommended for this plan, one that targets an outcome the plan addresses 3, and unless heat risk is clearly raised a cheaper one 2 or 1. Each card shows its reasons and the plan's suggested districts. A ticked action gets an owner (one of the country's departments) and a due date inside the plan's dates. Save plan stores them (plan_action, migration 039_plan_action, with the action's title, seasons and cost as the repository had them) and marks the plan saved.

Until it is saved the plan is a Draft in the Planning center, and Continue configuring reopens it; a saved plan made in the wizard has Edit plan.

Steps 1 and 2 are saved with PATCH /saved-plans/{id}, step 3 with PUT /saved-plans/{id}/districts and step 4 with PUT /saved-plans/{id}/actions then PATCH status: saved (each GET returns the ranking); the step and the draft live in the URL, so a reload or Back returns to the same step. Plans saved from the dashboard before the wizard are kept as saved, with no dates, under extreme heat and their own outcome (migration 037_plan_configuration).

Planning center

The Planning center (/plan, "Planning center" in the sidebar) builds one sentence: "We're planning together for the impacts of [climate hazard] on [health domain] in [place]." Both choices start empty, and Start planning together stays disabled until both are made. The enabled options are the hazards and domains the place's models cover; the roadmap's others (Flooding, Air pollution; Vector-borne diseases, Heat-related illnesses, Non-communicable diseases) are listed as "(coming soon)" and cannot be chosen. The outcome is picked on the dashboard, which opens on the first one the models estimate.

Planning starts in Madhya Pradesh for India and in Kajiado, its co-design county, for Kenya (PREFERRED_START_AREAS in web/src/features/planning/contextFamilies.ts). Switching country on the dashboard still opens the country or state as a whole.