Taos Engine · Analysis Tutorials

Analysis Tutorials · Step 1

Your First Graph

You'll learn: the graph editor's core moves — adding nodes with right-click, connecting typed ports, setting params, running, and reading the result — by building a genuinely interesting map: how fast every patch of ocean has warmed or cooled since 2016.

The question we're answering: "What's the sea-surface temperature trend, per decade, everywhere on Earth?" Four nodes: a data source, a reduction, and two sinks.

layer(sst) ──► trend ──┬──► display   (the map)
                       └──► answer    (the numbers)

Open the editor#

  1. Clone Earth Explorer, run npm run dev, and open the page it serves.
  2. Click ⚗ Analysis in the top bar, then the Graph tab. The panel widens into a toolbar, an issue line, and an empty canvas.
  3. Try the navigation before adding anything: drag the empty canvas to pan, scroll to zoom (it anchors at your cursor), press ⟲ in the toolbar to reset.

Add the source#

  1. Right-click anywhere on the canvas. The add-node menu opens, grouped the way the language is organized — sources, transform, reduce, combine, sinks. Hover any entry for a one-line description.

  2. Pick layer (under sources). A node card appears at the click point with the explorer's first layer, sst, already selected.

  3. Set its params (they commit when the field loses focus):

    • start → 2016-01
    • stepMonths → 1 (monthly frames — trends like dense sampling)
    • leave end empty (defaults to now) and component at — (that's only for vector layers like wind).

    Notice the blue dot on the node's right edge — that's its output port, and blue is the color of a stack (time × grid). Port colors are how you read a graph at a glance.

Reduce time away#

  1. Right-click to the right of the layer node and add trend (under reduce). Its input port value is blue too — it wants a stack.

  2. Drag from the layer's output dot onto trend's value dot. While you drag, watch the other ports: anything that would reject the connection dims out. Release on the dot and a colored edge appears.

    trend fits a least-squares slope through every cell's monthly values and reports it per decade — turning ~120 maps into one map of "°/decade". Its output dot is lavender: a field, one grid with no time axis left.

Sink it#

  1. Add display (under sinks) and connect trend → display.value. Set its title param to something like SST trend per decade.
  2. Add answer and connect trend → answer.value too (one output can feed many inputs). Set label to SST trend per decade. The label is required — leave it empty and the node turns red; hover it to read the validation message. That red-badge feedback is live on every edit.

Run it#

  1. Press ▶ Run. The status line below walks through Loading data… (the baked OISST stack, a few seconds) → Computing… → done · ~7M cell-ops.

    Three things happen:

    • The map takes over the display: a diverging red/blue map on a symmetric legend — red where the ocean warmed, blue where it cooled. The legend range came from the data's percentiles automatically, because trend marks its output as a relative (signed) quantity.
    • The answer line appears: something like SST trend per decade: mean +0.4°F · p5 -1.8°F · p95 +4°F · 67% coverage, followed by a second line reporting how much of that map survives a significance test — 41% of 38214 testable cells reach p ≤ 0.05 (24% reach p ≤ 0.01); median n_eff ≈ 14.2. (Formatting follows the explorer's °F/°C toggle.)
    • The stats line under the legend updates with the displayed field's min/mean/max.

  2. Read it like an oceanographer: the global mean is positive (the ocean is warming), but the map is the real story — warming isn't uniform, and a few regions actually cooled over this window. That's exactly why the engine always offers a map next to a number.

    Now read the second line, which is the difference between a picture and a result. trend fits a slope in every cell whether or not there is evidence for it, so a fair part of any trend map is noise that happens to lean. The n_eff figure is why: sea temperature is persistent, so consecutive monthly frames are not independent observations, and the test discounts the frame count by the residuals' lag-1 autocorrelation before computing p. A cell with 120 frames may carry only ~14 independent samples.

    Set display's stipple param to 0.05 and re-run. Every cell keeps its colour, and the ones whose warming is not distinguishable from zero get a dot screen over them. On a ten-year window that is most of the ocean — and seeing the pattern and its weakness at the same time is the point. The legend caption says which test ran and how many cells it marked.

    There is also trend's significance param, which removes the failing cells instead of marking them. Prefer stipple for anything you are going to look at: a blanked cell is indistinguishable from missing data, and the estimate is gone. Reach for significance when you want the weak cells excluded from something downstream — an areaMean over only the cells that passed, say.

    Neither one marks cells that could not be tested at all (too few frames). Untestable is not insignificant — it is the absence of a verdict, not a verdict of "no effect".

  3. To hand the map back to the normal explorer, just pick any layer or view in the top bar. Your graph stays in the editor.

Share it#

  1. Press 🔗 Copy link. The URL encodes your whole program (?prog=…); anyone opening it gets the panel on the Graph tab with this exact graph loaded and already running.
  2. To keep it locally instead, type a name in the toolbar's preset name box and press 💾 — it appears in the graphs… dropdown under mine, next to the built-in examples.

The finished program#

JSON
{ "nodes": [
  { "id": "layer1", "op": "layer", "params": { "layer": "sst", "start": "2016-01", "stepMonths": 1 } },
  { "id": "trend2", "op": "trend", "inputs": { "value": "layer1" } },
  { "id": "display3", "op": "display", "inputs": { "value": "trend2" }, "params": { "title": "SST trend per decade" } },
  { "id": "answer4", "op": "answer", "inputs": { "value": "trend2" }, "params": { "label": "SST trend per decade" } }
] }

(Your node ids will differ — they're generated as you add nodes and only need to be unique.)

Experiments before moving on: change layer to ice and re-run (the Arctic's trend map is sobering); shorten the window to start: 2022-01 and watch cells go gray — with fewer than 8 monthly frames per cell, trend refuses to fit a line, and honest no-data beats a fake slope.

Next: Correlation Maps — two layers at once, and the difference between temporal and spatial correlation.