TBB cluster explorer

The geostationary meteorological satellite that made these observations. It holds a fixed longitude over the equator and rescans East Asia on a repeating schedule, which is why one location can be followed through a whole day at 10-minute steps.

The instrument aboard it. AHI scans in several spectral bands; this demo uses one of them.

The infrared channel used here, centred at 3.83 µm.

What the detector actually measures: energy arriving per unit area, solid angle and wavelength. Brightness temperature is calculated from it.

The temperature an ideal black body (an object whose glow depends only on its temperature) would need to emit the measured radiance in this band. Over opaque cloud it describes the cloud top; over clear sky it is influenced by the surface. Across this table it runs 258.3–321.6 K.

The preprocessing that produced the matrix: keep the 39 × 121 grid cells over the coastal box, keep every scan the archive holds for the day, and stack them into one table. That table is the design matrix.

One cell of the latitude–longitude grid, 0.01° by 0.01°. There are 4 719 of them and each one is exactly one row of X.

One complete image of the box at one instant. 142 of them are present; a complete day would hold 144.

A scheduled slot in which the full-disk scan is suspended, so no image exists. 2 such slots fall inside this day, which is why the columns are not evenly spaced and why every column carries its own timestamp.

Japan Standard Time, nine hours ahead of UTC. Every clock label on this page is JST; the tables also print the UTC timestamp the payload actually stores. The scans on this page are 2024-03-03 09:00 → 2024-03-04 08:50 JST, which is the single UTC day 2024-03-03.

The 4 719 × 142 table of brightness temperatures. Rows are samples (locations), columns are features (scans). There is no target column.

X[i,:] is one location's day; X[:,j] is one scene; x_ij is one temperature. The three colours in the Design matrix section always appear beside that notation, so the encoding survives without colour.

The payload stores each cell as an integer of the source value × 100 to keep the file small. The page divides it back out and shows the value in kelvin (K).

Learning from inputs with no answer key. Nothing is trained to predict a known answer, so results are checked for internal consistency and against independent information, such as the map.

Principal component analysis: rotates the data onto new axes ordered by how much of the variation each one captures; this page draws the first two. Distances and directions in a PCA picture are meaningful, and the first two components hold 97.6 % of this matrix's variance.

A nonlinear method that keeps each point's nearest neighbours close together. Here it is fitted on the first thirty principal components, not on the raw columns.

The t-SNE setting for roughly how many neighbours each point takes into account. 5 values are precomputed: 5, 15, 30, 50, 100.

Another nonlinear neighbourhood method, fitted here on the raw matrix. It tends to produce tighter, more separated groups than t-SNE from the same data.

The UMAP parameter setting the size of the neighbourhood it treats as local. 5 values are precomputed: 5, 15, 30, 50, 100. UMAP's other setting, min_dist = 0.5, is held fixed across all of them; it controls how tightly points are allowed to pack together in the embedding, not what is neighbouring what.

A 2-D coordinate pair per location produced by one of those methods. All 11 embeddings on this page were computed offline; the browser only reads them.

Partitions points into k groups by repeatedly assigning each point to the nearest group centre and recomputing the centres. It is the only fit that runs in the browser, using k-means++ initialisation and a fixed seed so the same k always gives the same picture.

The number of clusters, set by the reader between 2 and 40.

After K-means, groups are renumbered C1…Ck by ascending mean TBB, so C1 is always the coldest group and the blue-to-red cluster colours mean the same thing in every panel and in every method. The convention is the original notebook's, which also sorts its clusters by ascending cluster mean.

A number comparing two partitions of the same points: 1 when they agree completely, 0 at the agreement chance alone would produce. The Conclusions section uses it to compare the three embeddings' clusterings.

The share of side-by-side grid cells that fall in the same cluster. It is the page's own measure of how spatially coherent a partition is, compared against the chance level implied by the cluster sizes.

The cluster colours run blue → red, from the coldest group C1 to the warmest group Ck, because the underlying quantity is ordered and what a reader is actually comparing is hot against cold. The same scale carries the clusters everywhere they appear: the three embedding scatters, the chips, the map, the curve panel and every table swatch. Neighbouring clusters can look alike; the tables beside each coloured panel give the exact values.

A second, one-ended blue scale, used only for the scan image in the Design matrix section. It encodes a brightness temperature in Kelvin rather than a cluster, which is why it does not share the clusters' blue-to-red scale.

The colour range used by the scan image in the Design matrix section: the 2nd to 98th percentile of every cell in the matrix (270.3–300.3 K), computed once and reused for every timestamp, with the few cells beyond either end clamped. It exists because the extremes of the matrix are rare enough to flatten every individual scene if the ramp is stretched to them.

Every colour-encoded panel on this page ships a table of the same numbers, behind a disclosure or beside the chart. Colour is the summary; the table is the evidence.

On the curve panel, the range containing the middle 80 % of the selected locations at each moment, showing how much they differ from one another.

A static OpenStreetMap image, fetched once at build time, cropped to the data box and embedded in this file. It serves only as a backdrop for judging the clusters.

Clicking a dot, a cluster button or a point on the map isolates one cluster across the embeddings, the map and the curves at once. Escape, or “All clusters”, clears it.