Help and methods · 10

Ocean regimes

Ocean regimes, under Explore, divides the sea surface into areas whose temperature statistics are alike and ranks those areas by how distinct they are. It answers a question the daily maps cannot: not what the sea surface looked like on a day, but where it behaves differently over three decades. The result is a classification of the archive, computed once on the server; nothing is fitted to the day on the map.

Using the tool

Press Ocean regimes under Explore. The regimes draw over the map in their colours and the card opens with the tree, the table and the controls. Drag the handle at the left of the card header, or use the arrow keys on it, to move the card.
Choose the number of regimes, 2 to 12. The dashed line on the tree shows where the cut falls, and the sentence beneath the selectors reports how much of the variance between blocks the cut removes.
Hover a row of the table to outline that regime on the map. Analyse opens the regime’s largest polygon as an area card with the full method menu, as for any other selection.
Within restricts the classification to one area: any of the seven LMEs: Benguela Current, Agulhas Current, Canary Current, Guinea Current, Somali Coastal Current, Red Sea or Mediterranean Sea; any open area, coastal buffer or polygon card; or anything saved under Regions & selections, including shapefiles, GeoJSON and WKT brought in through Import geometry (also offered beside the selector). An area needs at least 20 ocean blocks of 0.25°, about 14,000 km²; smaller areas are listed as too small, and a point has no regimes. Inside an area the statistics are standardised over that area alone, so its regimes describe its internal contrasts.
The downloads give the polygons as GeoJSON and the regime properties as JSON for the cut shown.

The Ocean regimes card with the Benguela Current LME chosen under Within: the tree coloured by regime, the cut at six regimes and the ranked table, with the six regimes drawn along the Namibian and South African west coast. — **Regimes inside the Benguela Current LME.** Six regimes, each branch of the tree below the cut in the colour of its regime on the map and in the table.
1Within and Regimes selectors.
2The Ward tree, cut at six regimes.
3Regimes ranked by distinctness, with the features that set each apart.

The classification covers the expanded SST domain, 26°W–60°E and 50°S–51°N. Each LME has a separately computed classification using its original boundary and holes. Selection is by block centre, so the displayed boundary follows the 0.25° grid. Numbers and colours belong to the selected classification: regime 1 in one LME does not mean the same regime as regime 1 in another. Requests extending beyond the published grid are refused.

How the classification works

The domain is divided into 0.25° blocks of five by five archive cells. A block counts as ocean when at least half its cells are ocean on 90 % of the days of the period, which is the 1991 to 2020 baseline. Each ocean block is described by fifteen statistics of its daily SST series over that period:

Climatology. The mean SST and the two coefficients (cosine and sine) of a single annual harmonic fitted by least squares to the daily series. The annual cycle enters as two linear coefficients rather than as amplitude and phase, because phase is an angle and cannot be averaged or standardised as a number.
Variability. The standard deviation of the daily residuals from that harmonic, and the standard deviation of the complete-year annual means.
Change. The trend of the annual means in °C per decade, from the same AR(1) fit as the trend layer, over all complete years.
Fronts. The magnitude of the gradient of mean SST, in °C per 100 km.
Heatwaves and cold-spells. Days per year, events per decade, mean duration and mean category, from the heatwave3 category fields, for marine heatwaves and for cold-spells separately.

Each statistic is standardised across blocks to zero mean and unit variance, then weighted so that each of the six groups above contributes equally to the distance between two blocks; otherwise the eight event metrics would outvote the single mean. In that space the blocks are joined by Ward’s method, which at every step merges the two groups whose union adds the least within-group variance. For a domain or area with more than 6,000 complete blocks, k-means first reduces the blocks to 2,000 micro-clusters, and the tree is built on those. Smaller areas are linked directly. The card reports the block count and the method used for the chosen scope. Cutting the tree at n gives n regimes, and the cuts are nested: the regimes at six are splits of the regimes at five.

The dendrogram is the top of that tree, truncated to forty leaves. Height is the Ward linkage distance, proportional to the square root of the variance a merge adds; a long vertical stem above a group means the group stays separate over a wide range of cuts. The branches below the cut line take the colour of their regime, so the map, the tree and the table share one palette. Variance removed is the share of the total variance between blocks that lies between regimes rather than within them; it rises with every added regime, and the shape of that rise, not any threshold, is the guide to how many regimes the area supports.

Distinctness is the distance from a regime’s centroid to the nearest other centroid divided by the root-mean-square distance of its blocks from their own centroid. A value above one means the regime is further from its nearest neighbour than its own blocks are from each other. The three features listed as distinguished by are those whose standardised centroid values depart furthest from zero, with the sign of the departure. Polygons are the dissolved blocks of each regime; enclaves smaller than four blocks are left out and counted, except in selected areas with fewer than 400 blocks, where single-block parts are retained.

Reading the result

The result depends on the feature set, the standardisation and the weights; all are recorded in the product manifest, and none is offered as a setting, because changing them changes the regimes. Spatial contiguity is not enforced: a regime may occupy both basins where their statistics agree. Trend and interannual variability use all complete years (1991–2025 in the expanded release), a longer window than the other statistics. Method version 2 corrects the previous trend-property conversion: values are in °C per decade and are one tenth of the erroneously multiplied version 1 values. This unit correction cancels under standardisation; expanding the domain changes the classification itself. Ward on micro-clusters moves boundaries slightly relative to Ward on every block. The regimes are statistical summaries of OSTIA at 0.25°; they are neither ecoregions nor management units, and a regime boundary is not an observation of a front.

The full definition, with its parameters and limits, is scientific decision record 0013 in the repository; the processing runs on the server with a fixed seed and is reproducible from the archive.