cartography

Background expertise for maps and figures that survive review.

Projections

• Global fields: Robinson or Equal Earth for presentation (area impressions matter); PlateCarree is a working view, not a publication default, because it exaggerates high latitudes.
• Polar: NorthPolarStereo / SouthPolarStereo with set_extent and a circular boundary path; never show polar fields on PlateCarree.
• Regional midlatitude: LambertConformal centered on the domain.
• Always pass transform=ccrs.PlateCarree() (or the data's actual CRS) to every plotting call on a GeoAxes. Omitting it is the classic silent bug: data draws in projection coordinates and lands in the wrong place or not at all.
• Global longitudes need a cyclic point (cartopy.util.add_cyclic_point) or a seam gap appears at the dateline or prime meridian.
• Coastlines at the resolution the domain needs ('110m' global, '50m' regional); gridline labels on, ticks meaningful.

Colormaps: perceptually uniform, matched to the variable

Rules: sequential colormaps for magnitudes, diverging only when the field has a physically meaningful center (zero anomaly, zero trend, zero divergence), and the center pinned there (vmin=-vmax, or TwoSlopeNorm when the range is asymmetric). Never jet or rainbow; perceptual banding manufactures features that are not in the data.

cmocean table (fall back to viridis/cividis and RdBu_r when cmocean is unavailable):

State units on every colorbar label; a colorbar without units is a finding in review.

Figure sizes

Set the true final size at creation (figsize=), use constrained_layout=True, export vector (PDF/SVG) or >= 300 dpi raster. Do not create a huge figure and shrink it in a document; fonts and line widths shrink with it.

Stippling and significance marking

Mark per-cell statistical significance on trend or composite maps with stippling or hatching over the field:

cf = ax.contourf(lon, lat, slope, levels=levels, cmap=cmap,
                 transform=ccrs.PlateCarree())
ax.contourf(lon, lat, pvals < 0.05, levels=[0.5, 1.5],
            colors="none", hatches=[".."], transform=ccrs.PlateCarree())

• The caption defines the marking, always: which test, which level ("stippling: p < 0.05, Hamed-Rao modified Mann-Kendall"), and what the stippled state is.
• When most of the field is significant, stippling the significant area blackens the map; mark the NON-significant regions instead and say so.
• Per-cell significance is not field significance; basic-statistics carries the FDR rule for map-wide claims.

Multi-panel layouts

• Build with plt.subplots(..., subplot_kw={"projection": proj}, constrained_layout=True); label panels (a), (b), (c).
• Panels that invite comparison share one colorbar: same levels and norm on every panel, one fig.colorbar(mappable, ax=axs) for the set. Per-panel colorbars with different ranges on comparable fields is a review finding; use them only when panels show different variables.
• Seasonal panel sets (DJF/MAM/JJA/SON) follow the QS-DEC convention from basic-statistics; label seasons, not month triplets.

Hovmoller diagrams

Time against longitude (or latitude): pcolormesh or contourf with time on the vertical axis, stated direction (increasing downward matches reading order for forecasts; increasing upward matches time series intuition; pick one and label it), anomalies on a centered diverging map, and the averaged-over band stated in the title (for example "5S-5N mean").

Uncertainty visualization

Per the specification's cartography requirements (docs/SPECIFICATION.md in open-science-pillars/marketplace), this section is load-bearing, not decoration:

• Interval bands on time series: fill_between around the line, the method and level in the label ("95% CI, block bootstrap"), band drawn under the line, alpha kept legible in print.
• Spread maps: an ensemble mean map is accompanied by a spread map (standard deviation or IQR, sequential colormap), same projection and extent, side by side.
• Agreement maps and hatching: for multi-member results, map the fraction of members agreeing on sign; hatch regions of LOW agreement (state the threshold, for example "hatching: fewer than 80% of members agree on sign"). Hatching low agreement keeps the visually quiet regions trustworthy.
• The pairing rule: any interpolated or machine-learning-derived surface presented to stakeholders ships with an adjacent uncertainty map (prediction standard error, cross-validation error, or ensemble spread). A kriged or ML surface without its uncertainty panel is not presentable.
• Every band, hatch, and stipple is defined in the caption with method and level; an undefined marking is worse than none.

Must NOT

• Never use jet or rainbow colormaps, or a diverging map without a physically meaningful, pinned center.
• Never omit transform= on GeoAxes plotting calls.
• Never plot comparable panels with unshared color scales or unlabeled colorbars.
• Never mark stippling, hatching, or bands without defining test, level, and meaning in the caption.
• Never present an interpolated or ML-derived surface to stakeholders without its adjacent uncertainty map.
• Never export print figures below 300 dpi or with fonts under 7 pt at final size.