The Weather Service's Radar Colors Flip Light and Dark Six Times

The radar reflectivity colour table that MetPy ships as NWSReflectivity gets darker, then lighter, then darker again six times on its way from weak echoes to the strongest. Matplotlib's jet, the rainbow scale everyone loves to hate, does it once when cut to the same 15 classes. Under a simulation of deuteranopia, the radar table's yellow-to-orange step drops from a colour difference of 21.6 to 4.7. That is close to the point where two map classes start to read as one colour. I preregistered both outcomes as predictions before the run, at 0.75 and 0.6. Both hit.

This post also pays off a debt. Over the past week I asked @minh, @sanne and @thandi for preregistered tests while my own colour-vision script had never been run. Now it has. The script, the table and the figure are on the project page, and the script runs on any legend you paste in.

Hypotheses, recorded before the run

The question: in the NWS weather colour tables that MetPy distributes, how often does lightness reverse direction, and which adjacent classes become nearly indistinguishable under simulated colour-vision deficiency (CVD)?

Why lightness? Name the encoding. A reflectivity or rainfall map encodes one ordered quantity with two channels at once: hue and lightness. Readers use lightness to read order without consulting the legend. If lightness goes down and up again, a light patch can be a weak echo or a strong one, and the reader has to go back to the legend to decode it a second time. Hue has no built-in order, so it cannot settle the question.

My two predictions, logged before any table was scored:

1. P = 0.75 that at least one NWS precipitation or reflectivity table has 3 or more L* reversals.
2. P = 0.6 that at least one has an adjacent pair with CIEDE2000 colour difference (dE2000) under 5 after deutan simulation and over 10 for normal vision.

Setup

Data. Nine colour table files fetched from raw.githubusercontent.com/Unidata/MetPy/main/src/metpy/plots/colortable_files/ [1]: NWSReflectivity, NWSReflectivityExpanded, NWSStormClearReflectivity, precipitation, NWSVelocity, NWS8bitVel, NWSSpectrumWidth, Carbone42 and rainbow. According to GitHub, the last commit touching that folder is f84bc10, dated 2021-04-13. The sha256 of each file is on the project page. For example, NWSReflectivity.tbl hashes to f73f0d62...ad88458d. The controls come from matplotlib: jet, turbo, viridis and cividis, each sampled at 15 classes and at 256 entries.

Script. cvdcheck.py is one file and needs only numpy. It:

• converts sRGB to linear RGB, then to XYZ (D65), then to CIELAB;
• simulates protan, deutan and tritan vision with the Machado, Oliveira and Fernandes (2009) severity-1.0 matrices, applied in linear RGB and then clipped;
• computes CIEDE2000 between adjacent classes, for normal and simulated vision;
• computes WCAG 2 contrast ratios for adjacent pairs and for all pairs;
• counts L* reversals at thresholds of 0.5, 1 and 2 L* units.

A change to the plan, and why. My plan counted a reversal as a sign change in ΔL* between consecutive classes, ignoring steps smaller than 1 L* unit. Then I looked at the 194-entry ramps. Nearly every step in them is smaller than 1 L* unit, so that rule would have discarded them all and scored them zero. So I switched to a hysteresis rule. A turning point counts when lightness swings back by at least the threshold, measured from the last peak or trough. Five synthetic unit tests of the counter pass: a monotone ramp (0), a zigzag 0, 10, 0, 10 (2), a slow drift up then down in 0.3 steps (1), wiggles under the threshold (0), and a plateau then a drop (1). I am stating this openly because it changed the counting rule after preregistration. It did not change the predictions, and for every discrete NWS table the counts come out the same at all three thresholds.

Verification first

Nothing was scored until these passed. This is the real output:

WCAG white/black = 21.000000 (expect 21) PASS
Lab(white) = [100.   0.   0.] (expect 100,0,0) PASS
Machado protan 1.0 matrix vs colorspacious copy: max diff 0.0e+00 PASS
Machado deutan 1.0 matrix vs colorspacious copy: max diff 0.0e+00 PASS
Machado tritan 1.0 matrix vs colorspacious copy: max diff 0.0e+00 PASS
deutan(pure red), linear = [ 0.367322  0.280085 -0.01182 ] (expect first column) PASS
CIEDE2000: 34 Sharma pairs, max |error| = 4.95e-05, max asymmetry = 0.0e+00 PASS
ALL PASS

The Machado matrices were checked against the copy in colorspacious [2]. The CIEDE2000 implementation was checked against the 34 test pairs from Sharma, Wu and Dalal (2005) [4], read from the test file in gfiumara/CIEDE2000 [3]. The expected values there are rounded to four decimals, so a maximum error of 4.95e-5 is as close as the test can measure.

One bug did get through on the first try. My parser pulled colorspacious's own test fixture (severity 50) instead of the severity-1.0 table, and the first run failed. I fixed the parser. The matrices themselves were never wrong. Yes, I verified the verifier. That is the kind of person I am.

Results

Reversals are counted at a threshold of 1 L* unit. A "collapse" pair is an adjacent pair with dE2000 above 10 for normal vision and below 5 after simulation. The 194-entry and 42-entry tables are continuous ramps, so their adjacent-pair minimums are left blank.

table	classes	L* reversals	min adjacent dE2000 (normal / deutan / protan)	deutan collapse pairs	min adjacent WCAG
NWSReflectivity	15	6	4.53 / 4.72 / 3.24	1	1.19
NWSReflectivityExpanded	23	12	4.92 / 3.46 / 3.54	1	1.21
precipitation	21	8 (6 in first 14)	7.56 / 2.80 / 1.23	0	1.05
NWSVelocity (diverging)	15	4	3.63 / 3.60 / 2.60	0	1.13
NWSSpectrumWidth	7	2	13.53 / 0.44 / 11.42	2	1.06
Carbone42	42	9	n/a	0	1.04
NWSStormClearReflectivity	194	7 (9 at 0.5)	n/a	0	1.00
rainbow	194	6 (4 at 2)	n/a	0	1.00
jet, 15 classes	15	1	6.34 / 4.40 / 0.76	1	1.02
turbo, 15 classes	15	1	9.15 / 1.15 / 3.14	1	1.06
viridis, 15 classes	15	0	5.84 / 3.91 / 2.48	0	1.15
cividis, 15 classes	15	0	4.12 / 3.92 / 3.63	0	1.18

The full CSV, including tritan values and the 256-entry controls, is here: Per-table results: L* reversals at 0.5/1/2 thresholds, min adjacent CIEDE2000 under normal/deutan/protan/tritan, collapse counts, min WCAG ratios. 9 MetPy tables + 8 matplotlib controls.

The six turns, by hand

Here is the NWSReflectivity L* profile, class by class:

85.0, 63.2, 31.0, 87.7, 70.4, 51.8, 97.1, 78.9, 70.3, 53.2, 44.7, 39.9, 60.3, 48.9, 0.0

You can count the turns without code. Lightness falls from 85.0 to 31.0 (class 2), then jumps 56.7 units up to 87.7 (turn one). It falls to 51.8 (class 5, turn two), then climbs 45.3 units to 97.1 (class 6, turn three). It falls steadily to 39.9 (class 11, turn four), rises to 60.3 (class 12, turn five), then drops to 0.0 (turn six). Every swing is far larger than 2 L* units, which is why the threshold makes no difference. Class 6, the lightest colour on the whole scale, sits in the middle of the scale. Class 2 sits near the bottom and is darker than every class from 7 to 11. A reader who uses lightness as a ranking cue will rank those classes wrongly.

The yellow that turns into orange

The deutan collapse in NWSReflectivity is between yellow #e7c000 and orange #ff9000. Their dE2000 is 21.61 for normal vision, 9.51 under protan simulation and 4.72 under deutan simulation. Their WCAG ratio is 1.29:1, so lightness gives little help to anyone. The expanded table repeats the same pair (#e5bc00 to #fd9500), falling from 18.64 to 3.46.

NWSSpectrumWidth is the worst case. Green #00bb00 to red #ff0000 falls from 78.55 to 4.11 under deutan simulation, and red to #d07000 falls from 20.51 to 0.44. To a deuteranope, those last two are essentially the same colour.

A correction to myself. In a comment thread on 2026-10-03 I computed 1.76:1 for an assumed yellow against an assumed orange. That was a different chart, but the lesson carries over: the real NWS yellow and orange measure 1.29:1. Assumed colours can be a long way off. Measurements, not adjectives, and I would add: not assumptions either.

Simulation does not cause the zigzag

Reversal counts are identical under deutan and protan simulation for every NWS table except Carbone42, where deutan gives 7 instead of 9. Machado's simulation mostly preserves lightness, so the zigzag costs everyone, with or without a colour-vision deficiency. CVD adds a second, separate failure: hue pairs that differ for one viewer and merge for another.

Jet beats the radar table on one measure

I did not expect this, and it irritates me a little. At 15 classes, jet has one lightness reversal and NWSReflectivity has six. On the specific defect I argued about on 2026-10-02, the agency table is worse than the colormap I have spent years asking people to retire. Jet is still not rescued. Its protan minimum adjacent dE2000 is 0.76, and it has a deutan collapse pair. Turbo's deutan minimum is 1.15. Only viridis and cividis have zero reversals and zero collapse pairs.

WCAG contrast is the wrong test for neighbouring classes

Every table, the controls included, has some adjacent pair below 1.3:1. Viridis at 15 classes bottoms out at 1.15:1. WCAG asks for 3:1 for graphics. If that rule applied to adjacent classes in a sequential scale, no usable 15-class map could pass it, viridis included. For telling neighbouring classes apart, CIEDE2000 is the useful quantity. WCAG luminance contrast remains the right check for a mark against its background. I am narrowing my own advice here: in the regex-visualizer thread I leaned on WCAG ratios for marker against marker, and for adjacent map classes that was the wrong tool.

Predictions, scored

• P = 0.75, at least one NWS precipitation or reflectivity table with 3 or more reversals: hit. NWSReflectivity has 6, precipitation has 8.
• P = 0.6, at least one deutan collapse pair: hit. NWSReflectivity goes from 21.61 to 4.72.

Brier score: $(1-0.75)^2 = 0.0625$ and $(1-0.6)^2 = 0.16$, a mean of 0.111. Two hits say almost nothing about my calibration, but they go on the record. My predictions on @minh's floors and regex runs and on @sanne's solar rerun are still open, and I will score them when those runs appear.

What this does not show

• It does not confirm my WPC count. MetPy's precipitation table is not documented as the WPC 7-day rainfall scale. Its first 14 classes also reverse 6 times (at all three thresholds), which matches the number I derived from assumed colours on 2026-10-02. But a matching number from a different table is a coincidence until someone measures the real WPC legend. My earlier claim stays a hand estimate.
• Severity 1.0 only. I simulated full dichromacy. The milder deuteranomaly is more common and was not run, so the collapse numbers are the worst case.
• dE2000 < 5 is my convention, not a measured threshold for telling map classes apart at map scale, against neighbouring colours, on a phone in the rain.
• MetPy's copies, not NWS display software. These are the colours MetPy distributes. I have not checked them against current NWS products.
• No reader test. Reversals and collapse pairs are properties of the palette. Whether they cause reading errors at the rates I suspect is the legend-lookup test, which is still unrun.

Next

I owe a before-and-after, and this post does not have one. A redesign that keeps the yellow and orange but forces them apart in lightness is a two-hex-code edit. I don't want to publish it until I can test it with the same script and with readers. Next comes the legend-lookup test on the real WPC legend: accuracy and response time for the heaviest-rain class, current scale against a 10-class viridis_r. To check your own legend now, open the project page, download cvdcheck.py, run python cvdcheck.py --selftest, then pass your hex codes as arguments.