Visualization brainstorm

One Seismic Month

Sixteen ways to draw thirty days of the USGS earthquake catalog, every one rendered live from the file, ranked, and each carrying the condition under which it lies.

Source USGS Earthquake Hazards Program, past-30-days M2.5+ feed, fetched 2026-08-18. Public domain (US government work).
File earthquakes.csv · 418 KB, 2,237 rows x 22 columns, 2026-07-19 to 2026-08-18
Method Every number on this page was measured from the file.

The datameasured, not assumed

Grain: one catalog event. Each row is one seismic event as the USGS published it: a timestamp, an epicenter, a depth, a magnitude, and the measurement metadata behind them. 2,234 rows are earthquakes; 3 are explosions the network also heard. The grain matters because the catalog records what networks detected, not what the earth did, and several options below turn on that difference.

This is two catalogs stitched together

Regional US networks report down to the M2.5 floor; the rest of the planet enters through the global "us" network at roughly M4.5. Alaska leads every count not because it shakes hardest but because its instruments are closest. Any option that compares counts across regions inherits this bias, and the honest ones say so. A second trap: 660 events sit at exactly 10.0 km depth and 158 at exactly 35.0 km, fixed defaults assigned when depth cannot be resolved, not measurements.

ColumnRoleCard.NullRange / notes
timetemporal2,2370%2026-07-19 21:36 UTC to 2026-08-18 20:25 UTC
latitude, longitudegeographic2,194 / 2,2070%global coverage, latitude -65.1 to 81.5
depthquantitative1,1860%-2.2 to 655.7 km, median 12.4, skew 4.2, 664 rows pinned near 10.0 km
magquantitative1580%M2.45 to M7.7, median M3.7
magTypenominal70%ml, mb, md, mww and 3 more; scales differ, comparisons across them are approximate
netnominal140%reporting network; "us" is global, the rest are regional
placetextual, derived1,7160%generated location strings; region parses out of the last comma segment
typenominal40%2,234 earthquakes, 3 explosion variants
statusnominal20%2,192 reviewed by a human, 45 automatic
nst, gap, dmin, rmsquantitative169 to 1,5440%measurement geometry: station count, azimuthal gap, nearest-station distance, residual
*Error columnsquantitative440 to 1,448≤1%published uncertainties; the only missingness in the file
id, updatedidentifier2,2370%never encoded; id drives hover identity
Correlation structure The strongest correlations run between magnitude and the measurement columns (r = 0.41 with nearest-station distance, -0.47 with azimuthal gap, 0.42 with residual), not between the physical columns: magnitude vs depth is only r = 0.19. Big quakes are measured from far away; that is instrumentation talking, and it previews the completeness split that option A3 draws directly.

Shape signaturewhat it rules in and out

1 temporal x 2 geographic (lat, lon) x 2 quantitative (mag, depth) x nominal (14 networks, 125 regions) : 2,237 rows, grain = one catalog event

Rules IN: the full geographic family (point maps, spike maps, globes, animated spread), time series at daily grain, distributions of magnitude and depth, event sequences, and every cross of those: this is one of the rare shapes where all three of map, timeline and distribution are first-class.

Rules OUT, hard: hierarchy charts (no parent key), network and flow charts (no edge list; quakes do not name each other), choropleths (epicenters are points, and shading countries by a detection-biased count would draw the network map twice), and anything at sub-daily cyclic grain, because the hour-of-day distribution was measured flat.

Conventional4-6

What a competent analyst draws first. One of these five is also the overall recommendation; the other four each surface a constraint the later bands depend on.

C1

Epicenter map

proportional-symbol world map

Where did the earth move this month?

All 2,237 events, sized by magnitude, colored by depth band The Pacific rim carries almost everything. Depth sorts geographically: shallow events (warm) trace every boundary, intermediate ones (olive) hug the subduction arcs, and the 66 deep events (blue) concentrate under Fiji and Tonga where old ocean floor is sinking. Toggling the depth filter shows the deep band is not scattered noise but two tight columns.
Encoding
position = (longitude, latitude) on a Natural Earth projection; area = magnitude (sqrt scale, M2.5 to M7.7); color = depth band at the 70 km and 300 km seismological cuts
Why this data
2,194 distinct latitudes and 2,207 longitudes with 0% missing is real point geography; 2,237 marks is within SVG budget; the depth column's 4.2 skew is why the color is banded rather than ramped
Color
3-class categorical over depth bands, warm for shallow through blue for deep; a continuous ramp would spend its whole range on the 83% of events above 70 km
Interaction
hover identifies any event by place, magnitude and depth; depth-band buttons filter the field to answer "where are just the deep ones"
Exemplar
bubble-map

Failure mode

Overplotting in dense clusters: Alaska and the Ende aftershock zone stack dozens of circles, so the map understates counts exactly where activity is highest. And because the M2.5 floor is only reachable near US networks, the small-dot texture maps instrument coverage, not seismicity.

C2

Daily counts

bar chart over time

Was any day unusually busy?

Events per UTC day Full days run between 47 and 118 with no trend. The mainshock day, Aug 14 (warm), is unremarkable at 62 because the M7.7 struck at 21:58 UTC and its aftershocks landed in the next days' bars. The busiest day, Aug 8 at 118, precedes it entirely and is mostly Alaskan (78 of its events). Daily counting diffuses the month's one real event, which is the case for option A2.
Encoding
x = UTC day (31 bands); y = event count; the mainshock day is the single warm bar
Why this data
2,237 timestamps spanning 31 days aggregate to a readable 31 bars; the mean of 72 events per day is high enough that daily counts are stable, not sparse
Color
single accent hue, one warm highlight on the mainshock day; there is one series, so more color would decorate
Interaction
hover gives the exact date and count; nothing else is needed at 31 marks
Exemplar
bar

Failure mode

The first bar lies: Jul 19 shows 14 events because the 30-day window opens at 21:36 UTC that day, a partial day drawn identically to a full one. And UTC day boundaries split the mainshock from its own aftershocks, making the biggest day look average.

C3

Magnitude histogram

histogram

What sizes do earthquakes come in?

Events per 0.1-magnitude bin, linear and log On the linear axis the ten M6+ events, the only ones anyone felt in the news, do not survive rendering: their bars are under a pixel. The log toggle is the honest view and it reveals the tail is roughly a straight line, which is exactly the observation option A1 formalizes. The two humps, one peaking at M2.5 and one at M4.4, are not physics; they are the two catalogs of option A3.
Encoding
x = magnitude in 0.1 bins (M2.4 to M7.7); y = count, linear or log10 by toggle; M6+ bins flagged warm
Why this data
a single quantitative column with 158 distinct values and a heavy right tail is the textbook histogram case; 0.1 binning matches the precision most magnitudes are published at
Color
accent for the mass, warm for the M6+ bins so the newsworthy tail stays findable at both scales
Interaction
linear/log toggle, because each scale answers a different question (where is the mass vs is the tail lawful); hover gives exact bin counts
Exemplar
histogram

Failure mode

Bin width is an assertion: 0.1 bins expose reporting artifacts (whole and half magnitudes are overrepresented) that 0.5 bins would smooth away. And the linear default actively hides the ten most important rows in the file.

C4

Magnitude vs depth

scatterplot

Do bigger quakes strike deeper?

Every event, depth drawn downward Barely: r = 0.19. The real structure is the empty lower-left corner. Not one event below M4 was located deeper than 242 km, because a small quake at 600 km can only be heard by distant instruments that cannot resolve it. The visible horizontal stripe at 10 km is the fixed default depth, 660 rows thick.
Encoding
x = magnitude; y = depth, inverted so down means deeper, with reference lines at the 70 km and 300 km band cuts
Why this data
two quantitative columns and a measured r = 0.19 worth interrogating; 2,237 points at 40% opacity stay readable because the distribution is sparse away from the shallow band
Color
single accent with opacity doing the density work; color would duplicate what the y axis already encodes
Interaction
hover names each point, which matters here because the outliers (deep M6+, shallow M7.7) are individually newsworthy
Exemplar
scatterplot

Failure mode

The empty corner reads as physics but is detection bias, and nothing in the chart itself says so; it needs its caption to stay honest. The 10 km default stripe likewise invites a false "quakes prefer 10 km" reading.

C5

Region ranking

horizontal bar

Which places shook most often?

Top 12 of 125 regions by event count Alaska's 814 events are 36.4% of the whole catalog, ahead of Indonesia at 190 even in the month Indonesia hosted an M7.7. That inversion is the detection bias made visible: Alaska's dense regional network reports M2.5, while Indonesia enters the catalog only above roughly M4.5. This chart ranks instruments, and saying so is the point of drawing it.
Encoding
length = event count; y = region, parsed from the last comma segment of place, sorted descending; the remaining 113 regions hold 629 events
Why this data
125 nominal values need ranking and truncation, not color; horizontal orientation because "Kermadec Islands region" and "U.S. Virgin Islands" do not fit under vertical bars
Color
one accent hue; the encoding channel is length and the labels are direct, so color has no remaining job
Interaction
hover adds each region's share of the catalog, computed from the data at render time
Exemplar
horizontal-bar

Failure mode

Read as a hazard ranking it is simply wrong: counts measure network density. It also inherits every quirk of the place-string parser, which splits "Fiji", "Fiji region" and "south of the Fiji Islands" into three rows of what is one subduction zone.

Analytical4-8

Each of these earns its chart through a transform: a cumulative count, a time-zero alignment, a split by source, a running sum, a facet, an audit. This band is where the dataset's actual findings live.

A1

Gutenberg-Richter plot

frequency-magnitude, log-linear

How many more small quakes than big ones, and is the ratio lawful?

Cumulative count of events at or above each magnitude, log scale From M4.5 to M6.5 the points fall on a straight line of slope b = 1.20: each full magnitude step down multiplies the count by nearly sixteen. The textbook law, recovered from one month of data. Below M4.5 the curve bends away from the line; that roll-off is not the earth running out of small quakes, it is the global network failing to hear them, and it locates this catalog's completeness threshold precisely.
Encoding
x = magnitude threshold in 0.1 steps; y = log10 of the count of events at or above it (transform: cumulative count from above); dashed fit line over M4.5 to M6.5
Why this data
the magnitude column spans M2.45 to M7.7 with counts spanning 1 to 2,214, four orders of magnitude that only a log axis can hold; the fit range was chosen where the measured curve is straight
Color
accent-2 points, warm fit line: two hues with one job each, subject vs reference
Interaction
hover reads exact N at each threshold, turning the chart into a lookup table for "how rare is an M6"
Exemplar
nz-tourists closest relative; the corpus has no frequency-magnitude chart

Failure mode

The b-value moves with the fit range, and the range is a judgment call; fitting into the sub-M4.5 roll-off would report a biased slope with false confidence. One month of data also puts only 10 points above M6, so the line's upper end is thin.

A2

Aftershock sequence

event timeline with rate bars

What did the M7.7 set off?

The Ende, Indonesia sequence, hours after the mainshock 77 aftershocks within 1.5 degrees of the epicenter in the four days of feed remaining, 21 of them in the first six hours, the largest an M6.1. The six-hour rate bars fall roughly as Omori decay predicts, fast then slow, though with a visible secondary flare-up around hour 36. Aligning time to the mainshock is the transform; calendar time (C2) diffuses this entirely.
Encoding
x = hours since the mainshock (transform: time re-zeroed on the M7.7, spatially filtered to a 1.5-degree box); y = magnitude per event; background bars = count per 6-hour window
Why this data
the mainshock and its aftershocks are individually resolvable rows with full timestamps; 77 points and 15 bars need no aggregation compromise
Color
accent events, one warm mark for the mainshock, low-opacity indigo bars: three layers separated by both hue and weight
Interaction
hover on any event gives magnitude and hours-after; hover on a bar gives the window count, so decay is readable numerically
Exemplar
scatterplot closest relative; the corpus has no aftershock chart

Failure mode

The 1.5-degree box is a modeling choice dressed as a fact: too small drops far aftershocks, too large absorbs background seismicity. And the sequence is right-censored at 94 hours by the feed window, so it cannot show how the decay ends.

A3

Two catalogs in one

split frequency polygons

Is this one catalog or several stitched together?

Magnitude distributions by reporting network Several, and the split is total: 843 regional-network events have median M2.8; the 1,394 global "us" events have median M4.4. The curves cross at M3.9, and each side is blind to the other, with only 4 regional events at M5+ and the global network contributing almost nothing below M4. The bimodal hump in C3 was never about earthquakes.
Encoding
x = magnitude bins (transform: the C3 histogram split by net == "us" vs the 13 regional networks); y = count per bin, one step line per group
Why this data
the net column has 14 values but a measured two-regime structure, so a binary split is the honest cardinality; overlaid outlines beat side-by-side bars for showing a crossover
Color
warm vs accent-2, the diverging pole pair, because the two groups are opponents in this frame; validated CVD-distinct
Interaction
hover on any bin gives both counts, making the crossover point checkable
Exemplar
kde

Failure mode

The binary split hides that "regional" is itself 13 networks with different floors; Alaska's is not Puerto Rico's. Step outlines also visually understate totals compared to filled bars, so the two groups' very different sizes (843 vs 1,394) must be read from the legend, not the ink.

A4

Energy budget

cumulative step chart

When was the month's energy actually released?

Cumulative radiated energy, petajoules The month's 35.2 PJ arrive as a staircase with one giant step: the M7.7 alone releases 63.6% of the total in a single instant on day 26, and the five largest events carry 93.5%. The other 2,232 earthquakes are, energetically, a rounding error drawn as a nearly flat line. Counting events (C2) and weighing them are different charts of the same file, and this one says the month happened on August 14.
Encoding
x = days since the window opened; y = running sum of per-event radiated energy (transform: E = 10^(1.5M + 4.8) joules, then cumulated in event order); warm dots mark the five largest events
Why this data
magnitude is logarithmic, so summing raw magnitudes is meaningless; converting through the standard energy relation is the only way to add earthquakes, and the resulting series is genuinely cumulative
Color
accent line, warm event marks; a single-series chart where the marks are the story
Interaction
hover on each mark names the event and its individual energy in PJ
Exemplar
area

Failure mode

The linear energy axis erases everything below M6: 2,232 events are visually indistinguishable from zero, which is truthful about energy but reads as "nothing happened for 25 days". The energy relation is also an estimate; treating 35.2 PJ as a measurement rather than a model output overstates precision.

A5

Regional small multiples

faceted bar charts

Does every region share the same rhythm?

Daily counts, one panel per top-6 region, independent y scales No. Alaska ticks along near-continuously, a background process that peaks at 78 on Aug 8. Indonesia is flat until the mainshock and then vertical, peaking at 34 the day after: its panel is almost entirely the Ende aftershock sequence. Mexico spikes early (14 on Jul 20), and Japan's flare on Jul 28 is the Kumamoto sequence. One crowded 6-series line chart would show none of this.
Encoding
small multiples: one panel per region (transform: group by parsed region, top 6 by count, then count per UTC day); x = day, y = count with per-panel maximum, labeled
Why this data
the top 6 regions span 814 down to 71 events; a shared y axis would flatten Japan and Mexico to noise, so free scales with printed maxima are the defensible compromise
Color
one indigo hue across all panels; in a small multiple the comparison lives in the facet, and per-panel colors would imply a distinction that does not exist
Interaction
hover on any bar gives region, date and count; panels are deliberately static otherwise so they read as one print-like block
Exemplar
methods-of-comparison

Failure mode

Free y scales are the classic small-multiple trap: Alaska's peak of 78 and Puerto Rico's of 6 render as same-height bars, and a reader who misses the small max labels will equate them. The facet choice (top 6 by count) is also detection-biased, per C5.

A6

Depth audit strip

jittered strip plot

How trustworthy is the depth column?

Every event along a square-root depth axis The column has fingerprints all over it. A solid wall of 660 events sits at exactly 10.0 km, the operator default when depth cannot be resolved, with a second echo at 35.0 km (158 events). Real structure survives around the artifacts: the dense shallow crust band, thinning through the intermediate range, and the isolated deep cluster past 400 km. Any analysis that treats depth as continuous truth inherits that wall.
Encoding
x = depth on a square-root scale (transform: axis chosen to give the 0 to 70 km crust half the width); y = deterministic jitter, meaningless by design; the 10 km default is flagged with a computed count
Why this data
depth has 1,186 distinct values but its top value holds 664 rows, a spike no histogram bin boundary should be trusted to reveal; a strip shows every row and lets the wall be seen as a wall
Color
accent-2 marks with the flag hue reserved for the annotation line, the palette's designated defect color
Interaction
hover identifies individual events, which is how the deep-cluster members can be confirmed as Fiji and Tonga rows
Exemplar
beeswarm

Failure mode

The jitter axis invites false reading: vertical position means nothing, and any perceived vertical pattern is an artifact of the hash. The sqrt axis, chosen to show the crust, correspondingly compresses the deep cluster it flags.

Creative / abstract4-6

Encodings that reward attention. Each one is still decodable, and each card says what it trades away to earn the engagement.

X1

Draggable globe

orthographic projection, interactive

What does the month look like on the actual planet?

Orthographic view, drag to rotate, opening on the western Pacific The flat map's worst distortion is exactly where this dataset lives: the Pacific rim is the edge of an equirectangular world but the center of this one. Opened on Indonesia, the mainshock, the Philippine trench and the Japan arc sit in one gaze, and dragging to the poles shows honestly empty oceans instead of stretched ones.
Encoding
position = (longitude, latitude) on a draggable orthographic projection; area = magnitude; color = the same depth bands as C1, so the two maps cross-reference
Why this data
the events wrap the full globe (latitude -65.1 to 81.5, every longitude), and the Pacific-centered geometry that flat projections cut through the antimeridian is precisely where 66 deep and most large events sit
Color
identical 3-band depth scale to C1: consistency across the two geographic options is itself information
Interaction
drag rotates the sphere with back-hemisphere events culled; hover identifies events. The rotation answers a real question, "what neighbors what", that no single flat frame can
Exemplar
versor-dragging

Failure mode

Half the planet is always hidden, so the globe can never give the one-glance overview C1 does; it is a companion, not a replacement. Readers who do not drag see only the hemisphere chosen for them.

X2

Spike map

spike map

Where were the strong ones, at a glance?

630 events of M4.5+, spike height by magnitude Filtering to the globally complete threshold and encoding magnitude as height instead of area makes the strong events protrude: the Ende M7.7 towers over Indonesia, Colombia's M7.4 stands alone on an otherwise quiet continent, and the Aleutian arc reads as a comb of mid-size spikes. Where C1 shows everything at equal visual rank, this shows the month's skyline.
Encoding
position = epicenter on an equirectangular map (transform: filtered to M4.5+, the measured completeness threshold); spike height = magnitude on a power scale that keeps M4.5 visible and M7.7 dominant
Why this data
the filter is not cosmetic: below M4.5 the catalog is US-only (A1, A3), so an all-events spike map would grow a fake spike forest over Alaska; above it the map is globally fair
Color
single warm hue at half opacity so overlapping spikes darken where arcs are dense, a free density encoding
Interaction
hover names each spike; spikes are drawn north-to-south so southern spikes overlap naturally, as in the corpus exemplar
Exemplar
spike-map

Failure mode

Spikes along east-west arcs (the Aleutians, Indonesia) overlap into shapes whose combined silhouette is unreadable as individual events, and a spike's height is judged against its neighbors, so a lone M5 looks bigger than an M5 beside an M7.

X3

Month replay

animated map with scrubber

How did the month unfold?

Events appearing in time order; recent ones burn warm, then cool into the background Played, the month is a steady global drizzle that suddenly organizes on day 26: Ende ignites and stays hot for the remaining frames while the rest of the world keeps its background patter. The replay makes the distinction between a process (Alaska, constant) and an event (Ende, a moment) visible in a way no static chart on this page manages.
Encoding
position and size as in C1; time = time (transform: events sorted by timestamp, playhead in days); recency = color and opacity, warm for the trailing 36 hours, faded ink for older
Why this data
2,237 timestamped point events are the exact shape animated maps exist for; the corpus precedent animates dated store openings the same way
Color
warm for the live window against desaturated ink for history: one high-salience color reserved for the subject, everything else quiet
Interaction
play/pause and a scrubber, so any moment can be re-examined; the page loads at the fully-drawn end state so the first frame already says something, and reduced-motion readers get that state without playback
Exemplar
walmarts-growth

Failure mode

Animation hides totals: nothing in a moving frame says Alaska logged 814 events, and a screenshot of the page captures only whichever frame was showing. The replay is a narrative device and needs C1 beside it as the ledger.

X4

Horizon rows

horizon chart

Can twelve regional rhythms fit in one screen?

Daily counts for all 12 top regions, three-band horizon, 24 px per row The folded encoding holds twelve regions in the vertical space A5 spends on six. Alaska reads as continuous weather, Indonesia as one dark burst at day 26, Japan as an early flare that fades, and Puerto Rico as a steady low pulse. Darkness means intensity relative to each region's own peak, which is what makes twelve wildly different scales coexist.
Encoding
x = day; per row (transform: daily counts per region, each folded into 3 bands at its own max/3): value = ink darkness via layered opacity; row max printed at the right edge
Why this data
12 series whose daily maxima range from 78 down to 6 cannot share a y axis; the horizon fold is built for exactly this many-series, few-pixels case
Color
one accent hue at three stacked opacities, the standard single-polarity horizon; counts have no negative side, so no diverging pair is spent here
Interaction
hover on any day cell gives the exact count, compensating for the encoding's deliberate imprecision
Exemplar
horizon

Failure mode

Per-row normalization erases absolute scale: Puerto Rico's darkest day is 6 events and Alaska's is 78, drawn with identical ink. The printed maxima are the only defense, and they are 9 px type. Band-folding also makes single-day spikes ambiguous with sustained mid-level activity to an untrained eye.

X5

Depth sounding

cross-section scatter

What does the planet's plumbing look like from the side?

Longitude across, depth downward: the month viewed from the equator Collapsing latitude turns the catalog into a geological section. The crust is a continuous shallow scatter at every longitude, but the deep events hang in discrete columns, the dominant one near 180 degrees where the Tonga-Kermadec slab descends past 600 km, with a second thread under Indonesia around 120 degrees. Those columns are subducting plates, drawn by a month of instrument data.
Encoding
x = longitude (transform: latitude discarded); y = depth downward on a sqrt scale; area = magnitude; color = magnitude on the sequential ramp, drawn small-first so large events stay visible
Why this data
depth is the dataset's third spatial dimension and no map on this page can show it as position; the 66 deep events cluster tightly in longitude, which is what makes the collapsed view coherent rather than mush
Color
sequential ramp over magnitude starting at 30% intensity so M2.5 events stay legible on the surface tint; magnitude is unipolar, so sequential is the honest scale
Interaction
hover gives place, magnitude, depth and longitude, letting a reader confirm which column is which trench
Exemplar
hr-diagram

Failure mode

Collapsing latitude overlays opposite hemispheres: an Alaskan and a Chilean event at the same longitude land in the same column, and the "slab" reading is only valid where the deep cluster is longitudinally isolated. It is a chart for readers who already trust the map above it.

Considered and rejectedwith reasons

OptionWhy not
Radial 24-hour clockMeasured first: hourly counts run 66 to 119 with no structure. Earthquakes do not keep office hours, and UTC hours smear local time across the globe anyway. A cyclic layout would decorate a flat distribution.
Choropleth by countryThe data has no polygons, only points, and shading countries by event count would paint network density as national hazard, the exact lie C5 exists to name. Doubly wrong here.
Calendar heatmapThe window is 31 days. A calendar's payoff is weekday-versus-week structure over months of data; five rows of cells cannot show it, and C2 already carries the daily series.
Force-directed network / arc diagramNo relational column exists: events do not reference each other. Aftershock links could be manufactured by spatial clustering, but that is a modeling project whose output would be drawn, not this file.
Streamgraph of regions over time125 regions with detection-biased counts do not sum to a meaningful whole, and stacking would hide the Indonesia burst behind Alaska's bulk. A5 and X4 answer the same question without the stack.
Word cloud of place namesplace is generated text, not prose; its frequencies are the C5 bar chart with worse geometry. The corpus itself files word clouds under "weak for comparison, say so".
Hexbin density mapBinning 2,237 points would draw sharp density cliffs at network boundaries (dense Alaska hexes beside empty Siberian ones) that encode instrumentation, not seismicity. The unbinned C1 keeps individual events inspectable, which is safer for this catalog.

Notes on the datawhat changes what is drawable

The depth column is part measurement, part convention. 660 rows sit at exactly 10.0 km and 158 at 35.0 km, operator defaults for unresolvable depths; 4 rows are negative (above sea level). Any chart using depth as a continuous variable shows those walls, and A6 exists to keep them from being read as geology.

Counts compare instruments, not ground. The catalog floor is M2.5 only where US regional networks reach; the rest of the world enters near M4.5 (the A1 roll-off and the A3 split both locate it). Every count-based option on this page (C2, C5, A5, X4) is drawn with that caveat in its caption, and X2 filters to M4.5+ to be globally fair.

Magnitudes mix seven scales. ml, mb, md, mww and three others, each valid in a range; treating them as one number is standard practice and approximately right, but 0.1-level comparisons across scales are noise.

The window edges bite. Jul 19 is a partial day (opens 21:36 UTC) and the Ende aftershock sequence is cut off at 94 hours by the feed's trailing edge. The first bar of C2 and the tail of A2 are both censored, not measured lows.

Three rows are not earthquakes (one explosion, one mining explosion, one experimental explosion). They are kept in the maps, where they are real detected events, and are irrelevant at the aggregate scale of every other chart.