Citizen Science on Mars — and the Surface Coverage of Dark Regolith on the Seasonal South Polar Ice Cap

K.-Michael Aye

Freie Universität Berlin

Planetary Sciences and Remote Sensing — Freie Universität Berlin

Tom Ihro

Freie Universität Berlin

Tim Michaels

University of Wisconsin–Madison

Ganna Portyankina

Deutsches Zentrum für Luft- und Raumfahrt (DLR)

Candice J. Hansen

Planetary Science Institute

Megan E. Schwamb

Queen’s University Belfast

04 Jun 2026

Outline

1. Citizen science — what it is, why it works
2. Planet Four — mapping CO₂-jet deposits on Mars with 40 000 volunteers
3. Surface coverage — how much of the seasonal ice is darkened, and why it matters
4. A cautionary tale — how an “inter-annual trend” turned out to be a sampling artifact
5. Where we go next — a parallel albedo channel and an AI to scale it up

~60 s. Set expectations: this is a continuation of the EGU26 talk, but with room to breathe. The middle of the talk — the cautionary tale — is the part I most want the younger scientists in the room to take home. Roughly 45 min, leave time for questions.

I · Citizen Science

What is citizen science, practically?

• Problem: a simple but arduous, repetitive task — too much data to go through yourself
• Solution: split the task into screen-sized subtasks
• Present the data through a simple workflow to thousands of untrained volunteers
• Aggregate many independent judgements into one reliable measurement

~90 s. The core idea in one slide. Emphasize: it only works for tasks that are simple to explain but expensive to do at scale. Pattern recognition on images is the canonical example. The aggregation step is where the science rigor lives.

What’s in it for everyone?

For scientists

• Data volume grows exponentially — humans don’t scale
• Ideal for simple-but-arduous questions
• Labeled data is always better — CS is an efficient label factory for ML
• Genuine public-engagement opportunity

For citizens

• Participate in real research
• Independent of prior education
• A wide variety of fields now offer CS projects
• A core community forms around shared goals

~60 s. Two-sided value proposition. The ML angle is increasingly the strategic one — and it comes back at the end of this talk. Note that a small core of dedicated volunteers typically does most of the work.

Origins — Zooniverse & Galaxy Zoo

• Zooniverse took off with Galaxy Zoo (2007) — classify galaxy morphology
• Dozens of papers before planetary science even joined
• Galaxy shape distributions had to be redefined thanks to volunteer classifications
• The platform now hosts hundreds of projects across every discipline

~60 s. Historical anchor: citizen science at internet scale is ~15 years old and already changed extragalactic astronomy. Planet Four is one of the planetary descendants. Keep it short — it’s context, not the point.

II · Planet Four

Planet Four — citizen science on Mars

• Zooniverse-hosted volunteer mapping; 40 000+ classifiers
• Volunteers outline fans (directional CO₂-jet deposits) and blotches (elliptical) on HiRISE south-polar tiles
• Six Mars Years (MY 28–33) — 469 obs · 64 494 tiles (Aye et al. 2019)
• Each tile seen by ≥ 30 independent volunteers before it counts

~90 s. Anchor numbers: 6 Mars Years, 64 k tiles, 40 k volunteers, ~690 k features. Left image: the P4 marking interface, stacked. The 30-volunteers-per-tile rule is what makes the aggregation statistically meaningful — and what makes progress slow.

Science case — the Kieffer cold-jet model

• Spring sublimation of translucent CO₂ slab ice → high-pressure gas vents through cracks
• Subsurface regolith is ejected onto the bright seasonal ice as fan / blotch deposits
• Fan shape is aligned with the prevailing wind at the moment of eruption (Kieffer 2007; Portyankina et al. 2022)

~90 s. The mechanism. Gas pressure under translucent ice → vent → dust ejected → the fan shape encodes wind direction (and, we hoped, strength). This is why mapping these deposits is scientifically valuable: more wind data than we have ever had on Mars.

What volunteers map → wind data

Full HiRISE colour strip PSP_003092_0985 — fan deposits along the seasonal cap

Zoom — every dark fan is elongated in the same direction; that shared orientation is the wind signal

• Fans: directional — angle, length, spread → a wind vector
• Blotches: elliptical — no clear direction
• A season of fans = wind measurements at hundreds of sites
• “More wind data than we’ve ever had on Mars” — a known point in the season (to ~day precision via the imaging sequence), but never the time of day

~75 s. Make the link explicit: each fan is, in effect, an in-situ wind measurement at the surface. Top: the full HiRISE RGB strip (PSP_003092_0985). Bottom-left: a zoom into the fan field — every dark fan is elongated in the same direction, and that consistent orientation is the wind signal.

State the caveat out loud: we know the season (L_s at the time of imaging), but not the local time of day each jet erupted. Depending on local topography, surface winds at a site can shift through the day, so each fan samples an unknown moment. That ambiguity limits how tightly we can match the fans to GCM / mesoscale wind predictions — and it contributes to why direction matches the models well while inferred strength does not.

Wind directions — sharp and repeatable

Giza fan directions accumulating through the season — one colour per HiRISE observation

• Each observation’s fan directions form a razor-sharp peak (here ~305°, NW)
• Two observations at the same season coincide — L_s 185.5° vs 185.6° lie on top of each other
• The seasonal rotation is orderly, and consistent across the two Mars Years analysed (MY 29–30) (Aye et al. 2019); reproduced by mesoscale models (Portyankina et al. 2022)
• This is the most direct usable product: direction — while area or count are harder to interpret.

~75 s. The evidence behind the “two-tier reliability” claim that pays off on the contrast slide and in the conclusion. Left: an animation of fan-direction distributions at Giza, one colour per HiRISE observation, accumulating as the season (L_s) advances.

Two points: (1) Sharpness & repeatability — each distribution is a tight peak, and the two earliest observations (L_s 185.5 and 185.6) sit exactly on top of each other: independent measurements of the same wind give the same answer. (2) Orderly seasonal change — the peak rotates smoothly from ~305° (NW) early season toward ~0–10° (N) by L_s 224; the wind regime evolves, but predictably.

Aye et al. (2019) found this pattern consistent across the two Mars Years monitored; Portyankina et al. (2022) reproduced it with MRAMS mesoscale winds. So directions are the robust product — which is exactly why I keep area and count separate when I get to coverage.

Active regions around the south pole

• A HiRISE observation = one multi-filter image at one place and time
• A ROI = a region of interest, re-imaged over years — but at very uneven time sampling
• Seasonal campaigns at ~25 ROIs; enormous HiRISE data volume → the need for citizen science
• Remember this map — the uneven sampling across ROIs and years is the punchline later

~75 s. Plant the seed. The ROIs are not sampled uniformly — some have many Mars Years, some only one or two. Tell the audience to hold that thought; it returns in the cautionary tale.

From classifications to a catalog

Clockwise from upper-left: raw tile → fan markings → blotch markings → blotch reduction → fan reduction → catalog entry at 50 % majority vote. Full pipeline in Aye et al. (2019).

~90 s. The reduction pipeline is where rigor is enforced: DBSCAN clustering of fans and blotches separately, then “fnotching” to resolve ambiguous markings by vote ratio, then a cut threshold. Volunteers give raw clicks; the pipeline gives a defensible catalog. White background so the diagram reads.

Experts vs. citizens — the contrast problem

Blotch-area distributions on the same tiles — three science-team experts (GP, MES, KMA) vs the citizen catalog

• Each science-team member (“gold” data) marked several hundred tiles
• Significant disagreement between experts too

• Experience does not overcome the core ambiguity: what contrast defines a feature’s edge?
• Aggregated citizens track the experts — the crowd is not the weak link

What carries error — and what doesn’t: covered area and deposit counts inherit this edge ambiguity, but a fan’s measured wind direction is sharp and reliable.

~75 s. Honest about the limits. The hard problem is not “amateurs vs professionals” — it is that the deposit edge is genuinely fuzzy. Experts disagree with each other. So aggregating many volunteers is a reasonable, well-calibrated estimator, not a compromise.

Drive the framing home before we move to coverage: because the edge is fuzzy, the quantitative products — covered area (→ coverage fraction) and deposit counts — carry real uncertainty. But the directional product does not: a fan’s orientation is unambiguous even when its exact extent is not, so the wind directions we derive are sharp and reliable. Keep this two-tier reliability in mind for the coverage numbers that follow.

Results — the v3.1 catalog, open & reachable

• Over 40 000 volunteers contributed; a core of ~10–20 did most of the work
• 64 494 tiles · 469 HiRISE observations · MY 28–33 · ~25 ROIs
• Open release on Zenodo as part of the Planet Four Data Catalog

Two lines of Python via p4tools

from p4tools import io
fans = io.get_fan_catalog(version="v3.1")

~75 s. The deliverable: a citable, openly redistributable catalog plus a pip-installable access library. This is what makes the coverage analysis in the second half reproducible by anyone in the room. Transition slide into Part III.

III · Surface Coverage

Why surface coverage matters

• Fresh CO₂ ice albedo 0.6–0.8; deposited regolith ≲ 0.2
• Coverage directly modulates the seasonal-cap energy balance, sublimation rate, and jet feedback
• Modellers want a number
• The canonical 20–30 % (Piqueux et al. 2003) is within the active cryptic terrain — not a cap-wide mean
• Planet Four lets us replace the folklore number with a measured distribution

~90 s. The motivation for the whole coverage effort. The widely-quoted 20–30 % is real but local to cryptic terrain; people misuse it as a global average. Our dataset can finally give a defensible distribution. Set up the histogram next.

How we measure coverage

Per-tile fractional coverage (method by Tom Ihro)

1. Each fan/blotch entry → a Shapely polygon in HiRISE pixel coordinates
2. Coverage = union area of all polygons ÷ tile area
3. Aggregate per observation: median, IQR, std across its tiles

• Per-tile table: 64 494 tiles / 469 obs
• ROI labels via p4tools.io.get_region_names()
• Computed in pixel space (not ground-projected) — small effect at similar latitudes
• 10° L_s bins balance time resolution vs. sample size

~90 s. Methodology, kept tight. The union-of-polygons step matters: overlapping fans don’t double-count. Median (not mean) across tiles because the per-tile distribution is heavily skewed. Caveat the pixel-space choice up front so nobody asks later.

Overall coverage statistics

• Cap-wide median 4.88 % · mean 9.32 %
• Heavy right tail past 30 % — but only a small minority of tiles

~90 s. The headline number. Most tiles carry only a few percent; the right tail brushes 30 % but those are rare. The mean/median gap is the signature of the skew — quote the median as the honest central value. The 20–30 % figure as a cap-wide mean is simply wrong.

Inter-annual — the distribution repeats

Distribution shape is strikingly repeatable across MY 28–32. So far, so reassuring…

~90 s. One KDE per Mars Year over the cap-wide background. The eye sees the repeatability immediately. End on a deliberately reassuring note — because the next section is going to complicate it. This is the pivot into the cautionary tale.

IV · A Cautionary Tale

An apparent trend across Mars Years

• Aggregate everything: median coverage across all ROIs, per (MY × L_s) bin
• The most recent Mars Year looked different — a tempting “inter-annual trend”
• It is easy to say here: “coverage trends are changing year over year”

~90 s. This is the trap, presented exactly as we first saw it. We aggregated across ROIs to get one tidy number per year, and the latest Mars Year stood out. The honest first reaction was: a trend! A climate signal! Pause here — let the audience feel the pull of that interpretation before you pull the rug.

The catch — who was actually sampled?

Median coverage per ROI · rows = MY · cols = L_s · ★ = Ithaca (separate scale)

• Each panel is a south-polar map — one dot per ROI, coloured by median coverage
• MY 28–30: richly sampled — 16–25 ROIs each, dots across the whole pole
• This is the regime our intuition is calibrated on…
• …now watch what happens to the sampling in the later Mars Years →

~75 s. Set up the trap with the well-sampled years. MY 28–30 each have 16–25 ROIs spread across the cap — plenty to average over. Tell the audience to hold the picture of “dots everywhere” in mind, because the next slide pulls it away. Same figure, split top/bottom only so the maps are readable.

…and then the sampling collapses

MY 31–33 — the same maps, emptying out

• ROI count falls over the mission: 16 → 25 → 16 → 7 → 6 → 2
• MY 33: almost empty — only 2 ROIs (13 observations), one of them Ithaca, the 10× outlier (★)
• So the MY 33 “all-ROI average” is really “Ithaca and one neighbour” — not comparable to years built from 16–25 regions
• The apparent inter-annual trend was composition, not climate

~120 s. The reveal. The observation-vs-ROI distinction does real work: MY 33 has a respectable 13 HiRISE observations, which feels like enough data — but they fall in only 2 ROIs (Ithaca + Manhattan Classic). The apparent inter-annual change was a composition effect: the cross-ROI average is dominated by which regions happened to be imaged, and MY 33 happens to include Ithaca, the ~10× outlier. Earlier years average 16–25 regions; MY 33 averages 2. Textbook Simpson’s paradox. (Numbers verified from region_names_v3.1.csv.)

The fix — read it within each ROI

Inca City (low, canonical) vs Ithaca (~10× higher). Compare within a region across years — never pool across regions that differ this much.

~90 s. The correct analysis: hold the ROI fixed, look across Mars Years inside it. Each panel highlights one year with IQR error bars; faint grey is the other years. Within a region, coverage is far more stable than the bogus all-ROI “trend” suggested. The between-ROI spread (Inca City vs Ithaca) is the dominant axis of variation in the whole dataset — bigger than any inter-annual effect.

Ithaca up close — why it’s the outlier

HiRISE — Ithaca’s seasonal ice carpeted with large, dramatic fans

Fan directions accumulating through the season — all blown the same way

Ithaca’s ice is carpeted with large fans → ~10× a typical ROI’s coverage. Same physics, extreme expression — keep it separate in any cross-ROI aggregate.

~75 s. Make the outlier physical. The earlier slides said “Ithaca is ~10× typical” as a number on a heatmap — this is what that number looks like on the ground: a HiRISE subframe of Ithaca’s seasonal ice densely covered by large, often overlapping dark fans, each one a CO₂-jet deposit. This is why Ithaca dominates any cross-ROI average and why MY 33 — which only sampled Ithaca and Manhattan Classic — looked anomalous. Same jet mechanism as everywhere else, just an unusually productive site. Right: the fan directions accumulating through the season — all blown the same way.

A short ROI tour

Per-ROI (MY × L_s) heatmaps. Ithaca is the outlier (~10× typical) — it must be treated separately in any cap-wide aggregate.

~75 s. A couple of concrete per-ROI heatmaps so the audience sees real structure: the seasonal rise within a year is the strong, repeatable signal; the year-to-year differences within one ROI are modest. Ithaca’s magnitude is why pooling it with others poisons the average. Optionally flip through a few more live from the review document.

What we did next: match ROI + L_s — but first, the same ground?

Ithaca — every Mars Year’s footprints overlap

Manhattan Classic — same story

All six Mars Years’ footprints pile onto the same patch — MY 33 (yellow) right in the middle. Comparing years within a ROI compares the same ground, not different terrain.

~75 s. The safeguard our internal group insisted on before trusting any within-ROI year-to-year claim: do the HiRISE observations from different Mars Years actually cover the same ground? Footprints come from the RDR index (CORNER lat/lon), coloured by Mars Year. They pile onto the same lat/lon patch at both ROIs, MY 33 (yellow) included. That is the licence to compare MY 33 against MY 28–32 within a ROI: same place, different year. Without this check a “year effect” could just be a “looked-somewhere-else effect.” This is the first of the slides from my internal anomalies deck.

MY 33 — more eruptions everywhere, calm wind only at Ithaca

Ithaca vs Manhattan Classic — fans/obs (blue) & blotches/obs (red) per Mars Year, matched L_s ∈ [180°, 240°]

• Restrict to the same two ROIs MY 33 actually sampled, same L_s — compare like with like
• MY 33 activity ≈ ×2 in both ROIs → more eruption events: plausibly a shared / global driver
• But the fan → blotch flip is only at Ithaca → unusually calm local wind, so ejecta drops as blotches, not wind-shaped fans
• A hypothesis to test, not a cap-wide trend — this is the payoff of stratifying properly

~120 s. The redemption — and the constructive half of the lesson. Having killed the cap-wide artefact, we ask: restrict every Mars Year to the exact two ROIs MY 33 has (Ithaca + Manhattan Classic), in the same L_s window, and compare like with like. Two things emerge: (1) total marking rate ≈ ×2 typical in both ROIs (Ithaca ×2.04, Manhattan ×2.22) — a shared “more eruptions” signal, plausibly global. (2) The fan/blotch ratio collapses (fan_fraction 0.63 → 0.18) only at Ithaca — a local calm-wind anomaly: with little wind to shape them, the ejecta settle as blotches. Manhattan Classic, same MY 33, keeps a normal fan/blotch ratio — confirming the calm-wind part is local, not global. These are genuine, physically-readable hypotheses — exactly what proper stratification buys you. (Source: matched-ROI / matched-L_s re-analysis, v3.1 catalog.)

MY 31 Manhattan Classic — stronger, more directional wind

Manhattan Classic fan geometry by Mars Year — MY 31 fans are longer and narrower

• MY 31 fans ~24 % longer and ~26 % narrower than the MY 28–32 cluster
• That year’s activity sits at the low end → fewer eruptions, each under stronger, more directionally-focused wind
• Another single-(MY × ROI) signal — local, physical, testable

~90 s. Second matched-ROI finding. 6 obsids in matched L_s ∈ [180°, 240°]; fan distance ×1.24, opening angle ×0.74 vs the MY 28–32 cluster. Total marking rate at the low end of typical (217 markings/obs) — so fewer events, but each captures stronger and more directionally stable surface wind during MY 31’s fan-forming episodes at Manhattan Classic. A clean, opposite-sign companion to the MY 33 Ithaca calm-wind story.

MY 32 Ithaca — a coverage shape-shift (suggestive)

Ithaca per-tile coverage distribution by Mars Year — MY 32 (red ★) piles up at low coverage

• MY 32’s per-tile coverage piles up at low values — median 7.5 % vs the typical 14–18 %
• Fan geometry stays normal (distance, spread, fan-fraction all typical) → a tile-coverage shape shift, not a fan effect
• Only 3 observations → flagged suggestive: a target for follow-up once more MY 32 Ithaca data arrive

~75 s. The fourth — and weakest — signature, included for honesty. 3 obsids only (N = 272 tiles), so suggestive, not firm. The per-tile coverage distribution shifts left: median 7.5 % vs typical 13.7–18.4 %; only ~38 % of tiles reach ≥ 10 % coverage versus the typical 60–89 %. Crucially the fan-side numbers (distance, spread, fan-fraction) are all typical — so this is a change in how coverage is distributed across tiles, not a fan-geometry effect. Flagged as a study target once more MY 32 Ithaca observations exist. Naming a weak signal honestly, with its caveat, models exactly the practice this talk is about.

Four signatures, four readings

MY × ROI	n	Signature	Physical reading
MY 31 Manhattan Classic	6	fans ×1.24 longer · ×0.74 narrower · low activity	stronger, more directional wind
MY 33 Ithaca	5	activity ×2.04 · fan-fraction 0.63 → 0.18	more eruptions (shared) + calm local wind → blotch surge
MY 33 Manhattan Classic	5	activity ×2.22 · fan/blotch ratio normal	more eruptions (shared) + normal wind → calm wind is local to Ithaca
MY 32 Ithaca	3	tile-coverage shifted low	suggestive — needs more observations

MY 33 across both ROIs: the ×2 activity is shared (a candidate global driver); the fan→blotch flip is local to Ithaca (a local wind anomaly). Every one of these stays hidden until ROI and L_s are matched.

~90 s. The synthesis from the internal deck. All four signatures are confined to a single (MY × ROI) cell — none extend cap-wide, which is exactly why the naive all-ROI average manufactured a false trend. Two are robust (n = 5–6 observations); MY 32 Ithaca is flagged suggestive only (n = 3). The point for the talk: admitting the EGU artefact did not leave us empty-handed — proper stratification turned one fake trend into four concrete, physically-readable hypotheses worth chasing. That is progress, bought by a mistake — which sets up the closing slide.

The lesson — heterogeneous datasets

An average over a heterogeneous, unevenly-sampled population is not a measurement — it is an artifact of the sampling.

• Check the n behind every aggregate cell before you read a trend into it
• A change in a pooled statistic can be pure composition change (Simpson’s paradox)
• Stratify first: verify the signal within each stratum, in every stratum
• Only aggregate across strata you have shown to be comparable

~120 s. The take-home for the younger scientists. Say it plainly: we almost reported an inter-annual trend that was really “we imaged different places that year.” The discipline that saves you is boring — always look at the per-stratum sample size, always check whether the same subgroups are present before and after. This habit generalizes far beyond Mars. This is the slide I most want them to remember.

V · Where We Go Next

Off-active darkening — a parallel channel

• Beyond discrete fans/blotches, the background CO₂ ice darkens diffusely
• Mechanism: a thin homogeneous dust mantle from lighter grains carried by wider jet activity (Pommerol et al. 2011)
• Photometric quantification (north cap) by (Portyankina et al. 2013); recent N–S synthesis (Hansen et al. 2024)
• Planet Four does not capture this — it is a parallel contribution to the seasonal albedo budget

~75 s. Important caveat for any modeller using our numbers: P4 measures the discrete deposits, not the diffuse darkening. A complete albedo budget needs both channels. Be explicit so nobody over-reads the coverage fraction as the total darkening.

Conclusions

• Two-tier reliability: wind directions are sharp & reliable; coverage area and deposit counts carry real uncertainty
• v3.1 catalog is openly redistributable (Zenodo + p4tools)
• Cap-wide coverage: median 4.88 %, heavy right tail — not the folklore 20–30 %
• Within-ROI distributions repeat across MY 28–33
• Beware the all-ROI average — uneven ROI sampling fakes inter-annual trends
• Combine with off-active darkening for a complete albedo budget
• Now training an AI on P4 labels to extend coverage from the colour strip to the 5× larger RED channel

~90 s. The slide they photograph. Hit the three results, the one methodological warning (our cautionary tale, now distilled to a bullet), and the forward look. The ML model trained on Planet Four labels will let us run this coverage analysis on the much wider RED-channel HiRISE data — roughly 5× the area of the colour strip.

Verbal acknowledgments: 40 000+ Planet Four citizen scientists; HiRISE / LPL; Zooniverse; co-authors (Tom Ihro, Ganna Portyankina, Candice Hansen, Megan Schwamb, Tim Michaels); funding sources. Background is PSP_003092_0985 at 0.5 opacity over dark blue.

One last thing — for the students

• Mistakes are the essence of progress — every scientific advance is built on them
• We just dress them up: “insufficiencies,” “a hole in the inference logic,” “limitations of prior work” — masking words for the very same thing
• Strip the disguise and it is always a mistake, small or large, that brought us forward — this talk is proof: our MY 33 “trend” was wrong, and being wrong opened the real science
• So don’t be afraid of them. Run into them with open arms, screaming. A mistake will always teach you where to go next.

The real closing message — aimed straight at the younger scientists in the room. Deliver it slowly. Tie it to the talk: the entire second half exists because we first got the MY 33 story wrong and were willing to say so out loud. “Insufficiency”, “gap in the inference”, “limitation” — every paper’s polite vocabulary for a mistake that taught the field something. The fear of being wrong is the single biggest brake on a young researcher; give them permission to let it go. End on “where to go next”, then thank the 40 000+ volunteers and co-authors.

References

Aye, K-Michael, Megan E Schwamb, Ganna Portyankina, et al. 2019. “Planet Four: Probing Springtime Winds on Mars by Mapping the Southern Polar CO2 Jet Deposits.” Icarus 319 (February): 558–98. https://doi.org/10.1016/j.icarus.2018.08.018.

Hansen, Candice J., Shane Byrne, Wendy M. Calvin, et al. 2024. “A Comparison of CO\(_2\) Seasonal Activity in Mars’s Northern and Southern Hemispheres.” Icarus 419: 115801. https://doi.org/10.1016/j.icarus.2023.115801.

Kieffer, Hugh H. 2007. “Cold Jets in the Martian Polar Caps.” J. Geophys. Res. Planets 112 (E8): E08005. https://doi.org/10.1029/2006JE002816.

Piqueux, Sylvain, Shane Byrne, and Mark I. Richardson. 2003. “Sublimation of Mars’s Southern Seasonal CO\(_2\) Ice Cap and the Formation of Spiders.” J. Geophys. Res. Planets 108 (E8): 5084. https://doi.org/10.1029/2002JE002007.

Pommerol, Antoine, Ganna Portyankina, Nicolas Thomas, et al. 2011. “Evolution of South Seasonal Cap During Martian Spring: Insights from High-Resolution Observations by HiRISE and CRISM on Mars Reconnaissance Orbiter.” J. Geophys. Res. Planets 116 (E08007). https://doi.org/10.1029/2010JE003790.

Portyankina, Ganna, Timothy I Michaels, Klaus-Michael Aye, Megan E Schwamb, Candice J Hansen, and Chris J Lintott. 2022. “Planet Four: Derived South Polar Martian Winds Interpreted Using Mesoscale Modeling.” Planet. Sci. J. 3 (2): 31. https://doi.org/10.3847/PSJ/ac3087.

Portyankina, Ganna, Antoine Pommerol, K.-Michael Aye, Candice J. Hansen, and Nicolas Thomas. 2013. “Observations of the Northern Seasonal Polar Cap on Mars II: HiRISE Photometric Analysis of Evolution of Northern Polar Dunes in Spring.” Icarus 225: 898–910. https://doi.org/10.1016/j.icarus.2012.10.017.