Introduction

First observed in Viking images in 19771,2, slope streak occurrence and formation have been subject to a continuous debate ever since. Dark slope streaks form sporadically and fade over years and decades (~3 to ~20 Mars years, MY)3,4, occasionally turning brighter than the surrounding regolith3,5. Slope streak populations have been observed to occur in dusty, low thermal inertia equatorial regions6,7. In stark contrast, recurring slope lineae (RSL), first discovered in 2011, are predominantly located in rocky, southern, high thermal inertia highlands8,9. The formation of streaks and RSL has been explained by dry and wet mechanisms, such as airfall- or thermal cycling-induced dust avalanching and impact-, quake-, dust devil-, or rockfall-induced mass/dust movements (dry)10,11,12,13,14,15,16,17,18, as well as groundwater spring-, frost-, and brine-induced flows (wet)7,9,19,20,21,22,23. Recently4, and24 identified an enhanced streak formation rate in northern autumn and winter (Ls ~140°–270°). Similarly, RSL have been observed to predominantly form in southern summer (Ls ~270° 8). To date, it remains unclear whether slope streaks and RSL are different expressions of the same process, or fundamentally different features (e.g. refs. 5,25). If confirmed, a wet formation mechanism for streaks or RSL would indicate that present-day Mars’ hydrological cycle is substantially more pronounced than previously anticipated, with substantial implications on our understanding of martian weather, climate, surface evolution, and habitability. In addition, liquid water or brines on Mars’ surface would evoke serious concerns about planetary protection. Here, we fuse several decades’ worth of orbital data and use a global-scale deep learning- and data-driven approach to create the first global catalog of slope streaks, shedding new light on a fundamental, decades-old debate.

Results

Martian streak global geostatistics

We identify a total of 13,026 bright and 484,019 dark slope streaks with a wide range of morphologies in the pre-v01 version of the global MRO CTX (Mars Reconnaissance Orbiter Context Imager) mosaic26,27 (‘detected population’, \({{{{\rm{p}}}}}_{{{{\rm{D}}}}}\)), including previously unknown slope streak-bearing regions (Fig. 1; Supplementary Figs. S1, S2). There is no clear, universally applicable, quantitative way to separate ‘dark’ from ‘bright’ streaks, especially in CTX data, as the transition between dark and bright is continuous. Here, ‘dark’ streaks include recently-formed streaks that appear dark in CTX images, and ‘bright’ streaks refer to old, partially-faded streaks that appear bright in CTX images (Fig. 1A, B). The deep learning-driven detector is able to identify about 20% and 55% of all bright and dark streaks in a given CTX image, on average; at the same time, we estimate that about 6 % of all detections are duplicates caused by the spatial overlap of some of the CTX images that make up the mosaic (Supplementary Fig. S3), suggesting that the actual number of bright and dark streaks observable on the surface of Mars are ~60,000 and ~830,000 (at the CTX-scale, ~6 m, ‘statistically-corrected population’, \({{{{\rm{p}}}}}_{{{{\rm{C}}}}}\)). We note that the deep learning-predicted detections contain an average of 2.33 individual streaks per detection box due to streak co-location and overlap (see Fig. 1A, B, Supplementary Fig. S1), indicating there might be up to ~140,000 and ~1,900,000 individual, CTX-scale bright and dark slope streaks on Mars, which is several factors more than was previously estimated (~800,000 dark streaks28). Applying the power law relation between the number of resolved streaks and image resolution as established by Bergonio et al.29, we estimate there might be ~14 times more streaks resolvable at the HiRISE-scale (High Resolution Imaging Science Experiment, ~0.3 m30), i.e., ~2,000,000 bright and ~28,000,000 dark streaks.

Fig. 1: Global distribution of slope streaks and RSL on Mars.
figure 1

A through C: CaSSIS (Colour and Stereo Surface Imaging System89) and HiRISE color-infrared images of bright and dark slope streaks, as well as RSL; downslope direction is always pointing down- or side-wards; image IDs MY36_014887_165_0, MY36_014975_168_0, MY35_009504_159_0, MY35_010269_009_0, ESP_022689_1380, ESP_034830_1670. Raw image credits ESA/TGO/CaSSIS CC-BY-SA 3.0 IGO and NASA/MRO/UoA. D) through F): Maps of slope streak (bright, white; dark, black) and RSL (red) distribution (e.g. refs. 8,51) as well as bright-to-dark slope streak ratio per 10° by 10° quadrangle, overlain on a Viking merged color mosaic and MOLA topography81. Percentages of total population per streak hotspot indicated on the map. Our results agree well with earlier maps of streak-bearing regions3,5, while identifying a series of previously unknown streak hotspots, e.g. in Xanthe Terra and Lunae Planum (Supplementary Fig. S2).

Full size image

We identify streaks between ~40°N and ~20°S, with a mean latitude of 12.7°N and 16.7°N for bright and dark streaks, respectively, indicating that dark streaks are systematically shifted northwards relative to bright streaks. Notably, ~89% and ~90% of all bright and dark streaks are located north of the equator, and there are a factor ~37 more dark than bright slope streaks. Bright streak hotspots (>100 features per 5° x 5° quadrangle) are located in Arabia Terra - Terra Sabaea, E Kasei Valles, Ius Chasma, and Noctis Labyrinthus, where the Arabia Terra – Terra Sabaea hotspot is the most pronounced. Dark streak hotspots (>5000 features per 5° × 5° quadrangle) are located in Olympus Rupes - Lycus Sulci, Ceraunius Fossae, Arabia Terra – Terra Sabaea, Tartarus Montes, and NW Medusae Fossae – Marte Vallis, where the Olympus Rupes – Lycus Sulci hotspot is the most pronounced (holding ~68% of the global dark streak population). Only 15% of all 10 × 10° quadrangles across the surface host more bright than dark streaks (Fig. 1). Those regions might have experienced reduced streak formation rates in recent years; we note a particularly substantial percentage of bright streaks in Kasei Valles and Nili Fossae. The vast majority (~85%) of all bright streaks are located on Noachian terranes, only ~4% are located on Amazonian terranes. In turn, about ~50% of all dark streaks are located on Amazonian terranes, and only 31% on Noachian terranes (Fig. 2A).

Fig. 2: Global streak and RSL geostatistics.
figure 2

A Pie charts of terrane age distribution of bright streaks, dark streaks, and RSL; co-located bar chart indicates their distribution over the northern (black) and southern (gray) hemispheres82. B Histograms for slope streak (bright, tan; dark, solid black), RSL (red), and background (black line) distributions over a range of thermophysical properties31,81,83). Note the blue line at TES albedo 0.26, as spatially visualized in Fig. 3. Average values indicated. Note that the relatively low spatial resolution of most of the global thermophysical datasets might blur certain relations, most prominently for the MOLA slope angle dataset, which largely underestimates the actual slope angle at the streak-scale. C AFDs of estimated feature area; \(\surd 2\) binning is used, power law fits indicated by lines, power law exponents indicated by numbers. Note that the lack of small slope streaks is caused by the spatial resolution of the used CTX images (6 m); RSL are only resolvable at the HiRISE scale (0.3 m).

Full size image

Generally, bright and dark streaks are located at a) below-average topographic elevations, with average values of −126 m (bright) and −1126 m (dark); b) above-average slope angles with average values of 7° (bright) and 8.6° (dark); c) in above-average TES (Thermal Emission Spectrometer31) albedo terrain with average values of 0.27 (bright) and 0.28 (dark); and d) in above-average OMEGA (Observatoire pour la Minéralogie, l’Eau, les Glaces et l’Activité32) nanophase ferric oxide terrain with average values of 1.02 (bright and dark) (Figs. 2B, 3A). Globally, we observe a weak power law relation between the density of slope streaks and the mean slope angle (Supplementary Fig. S4). Dark slope streaks do not appear to feature a preferred slope orientation, but equator-facing slopes host slightly more dark streaks (global difference to pole-facing slopes is 4.1%, i.e., ~20,000 streaks, Fig. 2B); bright streaks feature a distinct west-facing orientation. We do not observe any substantial differences when considering northern and southern hemisphere populations separately.

Fig. 3: Global relation between streaks, RSL, and dust.
figure 3

A Map of slope streak (bright, white; dark, black) and RSL (red) distribution, overlain on TES albedo (<0.2 transparent, 0.2–0.26 gray, 0.26–0.32 blue31), highlighting several spatially coinciding, distinct cut-offs of slope streak and high albedo terrain (>0.26), such as in southern Arabia Terra, north-west Elysium Planitia, southern Amazonis, and north-western Tharsis. This implies that streaks can only form in the highest albedo (smallest particle size) terrains. Viking merged color mosaic in the background. B Joyplots (showing kernel density estimates) of slope streak (bright, tan; dark, solid black), RSL (red), and background (black line) MCD daily mean dust deposition rate on a horizontal surface for Ls 0°, 90°, 180°, and 270° (left34) and sensible surface heat flux for Ls 90° (northern summer) at 12, 3, 6, and 9 AM (right34). Average values indicated.

Full size image

Bright and dark streaks feature a strong correlation between estimated size, MOLA elevation (Mars Orbiter Laser Altimeter33), MCD diurnal surface temperature amplitude (Mars Climate Database34), albedo, and nanophase ferric iron: larger streaks appear to feature lower MOLA elevation and higher diurnal temperature amplitude, albedo, and nanophase ferric iron values (Supplementary Fig. S5). On average, bright streaks are larger than dark streaks, with an average estimated size of ~775 vs ~609 meters. We do not observe any general, longitudinal or latitudinal variations in streak size. Streak length and width are weakly linearly related; assuming a rectangular streak shape, we estimate a mean surface area for bright and dark streaks of ~0.047 and 0.034 km2, respectively. The area-frequency-distribution (AFD) of estimated streak area follows a power law with a negative slope of 3.9 and 3.8, respectively. Interestingly, an AFD derived for RSL presents a negative exponent of 1.1 (Fig. 2B). There is no correlation between streak size and the age of the host terrane (Supplementary Fig. S6).

Using the mean number of streaks per detection (2.33), we estimate that one bright or dark slope streak detection modifies a surface area of ~0.1 km2, on average. This indicates that the detected slope streaks modify an area of 49,705–103,500 km2 globally, i.e., 0.03–0.07% of the martian surface (for \({{{{\rm{p}}}}}_{{{{\rm{D}}}}}\) and \({{{{\rm{p}}}}}_{{{{\rm{C}}}}}\), respectively). Considering a streak survival time of 3–20 MY (where streaks are observable3,4), the global surface modification rate would be 15,568 to 2485 km2/MY and 34,500 to 5175 km2/MY, i.e., 0.01–0.002% and 0.02–0.004% of the surface per MY (for \({{{{\rm{p}}}}}_{{{{\rm{D}}}}}\) and \({{{{\rm{p}}}}}_{{{{\rm{C}}}}}\), respectively). Considering a streak thickness of ~1 m (i.e., the maximum value derived by14,15, an average streak detection would modify a volume of ~100,000 m3 or ~0.0001 km3, which would result in a total modified volume of ~50 to ~100 km3 (for \({{{{\rm{p}}}}}_{{{{\rm{D}}}}}\) and \({{{{\rm{p}}}}}_{{{{\rm{C}}}}}\), respectively). Considering a dust bulk density of ~2500 kg/m3 or ~2.5 × 109 t/km335, the total modified mass would account for ~1.3 × 1011 to ~2.5 × 1011 t, i.e., ~4.2 × 1010 to ~6.3 × 109 t/MY and ~8.3 × 1010 to ~1.3 × 1010 t/MY (for \({{{{\rm{p}}}}}_{{{{\rm{D}}}}}\) and \({{{{\rm{p}}}}}_{{{{\rm{C}}}}}\), respectively). References 36,37 report that around 4.0 × 108 to 2.9 × 109 t/MY of dust is exchanged between the martian surface and atmosphere, i.e., between ~1 to ~46% and ~0.5 to ~22% of the dust mass modified by streaks (for \({{{{\rm{p}}}}}_{{{{\rm{D}}}}}\) and \({{{{\rm{p}}}}}_{{{{\rm{C}}}}}\), respectively). Reference 38 reports a total dust mass of ~4.3 × 108 t lifted during a global dust storm observed in 1977, which is ~1 to ~7% and ~0.5 to ~3% of the estimated mass modified by slope streaks per year (for \({{{{\rm{p}}}}}_{{{{\rm{D}}}}}\) and \({{{{\rm{p}}}}}_{{{{\rm{C}}}}}\), respectively). These estimates indicate that slope streaks move several dust-equivalents of a global dust storm every MY. The injection of dust into the atmosphere by slope streaks has not been observed yet, but would make slope streaks an important contributor to the martian dust cycle and climate, potentially accounting for a substantial fraction of the surface-atmosphere dust exchange budget.

Endo- and exogenic drivers of streak formation

We observe that dark streaks and date-constrained impact events39 are partially co-located in dusty regions (Fig. 4A). To suppress the observational bias related to the detection of new impacts in dusty regions (e.g. refs. 39,40) we perform a statistical analysis of streak and new impact locations in the dust-rich Arabia Terra, Elysium Planitia, and Tharsis regions, which host both streaks and new impacts. We find that dark streaks tend to be located slightly closer to real impact sites than to randomly placed (synthetic) impact sites in Arabia Terra (statistically significant, P = 0.038) and Elysium Planitia (not significant, P = 0.231), on average, but not in Tharsis. In Arabia and Elysium, in circles with radii of 0.5, 2, and 5°, there are a factor ~1.5, ~1.3, ~1.1, as well as ~1.9, ~1.6, and ~1.3 more streaks around new impacts vs synthetic impacts, on average (Fig. 4B, see Supplementary Table S1), decreasing with increasing distance from a given impact. Beyond Arabia Terra and Elysium, we do not see a global relation between the spatial density of streaks and new impacts (Supplementary Fig. S7).

Fig. 4: Global relation between streaks, RSL, tectonic activity, and meteoritic impact flux.
figure 4

A Map of slope streak (bright, white; dark, black) and RSL (red) distribution as well as new impacts (blue stars39). InSight-recorded marsquakes (LF and BB events with BAZ, yellow diamonds42), and tectonic features (white lines82); TES albedo and Viking merged color mosaic in the background31. Red dashed lines outline the three areas used for the statistical analysis of streak/RSL and impact/quake proximity (Arabia Terra, Tharsis, and Elysium Planitia). B Joyplots (showing kernel density estimates) of minimum distance between streaks (bright, tan; dark, solid black) and real/synthetic impact/seismic events for Arabia Terra, Elysium Planitia, and Tharsis; differences between distance dark streaks – real impacts and dark streaks – synthetic points are statistically significant for Arabia Terra, but not for Elysium and Tharsis. Note that there are no InSight-recorded marsquakes located in Arabia Terra and only one in Tharsis (per41). We note that we do not consider the relation in Tharsis representative due to the availability of just one single data point (InSight event S0183a). Average values indicated. C Pie charts visualize the ratio between the number of streaks (b = bright, d = dark) within a 0.5°, 2°, or 5° area around real (tan/black) vs synthetic (brown) impact events/epicenters. RSL statistics not shown as they are not deemed statistically sound (see Supplementary Table S1).

Full size image

Some of the global streak hotspots are spatially co-located with InSight-recorded, tecto-genetic broadband (BB) and low-frequency (LF) events with back azimuth (BAZ)41,42,43,44, but a) not all streak hotspots feature a co-located epicenter and b) not each epicenter features a streak hotspot (Fig. 4A). We perform a statistical analysis of the two global streak hotspots with associated seismic BB or LF event(s) (Elysium Planitia and Tharsis). Dark streaks do not tend to be located closer to real quake epicenters than to randomly placed (synthetic) epicenters (no statistically significant difference, P = 0.167). In Elysium, in circles with radii of 0.5, 2, and 5°, there are a factor ~1.4, ~1,4, and ~1.0 less streaks around real epicenters vs synthetic epicenters, on average (Fig. 4B, see Supplementary Table S1). We note that a substantial fraction of the CTX images used by the global mosaic were acquired prior to the InSight mission, rendering our analysis not ultimately conclusive. We do not observe a general, global relation between the distribution of streaks and the geomorphic expressions of past tectonic activity, such as graben and fossae, despite those features’ steep slopes, but identify noteworthy exceptions such as Ceraunius, Tractus, Noctis, Ulysses, Elysium, and Labeatis Fossae (Fig. 4A). Similarly, we do not observe a general, global relation between rockfall locations45 and slope streak distribution other than a small number of co-located features in the Cerberus Fossae and Ius Chasma areas (Fig. 5).

Fig. 5: Global map of slope streak, RSL, rockfall, and valley network distribution.
figure 5

Map of slope streak (bright, white; dark, black) and RSL (red) distribution, overlain on a Viking merged color mosaic. Blue lines indicate valley networks as mapped by Hynek et al.47; yellow shapes represent confirmed rockfall locations mapped by Bickel et al.45.

Full size image

Slope streaks tend to be located in regions with above-average MCD wind speed, particularly in the morning (9 AM, statistically significant, P < 2.2 × 10−233, Fig. 6). We observe that slope streaks and active dust devils46 are strongly inversely correlated, i.e., regions with dust devil activity are nearly entirely devoid of slope streaks. Notably, regions of active dust devil occurrence appear to coincide with several of the globally distinct boundaries of slope streak distribution, such as located in NW Lycus Sulci, Syria Planum, N Phlegra Dorsa, E Terra Sabaea, and S Arabia Terra (Fig. 6A). Slope streaks are strongly associated with above-average MCD diurnal temperature amplitude regions, particularly pronounced in northern spring, summer, and autumn (Fig. 7). Slope streaks tend to be located in regions with above-average MCD atmosphere-to-surface heat flux during the night (12 and 3 AM), where the correlation disappears in the morning (6 AM) (Fig. 3B).

Fig. 6: Global relation between streaks, RSL, and atmospheric processes.
figure 6

A Map of slope streak (bright, white; dark, black) and RSL (red) distribution as well as CTX-observed dust devils (brown triangles46, covering MY28 to MY36); MCD surface wind speed in the background34. B Joyplot (showing kernel density estimates) of slope streak (bright, tan; dark, solid black), RSL (red), and background (black line) MCD surface wind speed for 9 AM, 12 PM, and 3 PM LST for Ls 90° (left), 180° (center), and 270° (right). Differences between dark streaks and background are statistically significant for all Ls and LST. Average values indicated.

Full size image
Fig. 7: Global relation between streaks, RSL, and the thermal environment.
figure 7

A Map of slope streak (bright, white; dark, black) and RSL (red) distribution; MCD thermal amplitude (noon-midnight) at Ls 90° in the background34. B Joyplot (showing kernel density estimates) of slope streak (bright, tan; dark, solid black), RSL (red), and background (black line) MCD diurnal temperature amplitude (left) and diurnal noon temperature (right) for Ls 0°, 90°, 180°, and 270°. The temperature range between 273 and 283 K that (theoretically) provides conditions for liquid water on the surface of Mars (in some regions) is indicated in blue. Average values indicated.

Full size image

We observe that slope streaks appear to be located in regions with slightly above-average WEH (water equivalent hydrogen, statistically significant, P = ~0), H2O, and H, particularly bright slope streaks. Notably, not all high-H2O/H/WEH regions feature slope streaks and many streak hotspots are located in low H2O/H/WEH regions (Fig. 8; e.g., in Elysium, 10°N 170°E, and E Tharsis, 20°N 110°W). Slope streak locations feature slightly enhanced water vapor column values in northern summer and autumn (Fig. 8B). The vast majority/parts of the slope streak population pass through the ~273 to ~283 K temperature band that theoretically allows for transient liquid surface water, in northern summer and southern summer, respectively (northern summer is particularly pronounced, Fig. 7B). Streak locations do not show any substantial overlap with the valley networks mapped by Hynek et al.47, expect for a few locations in Naktong, Deva, Minio, Padus, and Sabis Vallis (Fig. 5A). We note that dark streaks are located in regions that see above-average dust deposition rates in southern summer (statistically significant, P = ~0), but below-average rates for the rest of the year (Fig. 3B).

Fig. 8: Global relation between streaks, RSL, and volatiles.
figure 8

A Map of slope streak (bright, white; dark, black) and RSL (red) distribution; FREND (Fine-Resolution Epithermal Neutron Detector84) water equivalent hydrogen (WEH) map and Viking merged color mosaic in the background. B Joyplots (showing kernel density estimates) of slope streak (bright, tan; dark, solid black), RSL (red), and background (black line) FREND WEH, GRS (Gamma Ray Spectrometer85) H2O, and GRS H values (left) and MCD water vapor column for Ls 0°, 90°, 180°, and 270° at 12 PM LST34 (right). Average values indicated. Note that the locations of bright and dark streaks tend to feature slightly higher FREND WEH values than GRS H values; note that GRS H and H2O are nearly (but not entirely) identical.

Full size image

Discussion

Broader implications for streak formation and Mars’ present-day habitability

Our observations discard three previously proposed dry streak formation mechanisms: dust devils (e.g. refs. 11,48), rockfalls (e.g. ref. 15), and thermal cycling (e.g. ref. 25) do not appear to play a globally important role in triggering slope streaks (Figs. 5–7). Dust devil, rockfall, and slope streak occurrence are predominantly inversely correlated, and slope streak populations experience higher-than-average diurnal temperature amplitudes around the year, not just during streak peak formation season in southern summer (Ls ~270°); in fact, the southern summer peak formation season experiences the least pronounced temperature amplitude difference (Fig. 7B).

Our observations question many of the hypothesized wet formation mechanisms: streak populations a) do not express any preferred slope orientation (Fig. 2B), challenging a potential CO2 frost related trigger21; b) only experience conditions suitable for liquid water in northern summer (Ls ~90°) – despite being located in regions that experience above-average diurnal thermal amplitudes and heat flux –, which conflicts with observations of peak streak formation occurring in northern autumn and winter (Ls ~140–270° 4,24); in fact, streak populations experience unfavorable conditions at Ls ~ 270° (statistically significant, P = 5.8 × 10−158) (Figs. 3B and 7B); and c) streak populations experience peak water vapor column values in northern summer and autumn (Ls ~90–180°), which also conflicts with observations of enhanced streak formation in northern autumn and winter (Ls ~140–270° 4,24) (Fig. 8B). We observe that streak populations feature slightly enhanced WEH, H2O, and H values at the global-scale (Fig. 8B), but note that recent, local-scale investigations using CRISM (Compact Reconnaissance Imaging Spectrometer for Mars) have not found any signs of hydration at slope streak locations49.

We identify three global-scale, statistically significant relations that support dry formation hypotheses for slope streaks: streak populations a) tend to be slightly closer located to new impacts than to the same number of randomly placed locations in some regions (Fig. 4B); b) consistently experience above-average (near-surface) wind velocities (Fig. 6B); and c) experience above-average dust deposition rates in northern winter (Ls ~270°) (Fig. 3B), coinciding with the seasonal formation peak4,24). Notably, we identify a distinct cascade of slope streaks on two specific slopes in the direct vicinity of a 140-m diameter crater that formed on September 18th 2022 (Ls 305°, during peak dust deposition, Supplementary Fig. S8, refs. 90,91) as well as new slope streaks close to other new impact events (ref. 92, Supplementary Figs. S9–S12), but also find slopes that do not host new streaks following close-by impact events (Supplementary Figs. S12–S14), which illustrates how impacts can trigger streaks, but only in sloped, dust-bearing terrain. We do not observe any relation between streak occurrence and tectonic marsquakes, noting that this particular comparison suffers from a substantial observational (spatiotemporal) bias due to the limited seismic data available (~2 MY). However, we identify a number of Fossae with an increased number of slope streaks (Fig. 4A), which might point to recent tectonic activity in those regions12. The lack of streaks in the majority of Fossae might also be rooted in a general lack of (fine) dust in those regions.

In contrast, RSL populations are preferably located on equator-facing slopes (Fig. 2B8,50) and experience conditions suitable for liquid water in southern summer (Ls ~270°), i.e., during their peak occurrence (Fig. 7B); however, RSL populations simultaneously feature below-average diurnal thermal amplitude, heat flux, and WEH, H2O, and H values, as well as average water vapor column values, rendering a wet formation mechanism less likely (Figs. 3B, 7B and 8B). We note that RSL locations experience peak dust deposition rates in northern autumn (Ls ~180°), i.e., at the onset of RSL peak formation, and dust storms have been shown to substantially enhance RSL formation rates51. We do not observe any global-scale relation between RSL locations, new impacts, and wind velocity (Figs. 4 and 6B), but identify an indistinct, spatial relation with dust devil occurrence, and – potentially – rockfall occurrence (Figs. 5 and 6A), as previously locally observed by, e.g. refs. 51,52. Here, dust devils and rockfalls might (jointly) trigger the displacement of material through turbulence and the transfer of kinetic energy (e.g. refs. 11,53). Recently, seasonal peak dust devil activity (number and size) has been observed in the southern highlands in southern summer (Fig. 6A46), coinciding with peak RSL formation. We note that one of the main limitations of this study is the lack of information on the timing of global-scale streak and RSL formation. Globally-distributed information about slope streak and RSL formation would allow for a more robust statistical comparison of streak and RSL occurrence and potential drivers.

Overall, our observations suggest that slope streak and RSL formation may be predominantly controlled by two independent, dry drivers, 1) the seasonal delivery of dust onto topographic inclines, and 2) the spontaneous activation of accumulated dust by energetic triggers – wind and impacts for slope streaks, as well as dust devils and rockfalls for RSL. These observations showcase how the seasonal dust delivery and the seasonal formation of streaks and RSL are inherently linked, while alleviating the likelihood of wet formation as proposed by preceding studies (e.g. refs. 16,17,50,51,52). The important role of fine dust in streak formation is further highlighted by the abundance of streaks in the highest-TES albedo terrain (Fig. 3A), the positive correlation between slope streak size and dust abundance (Figs. 2B, 3, Supplementary Fig. S5), and the lack of a relation between streak size and host terrane geologic age (Supplementary Fig. S6). Notably, ref. 54 report how variations in the particle size could explain albedo differences between high-albedo (e.g., Arabia) and intermediate-albedo terrain (e.g., Utopia). Per55, the dust in high-albedo terrain features a particle size less than 40 µm; recent work by Valantinas et al.56 suggest even smaller particle sizes, i.e., as small as 1 µm, which agrees with earlier observations of atmospheric dust57. Such small-sized particles could facilitate the formation of some of the observed complex, flow-like features of slope streaks, such as fingering and anastomosis (e.g. refs. 25,58,59,60). In addition, slope streak formation may be influenced by slow moving dense granular flows that are supported by dispersive pressure60, potentially explaining why slope streaks do not produce observable dust clouds during their formation. It remains unclear what amount of the dust displaced by slope streaks ultimately ends up being injected into the atmosphere.

Our results underline the fundamental differences between slope streaks and RSL, despite their visual resemblance. Streak and RSL populations occur on opposite hemispheres (north vs south), at different topographic elevations (mostly lowlands vs mostly highlands), in opposite thermal inertia terrain (low vs high), in different wind speed regimes (above-average vs below-average), in dissimilar diurnal thermal amplitude and heat flux terrain (above-average vs average), in different WEH, H2O, H, and water vapor column terrain (average vs below-average), and in terrain that provides suitable (theoretical) conditions for liquid water at different seasons (Ls ~90° vs Ls ~ 270°) (Figs. 1–8). Importantly, the exponents of the AFDs (streaks: ~3.8; RSL: ~1.1) bracket the exponents commonly experienced for landslides on Earth (~1.5 to ~361,62,63); we note that AFD exponents cannot be directly used to infer triggers or causes of mass wasting events, but the substantially different exponents add further weight to the conclusion that streaks and RSL are fundamentally different features and might point to different material properties (e.g., dusty, cohesion- vs sandy, friction-controlled material, e.g. refs. 18,63 (Fig. 2B).

Our findings suggest that martian slopes currently do not experience seasonal, transient flows of liquid water or brines, underscoring the dry, desert-like nature of Mars. This implies that slope streak and RSL locations are not likely to be habitable, alleviating strict planetary protection measures for future landed missions to those regions.

Methods

Few-shot learning-driven slope streak mapping

We select 40 globally-distributed MRO (Mars Reconnaissance Orbiter) CTX (Context Imager) images with a pixel scale of ~6 m64 and crop them into ~1000 × 1000 pixel patches. With those patches we create a training dataset by labeling a total of 1007 and 147 dark and bright slope streaks, following previously established workflows65,66,67,68. Here, a label represents a rectangular box drawn by a human expert around a slope streak. We label an additional 35 dark and 6 bright slope streaks for model testing and verification (~4% of the training set size; remains unseen during training). During labeling, we carefully separate bright and dark slope streaks following the definition provided in the main text and showcased in Fig. 1. We purposely minimize the size of the testing dataset, to maximize the data available for model training. Here, the usage of the term ‘few-shot learning’ accounts for the fact that the number of slope streak labels available for training and testing is lower than in traditional deep learning which commonly uses tens of thousands of labels. Importantly, past studies have demonstrated that a few-shot learning approach can successfully map geomorphic and atmospheric features across various planetary bodies and datasets45,46,69,70,71,72,73. A list of all used CTX images in shown in Supplementary Table S2 and their illumination and imaging geometries are listed in Supplementary Table S3.

We use the training dataset to fine-tune a COCO-pretrained (‘Common Objects in COntext’), off-the-shelf PyTorch 1.7 implementation of the YOLOv5x convolutional neural network (CNN) provided by ultralytics via GitHub (https://github.com/ultralytics/yolov5). YOLOv5 is a well-established, verified, cutting-edge object detection framework that has been utilized for a wide range of scientific applications, such as food science, robotics, medical image processing, and planetary science (e.g. refs. 71,73,74,75,76). YOLOv5 represents the best compromise between state-of-the-art performance and reliability, and has been shown to substantially outperform (performance and speed) earlier machine learning algorithms such as based on AdaBoost, Histogram of Oriented Gradients, and Region-based CNNs (e.g. refs. 77,78,79) as were used by earlier studies that reported on the automated detection of slope streaks80.

We apply substantial label augmentation during training to prevent model overfitting and minimize any negative impact of the relatively small number of labels, as is common practice in computer vision and few shot learning, utilizing label rotation, up-down flipping, shearing, up- and down-scaling, and brightness and contrast modifications. We do not apply left-right flipping to preserve a proper representation of the illumination conditions governed by MRO’s sun-synchronous orbit in the training data. The training of the CNN required ~15 min on one single NVIDIA RTX 3090 GPU (Graphical Processing Unit).

We evaluate the performance of the trained detector using the testing dataset and the three most common metrics used for object detector evaluation: recall (i.e., the number of detected streaks, e.g. 8/10), precision (i.e., the number of correctly detected streaks, e.g. 9/10), and average precision (AP, relation of recall and precision over a wide range of model confidence scores, i.e., the area under the precision-recall curve, which summarizes the performance of a detector in one single number, Fig. 9C). The CNN assigns a confidence score to each detection (0, low to 1, high), representing its confidence in the assigned detection and classification; higher confidence score populations usually feature lower recall but higher precision. At confidence 0.8, the trained slope streak detector achieves full precision (100%), a recall of 20 % and 55 %, as well as 0.98 and 0.50 AP (0.74 mAP, mean average precision for all classes) for bright and dark streaks in the testset, respectively (Fig. 9C). This indicates that the detector can be expected to make none to very few mistakes, while being able to (consistently) identify about 20 and 55% of all bright and dark slope streaks observed by CTX, at a confidence score of 0.8 and higher.

Fig. 9: Visualization of the deep learning-driven mapping workflow.
figure 9

A MRO CTX images are cropped into small patches and fed into the CNN, which (B) outputs predictions of slope streak location, size, and confidence score (true positives, TP, white boxes). The detector can miss slope streaks (false negatives, FN, yellow) or can make incorrect predictions (false positives, FP, red); note how larger streaks feature larger bounding boxes. C Testset performance of the detector is visualized on the left, indicating recall, precision, and AP, as well as the score cut-off of 0.8 used for this study. Image ID B22_018396_1890_XI_09N320W. Raw image credits to NASA/MSSS/ASU.

Full size image

At a score of 0.6 our detector outperforms (for dark streaks only, recall ~85%) the performance as reported by earlier studies that used a combination of feature selection and machine learning (binary) classification techniques and HiRISE images (ref. 80, recall ~79%), at a continuously high level of precision (~94%). However, given the substantial number of expected slope streak candidates we deliberately choose a confidence score of 0.8 for this study, to maximize the precision of the detector and minimize the amount of time required for the subsequent review efforts, as described below.

We deploy the trained detector in a customized image download and processing pipeline originally established by Bickel et al.69 and further refined by Bickel et al.70,72, and Conway et al.46. The pipeline successively downloads all of the 86,546 CTX images contained in the post-beta27, pre-v01 version of the global CTX mosaic (pers. comm. J. Dickson), which was later officially published in ref. 26, at https://image.mars.asu.edu/stream/, and tiles them into 1000 × 1000 pixel patches (Fig. 9A). At the time of data processing, the final version of the26 mosaic was not yet available, which required us to process the individual CTX images that made up the pre-v01 version of the mosaic instead. All patches are scanned by the streak detector, the geographic location of all detections are reconstructed using the image’s metadata, and all detections as well as their associated metadata are stored (including a size estimate, i.e., the diameter of the bounding box), including so-called ‘thumbnails’, i.e., crop-outs of all CNN-predicted bounding boxes that are used for the subsequent expert review (see Fig. 9B, Supplementary Fig. S1). We use YOLOv5’s default Non-Maximum-Suppression (NMS) algorithm to remove duplicate predictions in each patch, following computer vision best practice. Overall, the processing of the entire dataset requires ~2 months of processing time using one single NVIDIA RTX 3090 located in a regular desktop computer.

Due to the potential for geo-location and co-registration inaccuracies as well as partial overlap of the CTX images that make up the pre-v01 version of the global CTX mosaic (pers. comm. J. Dickson) there is a small chance that slope streaks located along the seams of individual images are detected multiple times (‘duplicates’). However, a case study in the Tharsis region suggests that the number of potential duplicate detections in the overall dataset is expected to be small, around 6 % of the total streak count (Supplementary Fig. S3). The slight location inaccuracies paired with the large number of slope streak detections impedes an automated and manual detection and removal of duplicates, which is why they are not specifically identified and removed. We note that duplicate detections are evenly distributed across the mosaic and globe and, thus, do not systematically affect the geostatistical science analysis.

After processing, the detector identified 661,144 slope streak candidates (40,855 bright, 620,289 dark) at a confidence score of 0.8 and higher. At confidence 0.7, 0.6, and 0.5 the detector identified 1,494,416, 2,015,709, and 2,463,542 slope streak candidates, respectively.

Candidate review

We note a number of false detections among the bright and dark slope streak detection candidates, which closely resemble the visual appearance of streaks, such as wind streaks, linear shadowed topography, dunes, and mountain ridges (Supplementary Fig. S15). Past experience with object detectors being applied to very large and heterogeneous orbital image datasets has shown that the testset performance is not always fully representative of the ‘real world’ performance of the detector in the full dataset, due to the inherently limited representability of the strictly size-constrained training and testing datasets as well as unknown/unexpected anomalies such as related to imaging artefacts, weather-related phenomena (clouds, etc.), and a-priori unknown look-alikes (see, e.g. refs. 45,70,71,72). The increased number of false detections as achieved by the bright streak detector (Fig. 9C) can likely be attributed to a) the smaller training dataset, which in turn is caused by the relative scarcity of bright streaks (as noted by other authors, such as, e.g. refs. 3,5), and b) the more subtle radiometric contrast between bright streaks and the surrounding slope (compared to dark streaks). The presence of false positives in the output of both detectors underlines the importance of the manual review of all candidates and the removal of all false positives.

In order to remove all obvious false detections and ensure the integrity of the scientific analysis, we manually review all CNN-identified candidates. The review consists of a human expert looking at every single slope streak candidate thumbnail, one-by-one. Due to the overwhelming amount of candidates at lower scores (e.g., >1 million candidates are score 0.7) we decide to only review all candidates with score 0.8 or higher, i.e., all very high confidence detections (n = 661,144). Overall, the manual review of all candidates required ~two weeks, achieving a reviewing rate of ~183 detections/minute or ~3 detections/second, assuming an 8-hour workday and 5 workdays per week. We identify a total of 13,026 bright and 484,019 dark slope streaks, indicating a detector precision of ~32% (bright) and ~78% (dark) (Fig. 9C). The testset precision at score 0.8 is 100% for both types, indicating there is a drop of performance by 68% for bright and by 22 % for dark streaks, as was expected based on previous experience. After expert review, we conclude the validation of the dataset by examining all remaining slope streak detections as plotted on the26,27 global CTX mosaic, considering all detections in their geomorphic context and removing any remaining outliers. All analyses and results (as well as the associated dataset) utilize the fully reviewed streak catalog.

Verification of the reviewed mapping results

The quality and reliability of the produced global slope streak dataset rely on the a) combined quality of the used CTX images, b) performance of the slope streak detector, and c) quality of the manual expert review. We do not find any direct or indirect evidence for any negative effect of the scanned CTX images on the mapping outputs, such as potentially caused by cloud cover, shadowing, haze, and data artifacts. We point out that26,27 carefully selected the images that make up the global mosaic to minimize any such negative impacts. The detector was carefully trained and its performance thoroughly tested and assessed using state-of-the-art approaches and heterogeneous, geographically-distributed training data. We do not expect any observational bias for the dark slope streak class; the very wide range of slope streak morphologies identified by the detector (Supplementary Fig. S1) provides evidence for a robust and capable detector.

Notably, the preferably more equatorward distribution and W-facing orientation of bright slope streaks is likely to be attributed to a bright slope streak detector that is more successful in retrieving bright streaks on well-illuminated, Sun-facing slopes, given the very faint appearance of bright slope streaks (Fig. 1A, C) and MRO’s sun-synchronous orbit (imaging takes place in the early local afternoon). Potentially, the higher mean length as reported for bright streaks could be attributed to the lower performance of the bright streak detector (compared to the performance of the dark streak detector) resulting in systematically less precisely placed bounding boxes, leading to larger size estimates. We point out that the evenly-distributed orientation of dark streaks (Fig. 2B) implies that there is no general, CTX imaging-related bias regarding slope shadowing, as further verified by qualitative testing – and as expected based on the careful selection of mosaic images performed by Dickson et al.26,27.

We expect the manual review to capture and remove the overwhelming majority of all false detections contained in the original candidate dataset. In confirmation of that expectation, our derived streak distribution maps agree exceedingly well with previously published maps of slope-bearing terrain (e.g., refs. 3,5, Supplementary Fig. S2), which further increases our confidence in the derived maps and relations. Notably, we identify a number of previously unknown streak hotspots, such as located in Xanthe Terra, Lunae Planum, E Kasei Valles, and Syrtis Major (Fig. 1, Supplementary Fig. S2). We note that our mapping approach is relying on a stitched mosaic consisting of CTX images taken several years apart (2006–202026,27), and at different times/seasons; as a result, our slope streak catalog does not represent one snapshot in time, but a wide time-range. Much of the remaining ambiguity could be addressed with globally-distributed information about slope streak and RSL formation times.

Streak number and size estimate correction

The CNN reports a size estimate for every slope streak detection, representing the diameter of the CNN-predicted bounding box, which closely encompasses the streaks, i.e., represents a proxy for streak length (see Supplementary Fig. S1). Using the spatial resolution of the associated CTX image, we convert the diameter from pixel to meter. Following a previously established methodology65,66 and in order to improve the accuracy of the relation between bounding box diameter and actual streak length, we manually measure the actual length of a randomly drawn subsample of 1000 slope streaks and relate it to the respective bounding box diameter (linear relation), deriving the following relation:

$${{length}}_{{actual}}=0.8774*{{length}}_{{CNN}}-61.503$$
(1)

We use this relation and correct all ~500,000 streak length estimates; all scientific analyses use the corrected length estimate. As the size of each bounding box is dominated by the largest contained streak, this relation might not be entirely representative for bounding boxes that include multiple streaks with distinct streak length differences. We further establish a linear relation between the length of streaks and their width, measured at their morphological center, using the same randomly drawn subsample of 1000 slope streaks:

$${width}=0.0461*{{length}}_{{actual}}+19.038$$
(2)

In addition, we use the subsample to quantify the average number of individual slope streaks contained in one single detection (see Supplementary Fig. S1), calculating an average of 2.33 streaks/detection. Multiple, spatially co-located slope streaks are not always easily discernible, such as visualized by Fig. 1A, B, Supplementary Fig. S8. We note that the absolute counts of individual slope streaks as derived in this work might under- or over-estimate the actual number of slope streaks in a given region.

Area estimate

We use the corrected streak length and the relation between length and width to produce estimates of streak area for the entire dataset, assuming a simplified rectangular streak shape:

$${area}={{length}}_{{actual}}*{width}$$
(3)

For RSL, we manually measure a subsample of 100 semi-randomly selected, geographically-distributed features (taken from, e.g. refs. 8,51) and estimate their surface area using Eq. (3).

Geostatistical analysis & auxiliary dataset limitations

We compare our slope streak catalog to a large number of auxiliary spatial raster and vector datasets using Quantum GIS (QGIS). Specifically, we use TES surface albedo31, MOLA topography81, RSL locations mapped in MRO HiRISE (such as by McEwen et al.8,51), modeled environmental outputs produced by the Mars Climate Database (MCD34, assuming a climatology average solar scenario), date-constrained (new) impacts39, InSight tecto-genetic broadband (BB) and low-frequency (LF) events with back azimuth (BAZ41,42), the most recent version of the martian geologic map82, river valleys47, rockfall locations45, OMEGA nanophase ferric iron83, TGO FREND water-equivalent-hydrogen (WEH84), GRS H2O and H85, and dust devil locations (ref. 46, covering MY28 to MY36). We use the MOLA topography data to create two derived topography products, slope angle and slope aspect, using built-in QGIS algorithms. We quantitatively define the significance of differences between different populations and datasets using a two-sample T-test with a default P-criterion of 0.05.

Importantly, many of the used global auxiliary datasets feature spatial resolutions several factors below the size of CTX-detected slope streaks and/or depend on various assumptions and models, such as the MCD-derived outputs. For example, the global, km-scale, MCD-derived wind speed maps do not fully capture the complex influence of small-scale topography on wind speed and direction. We underline the importance of local, site-specific studies to verify the global-scale observations and interpretations made by this work.

Seismic & impact event distance analysis

Using a methodology developed by Bickel et al.70,86, we quantify the statistical relation between streak and date-constrained impact/marsquake epicenter locations (Fig. 10). First (‘Criterion 1’), we calculate the shortest distance between all streaks and a) all real impacts/quakes and b) the same number of randomly placed (synthetic) points (in the same area). Second (‘Criterion 2’), we count the number of streaks in 0.5, 2, and 5° radii around a) all real impacts/quakes and b) the same number of randomly placed (synthetic) points (in the same area). If there was a physical relation between impacts/quakes and streak occurrence, the average distance to streak locations should be smaller, and the number of streak should increase in the proximity of real impacts/quakes (Fig. 10).

Fig. 10: Schematic of statistical analysis of streak and seismic/impact event co-location.
figure 10

A Schematic map of streaks (black dots, n = 22), impacts (blue, n = 4), and synthetic impacts (orange, n = 4); B Assessment of minimum distance between streaks and (real/synthetic) events; C Assessment of number of streaks in n° bins around (real/synthetic) events. A relation between an impact/seismic event and streaks would result in smaller minimum distances and increased numbers of streaks closer to the events. D Real example from Arabia Terra, including bright (white dots) and dark streak detections (black dots); diameter of circles does not properly reflect the 0.5, 2, and 5° radii used throughout this study. Note how the individual streak detections cluster along topographic inclines, mostly along crater walls. Viking merged color mosaic in the background.

Full size image

We note that this method is strongly affected by the quality and spatiotemporal availability of the impact and marsquake data; ref. 39 acknowledge that their impact catalog is likely incomplete, particularly in the southern highlands; in addition, a CTX image used in the mosaic might have been taken prior to an impact even at the same location. However, there is a substantial number of date constrained impacts available for a statistical analysis (n = 1203) which have been mapped over the same time period as the images for the global CTX mosaic have been acquired, indicating there is a robust statistical foundation for our analyses.

In turn, a substantial fraction of the CTX images in the global mosaic were acquired prior to the InSight mission (late 2018 to late 2022), making a meaningful comparison with the seismic data challenging. Tecto-genetic quakes are likely to periodically re-occur in the same regions, but their periods are unknown and might exceed the survival time of slope streaks. In addition, InSight’s records are occasionally noisy, i.e., might not include all tectonic events in the monitoring period, do not cover all of Mars’ surface (core shadow), and suffer from substantial localization uncertainties (e.g. ref. 87). Overall, the geostatistical comparison of streak and marsquake locations is not expected to be entirely conclusive. In order to compensate the data limitations, we also compare streak locations with the distribution of permanent geomorphic expressions of geologically recent tectonic activity, such as graben and fossae.