Introduction

The search for evidence of the current presence of liquid water on the Martian surface is essential for advancing Mars scientific research and future exploration missions1. Recurring slope lineae (RSL) are one type of low-albedo slope lineae2,3 observed on the surface of Mars, thought to result from transient flows of liquid water or dry particles4,5. There are two other dark streak types on the Martian surface: circumpolar dunes spot flows6 and dark equatorial slope streaks7. However, in the south mid-latitude region on Mars, the appearance and disappearance of RSL are closely related to seasons on Mars1,8,9: RSL in the southern mid-latitude regions appear and grow during the Martian maximum solar radiation input4,10 and gradually disappear during the Martian minimum solar radiation input. The duration of an RSL cycle is typically one Mars year (MY)11,12.

Models describing the origin and circulation of RSL can be divided into “wet” and “dry”4,5. The wet model describes RSL as flows of water or brine4,8,9,13. According to this model, solar irradiance leads to surface heating14 and thawing of the subsurface layer, activating percolation and capillary suction15. This process describes the growth as unconfined fluid flow through a porous medium16 which helps explain the seasonal dependence and slope exposure of RSL3. Events such as Marsquakes or impacts may lead to the rupture of the aquifer17,18. The seasonal melting of ice due to increased rock temperatures can result in subsurface water percolating during the warm season19. The water source may thus be melted subsurface ice located under the surface10,20 or liquid brine deposits3,19. Tectonic and impact-related structural faults may form natural drainage channels that allow the uplifting of the brine from these saline aquifers. Recent spectral analyses suggested the presence of hydrated salts, such as magnesium perchlorate, sodium perchlorate, and magnesium chlorate in RSL13 implying a potential correlation with brine activity. However, the detections of perchlorates reported by Ojha et al.13 have been challenged. Leask et al.21 and Vincendon et al.12 have raised concerns about the reliability of these detections, pointing to possible column artifacts and noise issues in the CRISM data. RSL typically occurs along the steep slopes of ancient bedrock alcoves, especially on the walls of impact craters14 and Valles Marineris22. Moreover, studies of RSL’s aspect show that approximately 60% appear on equator-facing slopes3. Decreased water influx, freezing and gradual evaporation may cause the streams to fade. The wet model thus implies the presence of significant groundwater deposits with briny or freshwater composition on Mars10,15.

In contrast, the dry model11,23 hypothesizes that RSL are formed due to downslope dry granular flow, such as sand flow24,25. Under this hypothesis, sand particles become unstable and trigger avalanches due to temperature changes and gas flow26. RSL grow with the granular flow26 and fade with dust removal27, weathering23, or overlapping by material12,26. Besides, the dry model explains the formation of some RSL on the dunes (e.g.28), without visible connection to the bedrock outcrops. However, these spots are not widely spread and probably could be explained by the close underlying liquid sources. Also, the dry model fails to explain certain morphological characteristics, such as RSL disruption in areas of high, rough, or rocky terrain, as noted by Lark et al.5.

From 2014 to 2022, corresponding to MY 32 to 36, a significant global dust event transpired during the summer of MY 34, equivalent to Earth year 201829,30. Concurrently, atmospheric temperatures on Mars experienced a marked increase31,32 with the southern high latitudes witnessing a temperature surge of approximately 40 K31. The number and spatial distribution of RSL notably expanded, with observations extending beyond the mid-latitudes of the southern hemisphere to include equatorial regions24.

The High Resolution Imaging Science Experiment (HiRISE) images with a high resolution of 0.25 m/pixel are essential for RSL detection and measurement, especially for narrow RSL. HiRISE, as a sun-synchronous orbiter33,34 can capture multi-temporal surface images with similar sun incidence and emission angles. RSL can be distinguished from their surroundings by their morphological25,35 and spectral features13,35. However, the unique morphology and seasonal occurrence of RSL pose challenges for manual digitalization and quantitative analysis, especially for vast time-series datasets.

Therefore, this study focuses on the effect of the MY 34 global dust event on the RSL spatial distribution using time-series analysis of RSL and exploring its implications for revealing new evidence of contemporary water activities on Mars. Stillman et al.11 employed the Mapping and Automated Analysis of RSL (MAARSL) for automatic mapping of RSL, complemented by manual mapping. However, a notable limitation of this method is its failure to account for variations in each image surface composition, which influence reflectance properties and, consequently, the intensities observed. This necessitates the development of a new model tailored to each individual HiRISE images. In response to this limitation, a deep-learning approach based on the Siamese network36 is developed for the automatic detection of RSL using HiRISE images collected between MY 32 and MY 36. Time-series variations of RSL in this period at two representative RSL sites, Palikir and Raga craters1 are quantitatively analyzed. Based on the time-series analysis results and association between RSL and bedrock outcrops, a scenario involving bedrock aquifer melting or brine uplifting for RSL formation on Mars is suggested as presumptive. This hypothesis is further supported by the association between RSL and a landslide in a particular study area and the time-series variations in the spectral signatures of hydrated salts within RSL.

Time-series variations of RSL

Since the discovery of RSL on Mars, specific regions have been regularly monitored9 such as Palikir and Raga craters shown in (Fig. 1). Our study systematically investigated these two craters, given the extensive coverage of HiRISE images for MY 32–36 period. Palikir crater, with a diameter of 15 km and depth of ~ 1.5 km, is located at (41.58°S, 157.83°W). The selected study area in this crater lies on its eastern wall (Fig. 1b), chosen based on the availability of HiRISE images. Raga crater, with a diameter of 3 km and depth of 654 m, is located at (48.10°S, 117.57°W). This crater is small enough to be covered by a single strip of HiRISE images, facilitating the exploration of RSL in the entire crater. In addition to RSL, in the studied craters, well-distinguished bedrock outcrops are located in the upper part of the slopes, and the north-facing inner slope of Palikir crater is covered by a possible landslide (see details in Fig. 5).

Fig. 1
figure 1

(a) Locations of Palikir and Raga craters on a MOLA global map; (b) Palikir crater with a CTX image as the base map and HiRISE footprints; (c) Raga crater with a HiRISE image as the base map. Panels on the right in (b,c) show enlarged views of RSL appearing in southern summer and gradually disappearing afterwards. The software ArcGIS 10.6 (https://www.esri.com/) was used to create the maps in the figure.

Full size image

For a comparative analysis of changes in RSL across different MYs, the solar longitude (Ls) values of the images selected from each MY should be similar. Details of the HiRISE images used for the time-series analysis of RSL at Palikir and Raga craters are listed in (Table S1). For Raga crater, suitable summer data for MY 35 are missing, and this period is thus excluded from the analysis.

To investigate the time-series variations of RSL using HiRISE images from different MYs, we developed a deep-learning approach based on the Siamese network36 for the automatic detection of RSL. Details of the deep-learning approach are presented in the Methods section, and an evaluation of the results is provided in (Fig. S1). The results of the time-series analysis of RSL are outlined in the following sections.

Time-series variations of RSL across different MYs

The features of RSL in different MYs were extracted by first using a deep learning model, followed by manual verification to ensure accuracy. The area was then calculated from the shapefiles generated during this process. As shown in Fig. 2, MYs 32, 33, and 36 observed minor fluctuations in RSL’s coverage, representing seasonal cycle fluctuation. However, in MY 34, the RSL area significantly increased by ~ 170% and ~ 65% for Palikir and Raga craters, respectively, compared with the mean value of inter-storm years. After the global dust event, the RSL’s area gradually decreased and, in two MYs, returned to its pre-global dust event state.

Radiometric parameters (I/F values), meaning the ratio of reflected intensity to incident sunlight intensity, can be calculated by multiplying the digital number (DN) by a scaling factor and then adding an offset. I/F values of RSL exhibit a consistent cycle, with lower values (indicating darker surfaces) in MY 34, followed by gradual brightening in the subsequent MYs. The I/F values of RSL in MY 34 decreased by about 50, suggesting significant darkening. This darkening could be attributed to an increased volume of liquid or other substances, indicating that the volume of a substance or the force triggering the RSL significantly increased during the global dust event in MY 34 and gradually decreased afterward.

Fig. 2
figure 2

The time-series summer HiRISE images in the same scale of four areas show an example of yearly cycle variations of RSL across different MYs. The (a) and (b) panels correspond to Palikir crater, (c) and (d) to Raga. Red arrows in the first column mark features as references to indicate the area variation of the RSL. The time-series variations in the total areas of detected RSL are shown in (e) for Palikir and (f) for Raga across MY 32 to MY 36. The software ArcGIS 10.6 (https://www.esri.com/) was used to create the maps in (a–d).

Full size image

RSL variations with respect to slopes and aspects in different MYs

The RSL are mainly located in the upper steeper parts of the crater walls. The slope values follow a normal distribution and range from about 20° to 50°, an average slope of 33.2°, and a standard deviation (SD) of 5.3° (Fig. 3a–d). In MY 33, the area and number of RSL are significantly lower, ranging from 30°–40° slope inclination. Statistically, the difference between MY 34 and MY 33 is more evident in Palikir crater’s total area and RSL number. The difference in the number of RSL in Raga crater during the studied seasons is less noticeable.

We also examined dozens of distinct RSL within the Palikir crater and Raga crater and analyzed their slopes individually. Results show that some of these RSL are on relatively constant slope of approximately 30°, but some are on varying slopes. The RSL with maximum slope variation is on slopes from 27.1° to 50.4°. Figure S2 shows examples of the examined RSL and their detailed slope statistics.

The slope orientation (or aspect) of RSL was computed from SfS DEMs, as shown in (Fig. 3e–h). The RSL in the study area in Palikir crater were concentrated along the western- and northwestern-facing inner crater walls. The RSL in Raga crater are located along northern- and northwestern-facing inner crater walls. Over 80% of the RSL concentrated on aspects between 270° and 360° (northwest-facing). Palikir crater does not have a noticeable difference in RSL’s aspects in MYs 34 and 33 (Fig. 3e,f), while for Raga Crater, the non-global dust event year (MY 33) does not have well-defined peaks, and the values are homogeneously spread on northwest-facing slopes (Fig. 3h). This behavior may indicate stable annual RSL activity, which covers a vast area on northwest-facing slopes in Raga crater. In contrast, the same activity in Palikir crater is limited. Figure S3 shows detailed examples of the RSL overlaid on the slope and aspect maps of the study areas in Palikir and Raga craters.

The observed concentration of RSL on northwest-facing slopes in Palikir and Raga craters is consistent with RSL in other craters at similar latitudes37. Their distribution on the northwest-facing crater walls appears attributable to the fact that the two craters are located in the southern hemisphere, with the highest temperature (over 250 K, Fig. S5) at around 13.0 h in the Local True Solar Time (LTST) within a day38 where the walls can absorb more solar energy than other orientated walls. Another reason for the observed concentration is the distribution of bedrock in the craters, which may serve as a potential source of RSL1. The northwest-facing walls of the craters (especially Raga crater) have more abundant bedrock exposures (Fig. 4) than the opposite direction, and their influence is discussed in the subsequent section.

Fig. 3
figure 3

(a–d) Show histograms of the area and number of RSL with respect to slopes in Palikir crater MY 34 (a) and MY 33 (b) and Raga crater MY 34 (c) and MY 33 (d). (e–h) show rose maps of the number ratio and coverage area of RSL with respect to aspect in Palikir crater MY 34 (e) and MY 33 (f) and Raga crater MY 34 (g) and MY 33 (h). The outer axis represents the aspect distribution, with 0° indicating a north-facing slope. The length of the bar on the number ratio axis indicates the percentage of RSL in a specific direction relative to all RSL.

Full size image

Bedrock aquifer melting for RSL formation on Mars

Association between RSL and bedrock outcrops

Previous research (e.g.1,22), has indicated that subsurface bedrock might be related to the formation of RSL. The RSL examples in (Figs. 1b, 2a,b) demonstrate that the RSL initiate around the steep bedrock outcrops, having a narrow shape that widens in the lower part to form an accumulation cone. In this section, we systematically investigated the associations between RSL and bedrock outcrops in the two study areas (Fig. 4).

The association between RSL and bedrock outcrops is better represented on a local scale (Fig. S4), where it is possible to calculate elevation dependencies between these features. Figure 4a,c show four elevation clusters of RSL in Palikir crater. These clusters represent nonuniform erosion of slopes mainly triggered by gullies39. The first cluster of bedrock outcrops related to RSL, located in the lower part of the slopes (-632 m mean elevation, 51 m SD), represents remnants of bedrock outcrops (-578 m, 52 m SD) with low gullies and RSL activity. The RSL in cluster 2 (-443 m, 22 m SD) differ from others, originating from well-connected bedrock outcrops (-379 m, 27 m SD). This area also does not show signs of highly active erosion. Cluster 3 is divided into two areas and has fragmented bedrock outcrops with a mean RSL elevation of -319 m (29 m SD). This cluster’s mean elevation of bedrock outcrops is -256 m (46 m SD). Cluster 4 has the highest erosion rate, indicating that the geological processes in this region are particularly active. This highest erosion could be due to a combination of factors such as the area’s topography, climate conditions, and the concentration of RSL, which are known to influence surface dynamics, with a mean elevation of -207 m (33 m SD). The elevation of bedrock outcrops in this area is -156 m (56 m SD).

In contrast, RSL in Raga crater are exclusively associated with the bedrock outcrops on the northwest-facing slopes, with no RSL identified close to the bedrock outcrops on the northeast- and south-facing slopes. The elevations of the RSL on the northwest-facing slopes exhibit only one peak at 1660 m with a unimodal distribution shown in (Figs. 4b and 6d), which indicates a continuous and uniform bedrock layer that exposed at the surface and consistent conditions and sources for RSL formation.

Fig. 4
figure 4

The distribution of bedrock outcrops (red lines) and RSL (black lines, mapped in MY 34) in Palikir (a) and Raga (b) craters. Histograms of RSL elevations in Palikir (c) and Raga (d) craters. Values in brackets indicate the mode of elevation and the associated number of RSL samples. Numbers 1–4 in (a) and (c) show RSL clusters related to the bedrock outcrops. Elevation histograms of the RSL (c,d) were plotted by sampling the elevations along each RSL in 1-m intervals. The software ArcGIS 10.6 (https://www.esri.com/) was used to create the maps in (a,b).

Full size image

RSL are predominantly aligned with the bedrock outcrops on sun-facing slopes in both study areas. Our results confirm the association between RSL and bedrock outcrops on slopes, as reported in previous publications1,23. The water source of RSL could be recharged from buried ice, atmosphere, or aquifer15. In early Martian wet climate (e.g.40), water could naturally percolate into the ground until it reached waterproof rocks, forming an aquifer (e.g.41). As the average Martian temperature decreased, a global cryosphere formed, effectively freezing and sealing groundwater in aquifers42. For shallow subsurface aquifers on sun-facing slopes, when the surface temperature exceeds the freezing temperature of water or brine on Mars, e.g., as low as 199 K (-74 °C)43 or 180 K (−93 °C)44 depending on the saturation condition of mixtures of salts in liquid water, the aquifer melted to form RSL on slopes. Additionally, the process of deliquescence, where salts absorb atmospheric moisture, could also play a role in RSL formation. The surface temperature at Palikir and Raga craters can reach 250 K (−23 °C) during the summer seasons (Fig. S5), as indicated by Thermal Emission Imaging System (THEMIS) data. Therefore, it is plausible to assume a scenario of bedrock aquifer melting and deliquescence45 for RSL formation on Mars.

Cluster 1 in Palikir Crater is located significantly lower than the main concentration of RSL and represents a possible remnant of bedrock outcrops. The concentration of RSL in this region is low. This pattern can be attributed to the blockage of rock faults at lower elevations by sediment, forcing water to emerge from faults at higher elevations. The distribution of RSL in Palikir Crater shows rising erosion activity and RSL concentration consistently from north to south for clusters 2 to 4, correlated with input solar radiation. To the south, from cluster 4, with a higher concentration of RSL and erosion activity, we found a landslide body whose formation was probably related to water activity.

Association between RSL and a large landslide in Palikir crater

The southern part of Palikir crater is encompassed by a large landslide (Fig. 5a), spanning a width of 8.5 km (east-to-west) and a length of 4.6 km (south-to-north), covering an area of approximately 43 km2. Figure 5b shows a DEM (50 m/pixel) generated from High-Resolution Stereo Camera (HRSC) images46 where two elevation profiles are derived—one inside the landslide region (red line) and the other outside the landslide region (blue line). A comparison of the elevation profiles in Fig. 5c shows a subsidence of approximately 300 m of the landslide from the crater rim. RSL (blue lines) terminated its southwest-ward extension precisely at the point where it would have entered the landslide region (Fig. 5a).

Different types of mass-wasting processes can be found on the surfaces of planetary bodies. One of the most common causes is the gravitational movement of material on unstable slopes triggered by seismic activity (e.g.47). In the following, we show through numerical modeling (Fig. 5d–f) that slope destabilization could also result from subsurface water interaction with structural faults48.

Fig. 5
figure 5

(a) Locations of the landslide (red line) and RSL (blue lines) in Palikir crater; (b) HRSC DEM covering the landslide, with colored lines indicating the locations of the elevation profiles; (c) Elevation profiles derived from the landslide and non-landslide regions on the DEM. (d–f) show numerical modeling of the landslide at the initial, middle, and final stages, respectively, with different displacement d. p is the pore pressure due to a possible fault within the bedrock resulting from uplifting brines. (g) shows the actual profile derived from the HRSC DEM, which matches the profile of the simulated landslide in (f). The software ArcGIS 10.6 (https://www.esri.com/) was used to create the maps in (a,b).

Full size image

Upon analyzing the simulation results generated from finite difference method (FDM), it was observed that with 7.0 MPa pore pressure from the bottom, the maximum displacement reached approximately 94 m. The delineation of the sliding surface, as illustrated in (Fig. 5f), resembles the actual HRSC DEM profile depicted in (Fig. 5g). The modeling results suggest that excessive water invasion by uplifting the brines results in significant water saturation, destabilizing the slope and inducing the landslide. Consequently, the landslide body acts as a low-permeability barrier, preventing the necessary hydrological conditions for RSL formation. The displaced material covers the bedrock outcrops and blocks the passage of groundwater from the bedrock aquifer to the surface, which was the water source for RSL. This could explain the absence of RSL on the north-facing slopes of Palikir crater, where the landslide is located.

Time-series variations in spectral signatures of hydrated salts in RSL

Ojha et al.13 found evidence for hydrated salts in RSL in a few places based on data from Compact Reconnaissance Imaging Spectrometer for Mars (CRISM)49. The identified hydrated salts include magnesium perchlorate, magnesium chlorate, and sodium perchlorate, suggesting water-related activities of RSL. We extended the spectral analysis to include seasons in which RSL are extensive and non-extensive to investigate the time-series variations in spectral signatures of hydrated salts within RSL.

To this end, we used two sets of CRISM datasets covering part of the study area in Palikir crater (Fig. 6a). One dataset was collected during the summer (Ls: 359°) of MY 30 (Fig. 6b), while the other was acquired during the autumn (Ls: 35°) of MY 33 (Fig. 6c). Both datasets had a spatial resolution of 18 m/pixel. The preprocessing of CRISM data was conducted using the CRISM Analysis Toolkit (CAT) embedded in ENVI to ensure data quality and suitability for subsequent analysis. Radiometric, atmospheric, and geometric corrections were applied to ensure robust and accurate scientific analysis. Referring to the HiRISE image, we selected two sampling sites within the RSL to retrieve their spectra from the CRISM datasets. As the spatial resolution of the CRISM datasets is coarser than HiRISE, the sampling sites were chosen to be those in which RSL are adequately broad or dense to fill a CRISM pixel (Fig. 6d).

Leask et al.21 indicated that in a small fraction (< 0.05%) of pixels in CRISM images, noise in the data is smoothed in a way that mimics real mineral absorptions, occurring most often at 1.9 and 2.1 μm. To minimize the risk of false identification of hydrated salts due to possible artifacts of the CRISM instrument, as discussed by Gosset50 and Vincendon et al.12, we avoided using the cherry-picking single pixel-based method. Instead, we adopted a spatial averaging approach to improve the reliability and robustness of the spectral analysis51. Specifically, for each sampling point on the CRISM image, we selected three adjacent pixels and calculated their mean reflectance values. This method reduces the influence of random noise and isolated pixel artifacts, providing a more representative spectral signature of the target area. Before calculating the average spectrum, we processed the spectral curves using a Savitzky–Golay smoothing filter. The averaged spectral curves, shown as bold blue and green lines in (Fig. 6e,f), provide a clear and accurate representation of the surface composition, enhancing confidence in the identification of hydrated salts. We also examined additional locations within the same CRISM image and found no discernible absorption features. Additionally, the raw data of adjacent pixels are also plotted in Fig. 6e,f as light blue and green curves. From the raw data, we did not observe obvious artifacts of column-dependent spikes as described in Leask et al.21. Our spectral analysis, which scrutinized observations from different seasons of CRISM data at two distinct sampling sites, revealed spectral signatures of hydrated salts52 in summer (Fig. 6e) and the absence of such signatures at the same sites in autumn (Fig. 6f). In Fig. 6e, the reflectance curves exhibit absorption features near 1.47, 1.91, and 2.12 μm, indicating the presence of H2O and suggesting the occurrence of hydrated calcium perchlorate, magnesium chlorate, and magnesium dichloride, corroborated by their laboratory spectra53 shown in Fig. 6g. The wavelength position of the observed absorption at 1.47 μm slightly deviates from the laboratory spectra of 1.43 μm, indicating a spectral mixture of hydrous salts and other minerals. In contrast, no apparent absorption features at the bands mentioned above can be found in the reflectance curves for autumn shown in Fig. 6f.

To be more informative, we calculated and plotted the ratio of summer and autumn reflectance of the two sampled RSL points in Palikir crater (Fig. S7). It shows that the ratios of reflectance for both points show distinct absorption features at wavelengths of 1.47, 1.91, and 2.12 μm. These spectral signatures clearly demonstrate the differences in the observed CRISM data between summer and autumn, suggesting variations in the surface materials during these two seasons. The presence of these features in the ratio plots (Fig. S7) clearly demonstrates that the observed CRISM data differ between summer and autumn, suggesting that the composition or abundance of hydrated salts varies with the seasons. This seasonal variation, observed across multiple seasonal cycles in different locations, may be linked to environmental processes such as temperature fluctuations, or water activity, providing valuable insights into the dynamic nature of RSL sites on Mars.

Fig. 6
figure 6

Time-series variations in spectral signatures of hydrated salts in RSL at Palikir crater. (a) Overview of the study area in Palikir crater, with the blue box indicating CRISM coverage; (b) CRISM data in summer (FRT0002038F, Ls: 359°) with sampling points in RSL; (c) CRISM data in autumn (FRS00037F73, Ls: 35°) with the same sampling points; (d) HiRISE image in summer for reference; (e) Spectra of the sampling points from the CRISM summer data, indicating the presence of hydrated salts; (f) Spectra of the same sampling points from the CRISM autumn data showing no obvious absorption features related to hydrated salts; (e,f) fine lines - per pixel spectra, and bold lines - mean of per pixel analysis; (g) Laboratory spectra of relevant hydrated salts for reference53. The software ArcGIS 10.6 (https://www.esri.com/) was used to create the maps in (a–d).

Full size image

Additionally, the spectral analysis of RSL in new locations—Tivat crater and Rauna crater—using CRISM data from summer and autumn are presented in Fig. S8 and Fig. S9. All CRISM data involved are listed in Table S2. Raga crater is not used here due to the absence of CRISM data. Figures S8 and S9 exhibit a similar pattern of time-series variations of hydrated salts as observed in Palikir crater.

This time-series spectral analysis further supports the scenario of bedrock aquifer melting for RSL formation. The hydration process may occur with the formation of hydrous salts in the aquifer. Brines can be stored in deep aquifers48 or preserved as subsurface ice or liquid3,10,19,20. In summer, as bedrock outcrop temperatures increase, subsurface ice begins to melt, causing brine to rise or percolate to form RSL. This process results in the deposition of hydrated salts on the Martian surface3 which can be detected through CRISM data. Conversely, when the RSL later disappears, the hydrated salts may become amorphous due to dust activity54 desiccated55 or be covered by dust in autumn or winter. Therefore, no obvious hydrous mineral signatures can be observed at the same site.

Conclusions and discussion

The time-series analysis of RSL in Palikir and Raga craters on Mars revealed a noteworthy increase in the number of RSL following the 2018 global dust event (MY 34), which also corresponds to the previous research8,11,24,27. The number of RSL in MY 34 increased by approximately 65–170% compared with the previous and subsequent MYs for Palikir and Raga craters, respectively, suggesting a strong correlation between the global temperature rise in the post-global dust event period and RSL formation. The time-series analysis of RSL indicates a significant increase in RSL activity during MY 34. However, there was no observed change in RSL distribution concerning aspect spreading, which would typically be expected with rising temperatures and increased atmospheric turbulence during this period. While dry mechanisms, such as dust involvement, might explain the increased RSL activity following a global dust event due to greater atmospheric dust deposition, they do not account for the absence of aspect spreading. Therefore, dry mechanism alone may not explain the behavior observed in MY 34. In the context of the wet model, this behavior could indicate stable yearly RSL activity originating from a single source related to the bedrock outcrops. One year of rising temperatures is insufficient to initiate the same process over a broader area.

The distribution of RSL located in the southern mid-latitude regions along northern- and northwestern-facing slopes indicates a strong dependency on surface solar irradiance (Fig. S5), consistent with previous research1,9,37. Spatial analysis of the relationship between RSL and bedrock outcrops in Palikir and Raga craters revealed a strong spatial association between the RSL and sun-facing bedrock outcrops in the southern mid-latitudes of Mars. This supports the hypothesis that melting of bedrock aquifers due to sufficient solar radiation is a plausible mechanism for RSL formation in the southern mid-latitude regions. Additionally, the numerical modeling of the landslide found in Palikir crater showed that water saturation in the upper part of the slope could cause destabilization and landslide formation. The proximity of the landslide to the RSL region and the absence of RSL within the landslide further strengthen the argument for bedrock aquifer melting. Moreover, spectral evidence of hydrated salts in RSL was found in CRISM data collected in the summer, while it was absent in CRISM data from the same sites in the autumn. The time-series variations in the spectral signatures of hydrated salts in RSL provide further validation for the mechanism of bedrock aquifer melting in RSL formation. To get closer to the solution for the formation of RSL, future observations such as high-spatial-resolution hyperspectral data with minimal endmember mixing are needed. Such data would provide direct evidence by clearly distinguishing the spectral differences between RSL and their surrounding background, thereby enhancing our understanding of the formation mechanism of RSL.

In summary, the presented results provide evidence to support the hypothesis that RSL is related to contemporary water activity on Mars, offering a possible scenario of bedrock aquifer melting for RSL formation. Note that it is also important to explore and potentially rule out other possible scenarios for RSL formation, which will be the focus of our future efforts.

Methods

Deep-learning approach for RSL detection

We developed a deep-learning approach based on the Siamese network36 to automatically detect RSL from the HiRISE images of different Mars Years. The inputs to the network are two images covering the same area. The two images are down-sampled through the U-Net56 structure to gradually extract feature information. Weights are assigned and shared between the two images. Afterward, a difference image from the two input images is obtained and up-sampled to generate the detection result by comparing it with the input labels. The difference between the two input images is calculated using a loss function equation, and weights are adjusted iteratively. Compared with the complete HiRISE image, the number of pixels in RSL (i.e., positive samples) is significantly smaller than that of non-RSL (i.e., negative samples) pixels, indicating the imbalance of positive and negative samples. To address this problem, a focal loss function57 designed for the imbalance between positive and negative samples was used. In addition, a spatial attention mechanism58 was introduced to adjust the response features, amplifying useful information and reducing the impact of non-useful information.

The HiRISE images covering Palikir crater, collected in both the summer and winter of MY 31, were used to construct the training and testing datasets. The images were split into a total of 5358 square patches and randomly divided into 70% and 30% to form the training and testing datasets, respectively. Figure S1 shows a pixel-based evaluation of the RSL detection results. Most RSL could be detected, although specific challenges were encountered in detecting the RSL edges. The accuracy evaluation primarily relies on balanced accuracy to address the issue of imbalanced positive and negative samples, defined as follows:

$$\:\text{S}\text{e}\text{n}\text{s}\text{i}\text{t}\text{i}\text{v}\text{i}\text{t}\text{y}=\:\frac{TP}{TP+FN}$$
(1)
$$\:\text{S}\text{p}\text{e}\text{c}\text{i}\text{f}\text{i}\text{c}\text{i}\text{t}\text{y}=\:\frac{TN}{TN+FP}$$
(2)
$$\:\text{B}\text{a}\text{l}\text{a}\text{n}\text{c}\text{e}\text{d}\:\text{A}\text{c}\text{c}\text{u}\text{r}\text{a}\text{c}\text{y}=\:\frac{Sensitivity+Specificity}{2}$$
(3)

where TP, FN, TN, FP is defined as true positive, false negative, true negative and false positive respectively, in the model’s detection outcomes.

The proposed deep-learning approach achieved a balanced accuracy of 81.8%. Most false detections occurred in hill shade areas, and missed detections were associated with highly narrow RSL (~ 1 pixel in the HiRISE image). After the automatic detection by deep learning, extensive manual checking and digitization were performed to further improve the automatic detection results.

SfS algorithm for pixel-wise high-resolution DEM generation and local slope and aspect calculations

We used the shape-from-shading (SfS) algorithm59,60 to generate pixel-wise high-resolution DEMs (0.25 m/pixel) based on the HiRISE images. The SfS uses light scattered from a surface to reconstruct 3D topography and can retrieve subtle topographic details from monocular images. This approach was developed for the 3D mapping of the lunar surface59,61 and adopted to Mars surface reflectance parameters affected by the scattering and absorption by aerosols in the atmosphere60. The SfS algorithm uses a high-resolution image, the corresponding coarse-resolution DEM, and information on the optical depth as inputs. A radiative transfer model adapted to the Martian atmosphere incorporates the atmospheric effects. The output is a high-resolution DEM with a resolution identical to the input image. Liu and Wu60 indicated that the SfS algorithm can achieve a geometric accuracy of ~ 2 pixels of the image resolution and greatly enhance topographic details. The test conducted at Victoria Crater using HiRISE imagery demonstrates satisfactory results, with a root mean squared deviation (RMSD) of 0.5 m, a maximum absolute elevation difference of 6.8 m, and a Pearson correlation coefficient of 0.785 between the computed shaded relief and the input image. The high-resolution DEMs generated by the SfS algorithm offer more detailed topographic information (Fig. S6) than DEMs generated using conventional techniques such as photogrammetry, facilitating detailed measurements of RSL geometry and the surface slopes and aspects.

Several methods have been used to calculate slopes and aspects algorithms (e.g.62,63), revealing slight variations in the calculated results. In this study, we used the classic third-order finite difference method, weighted by the Reciprocal of Squared Distance as proposed by Horn64 to calculate slopes and aspects.

Fig. 7
figure 7

a 3 × 3 surface scanning window.

Full size image

The algorithm used to calculate the slope and aspect is as follows:

$$\:slope\_degrees\:=\:arctan\sqrt{{\left(\frac{dz}{dx}\right)}^{2}+{\left(\frac{dz}{dy}\right)}^{2}}\times\:\frac{180}{\pi\:}$$
(4)
$$\:aspect\_degrees\:=\:atan2(\left(\frac{dz}{dy}\right),-\left(\frac{dz}{dx}\right))\:\times\:\frac{180}{\pi\:}$$
(5)

where the rate of change in the x direction \(\:\left(\frac{dz}{dx}\right)\) and the rate of change in the y direction \(\:\left(\frac{dz}{dy}\right)\) for the central pixel (defined in Fig. 7) is calculated with the following algorithms (zi (i = 1,2,…9) is the elevation value in cell i):

$$\:\frac{dz}{dx}\:=\frac{(\:4\times\:\frac{(z3+z9+2\times\:z6)}{wght1}-4\times\:\frac{(z1+z7+2\times\:z4)}{wght2})}{(8\times\:x\_cellsize)}$$
(6)
$$\:\frac{dz}{dy}\:=\frac{(\:4\times\:\frac{(z7+z9+2\times\:z8)}{wght3}-4\times\:\frac{(z1+z3+2\times\:z2)}{wght4})}{(8\times\:y\_cellsize)}$$
(7)

where wght1, wght2, wght3 and wght4 are the weighted counts of valid cells.

FDM method for landslide modeling

We used the finite difference method (FDM), recognized for its effectiveness in modeling slope stability65 for a 2D simulation of landslide formation within Palikir crater. The FDM is a numerical technique widely used for solving differential equations that arise in various fields of science and engineering. By discretizing continuous functions into a grid of points, FDM approximates derivatives using differences between function values at these points. Material properties were selected based on empirical values of terrestrial soil or inferred values from Martian activities, with the Young’s modulus set at 500 MPa for the surface soil and 2400 MPa for the underlying bedrock. The bedrock is considered elastic, while the surface soil is elastoplastic. Constitutive models were chosen to reflect the physical behavior of these materials, with the Mohr-Coulomb model applied to the surface soil and an elastic model to the bedrock. The permeability and depth of the bedrock were set at 10− 13 m/s and 750 m48. Additionally, the effective internal friction angle and cohesion of the surface soil were set at 30° and 1 kPa, respectively66,67 to accurately simulate the slope’s resistance to shearing. In the numerical model, the crater slope was first calculated to equilibrium under gravity without considering any water flow. No landslides were observed in this stage, then the landslide was triggered by increasing the pore pressure in the possible fault within the bedrock progressively. The simulation stopped automatically due to significant mesh distortion after slope failure occurred.

Spectral analysis

CRISM datasets were used to detect hydrated salts in RSL. The preprocessing of CRISM TRR3 data was conducted using the CRISM Analysis Toolkit (CAT) version 7.4 embedded in ENVI to ensure data quality and suitability for subsequent analysis. First, the raw data were imported and radiometric calibration was applied to convert digital numbers to radiance values. The observed radiances were then transformed into the radiance factor (I/F) by dividing them by the solar spectrum at the time of observation. Photometric corrections were performed by dividing the I/F value by the cosine of the incidence angle49. Atmospheric effects were minimized using the volcano scan correction technique68 to retrieve surface reflectance. Subsequently, map projection was carried out to align the images with the standard projection used by HiRISE for accurate spatial analysis. The relative reflectance of pixels was obtained by dividing the region of interest by the spectral flat from an unremarkable region, and the resulting reflectance curves were further smoothed using the Savitzky–Golay smoothing method69 to reduce noise. The reflectance curves from the CRISM datasets were compared with laboratory spectra to identify hydrated salts.