Implications of present ground temperatures and relict stone stripes in the Ethiopian Highlands for the palaeoclimate of the tropics

. Large sorted patterned grounds are the most prominent features of periglacial and permafrost environments of the mid and high latitudes, but have not yet been veriﬁed for the tropics. Here, we report on relict large sorted polygons (up to 8 m in diameter) and large sorted stone stripes (up to 1000 m long, 15 m wide, and 2 m deep) on the ~4000 m high Sanetti Plateau in the Bale Mountains, southern Ethiopian Highlands. For a systematic investigation of past and present frost-related processes and landforms in the Bale Mountains, we conducted geomorphological mapping both in the ﬁeld and on satellite 5 images. The sorted stone stripes were studied in more detail by applying aerial photogrammetry, ground-penetrating radar measurements, and 36 Cl surface exposure dating. In addition, we installed 29 ground temperature data loggers between 3493 and 4377 m to analyse present frost occurrence and seasonal temperature variations from 2017 to 2020. Finally, we ran a simple experiment and combined recent ground temperature measurements with meteorological data in a statistical model to assess the air temperature depression needed for the past formation of deep seasonal frost and cyclic freezing and thawing 10 on the plateau. Our results show that relict and modern periglacial landforms are common in the Bale Mountains. Nocturnal superﬁcial ground frost on the plateau occurs at 35-90 days per year, but the mean annual ground temperature (~11 °C) is far off from seasonal or permanent frost conditions. The modelling experiment suggests a minimum air temperature depression on the plateau of 7.6 ± 1.3 °C for the emergence of several decimetre deep seasonal frost. The stone stripes probably formed under periglacial conditions in proximity of a palaeo ice cap on the plateau during the coldest period(s) of the last glacial cycle. 15 We hypothesise that the slightly inclined and unglaciated areas of the plateau, the coexistence of regolith and large blocks, the occurrence of deep seasonal frost, as well as relatively dry conditions beyond the ice cap provided ideal conditions for frost heave and sorting and the formation of large sorted patterned grounds. The presence of these landforms and the associated air temperature depression provide further evidence for an ampliﬁed cooling of high tropical mountains during the last glacial period that is yet not well captured in global climate models. in the lab, and Armin Rist for the fruitful discussion.

4 https://doi.org/10.5194/esurf-2020-53 Preprint. Discussion started: 29 June 2020 c Author(s) 2020. CC BY 4.0 License. located south of the equator and high pressure cells have established over Western Asia and the Sahara, northeasterly trade winds from the Arabian Peninsula and Sea prevail in the Bale Mountains and cause only little precipitation. Along with the northward movement of the ITCZ from March to June, the main wind direction changes from northeast to southeast and brings moist air from the southern Indian Ocean to the Bale Mountains (Lemma et al., 2020). Although the Gulf of Guinea and Congo Basin are important moisture sources for the northern Ethiopian Highlands (Levin et al., 2009;Viste and Sorteberg, 5 2013; Costa et al., 2014), they seem of minor relevance for the Bale-Arsi massif (Lemma et al., 2020). The Sanetti Plateau and highest peaks of the massif experience occasional snowfall during the rainy seasons, but the thin snowpack usually melts within hours or days (Miehe and Miehe, 1994).
Like most of the other tropical mountains in Eastern Africa, the Bale Mountains are currently unglaciated. The present mean 0°C isotherm (a rough proxy for the modern snow line in the tropics) is located at least 300 m above the highest peak Tullu 10 Dimtu. However, latest glacial geomorphological and chronological studies provide clear evidence that the snow line was much lower during the Late Pleistocene and favoured the formation of an extensive plateau glaciation with outlet glaciers extending down into the northern valleys. Between 50-30 ka during the local Last Glacial Maximum in the southern Ethiopian Highlands, ice covered about 265 km² of the Bale and additional 83 km² of the adjacent Arsi Mountains. Two later glacial stages were dated to~18 and~15 ka. At~18 ka, the ice extent was slightly smaller than during the local LGM, but ice still covered the central part 15 of the plateau and northern valleys. Deglaciation in the region set in after~18-15 ka. The highlands remained probably ice-free over the entire Holocene (Ossendorf et al., 2019;Groos et al., in revision). Besides glacial landforms like moraines and roche moutonnées, also relict periglacial features have been reported from the Bale Mountains (Grab, 2002;Hendrickx et al., 2014).
Among those, large sorted stone circles (several meters in diameter) and stone stripes (several meters wide and hundred meter long) on the Sanetti Plateau are the most prominent ones. The formation of such large features is associated with freeze-thaw 20 processes and indicates decimetre to meter deep seasonal frost or sporadic permafrost with a thick active layer (see Sections 3.7, 4, and 5). Field observations and short-term ground temperature measurements between December 1989 andMarch 1990 verify that nocturnal frost near the soil-atmosphere interface still occurs. The most apparent results of the modern freeze-thaw cycles are the formation of needle ice along saturated stream banks and the presence of sorted stone nets in flat and poorly drained areas on the Sanetti Plateau (Miehe and Miehe, 1994;Grab, 2002). 25 Afroalpine herbs, grasses, Helichrysum dwarf shrubs, extrazonal patches of Erica, and giant lobelias cover the ice-free and barren Sanetti Plateau today (Miehe and Miehe, 1994). The plateau is home to many endemic plant species like the giant lobelia (Lobelia rhynchopetalum) (Chala et al., 2016) and mammal species like the Ethiopian wolf (Canis simensis) (e.g. Gottelli et al., 1994), giant mole-rat (Tachyoryctes macrocephalus) (e.g. Vlasatá et al., 2017), and mountain nyala (Tragelaphus buxtoni) that are restricted to the Ethiopian Highlands (Miehe and Miehe, 1994). Since these endemic species mainly populate 30 the upper valleys and Sanetti Plateau today, palaeoclimatic and -environmental changes like a severe cooling, expansion of the ice cover and periglacial area as well as depression of altitudinal vegetation belts must have directly affected their habitat and are therefore also of relevance in a geoecological context.

Mapping of periglacial landforms
Comprehensive and thorough geomorphological mapping of glacial and periglacial landforms provides crucial data for establishing glacial chronologies and reconstructing the palaeoclimate and palaeoenvironment of polar and alpine regions (Chandler et al., 2018). We evaluated maps, photographs, and field notes from previous studies dealing with periglacial processes and 5 landforms in the Bale Mountains (e.g. Messerli and Winiger, 1992;Miehe and Miehe, 1994;Grab, 2002;Umer et al., 2004;Osmaston et al., 2005) to compile evidence of relict and modern frost occurrence. Since periglacial landforms have yet not been described systematically, we performed extensive geomorphological mapping on the Sanetti Plateau, along the upper Harenna Escarpment and in the western, northern, and eastern valleys during multiple field excursions between 2016 and 2020 ( Fig.   2). In addition, we evaluated high-resolution WorldView-1 satellite images (pixel size = 0.5 m) provided by the DigitalGlobe 10 Foundation to identify features in remote and difficult-to-access areas of the mountain range. All periglacial features discovered in the field or on satellite images were geotagged and compiled in a catalogue (see Appendix A).
For a detailed analysis of the geomorphology, geometry, and size of the sorted stone stripes on the Sanetti Plateau, we conducted a manual aerial survey (~50 m above ground level) with a small quadcopter (DJI Mavic Pro) on the 30 th January 2020 at 2 pm local time. A total of 75 aerial images were acquired during the survey and processed with the photogrammetric 15 software OpenDroneMap (following the approach described in Groos et al., 2019) to obtain a high-resolution orthophoto (5 cm) and digital surface model (DSM, 10 cm) of the stone stripes. Five natural objects (rocks and dwarf shrubs) visible on the orthorectified WorldView-1 image and at least on three aerial images were used as ground control points (Fig. 1) for a rough georeferencing of the orthophoto and DSM . The necessary elevation information were extracted from the SRTM 1 Arc-Second Global dataset.

Ground-penetrating radar measurements
Investigating the internal structure of the sorted stone stripes by excavating a transect was not possible due to regulations.
Instead, we performed a ground-penetrating radar (GPR) survey between two stripes on the southern Sanetti Plateau on the 10 th February 2020 (Fig. 1). We made use of the Pulse EKKO PRO GPR with a 1000 MHz antenna (7.5 cm sensor width) manufactured by Sensors & Software Inc., which was originally purchased by another subproject of the Ethio-European DFG

25
Research Unit 2358 "The Mountain Exile Hypothesis" for geoecological investigations (for system settings see Appendix B).
The GPR was mounted on a compatible pushcart (Fig. 2). As survey setting, an exploration depth of 1 m and pulse length of 16 nanoseconds (ns) was applied for the first line and modified to 1.5 m depth and 24 ns pulse length for the following lines.
The starting point of the GPR measurement was located 10 m above the position of data loggers GT07-09 since the uppermost part of the volcanic plug was not accessible with the pushcart. Due to uneven terrain and several natural obstacles like smaller 30 stones and Helichrysum dwarf shrubs, the GPR profile between the two stone stripes was split into five separate lines varying between 3.8 and 38.5 m in length. The chaotic structure of the stones stripes prevented a GPR survey inside the troughs and coarse material. For analysis and visualisation of the GPR data, we used the software EKKO Project (version 5.0).

36 Cl surface exposure dating of periglacial landforms
Previous studies have demonstrated that the stabilisation age of relict periglacial features like rock glaciers and block fields can be successfully dated with cosmogenic nuclides (e.g. Barrows et al., 2004;Ivy-Ochs et al., 2009;Steinemann et al., in press). In analogy, we sampled two sorted stone stripes on the Sanetti Plateau, one about 5 km south and another one about 10 km west of Tullu Dimtu (Fig. 1), to determine the stabilisation age of these features. The results were originally published 5 by Groos et al. (in revision) in a palaeoglaciological context. We present them here briefly again as they are also of relevance for the discussion of the origin of the sorted stone stripes. From both stone stripes, we selected three columnar rocks for 36 Cl surface exposure dating (Table 1). To avoid distorting effects on exposure dating due to strong shielding in the trough-shaped stripes or toppling of rocks after the stabilisation phase, we only chose rocks that were sticking out a bit and were wedged between other rocks. The upper few centimetres of each target rock were sampled with hammer, chisel, and angle grinder for the subsequent laboratory analysis (Fig. 2). An inclinometer was used in the field for measuring the topographic shielding. For extraction of the isotope 36 Cl, the six samples were crushed, sieved and chemically treated in the Surface Exposure Dating Laboratory of the University of Bern. Total Cl-and 36 Cl-concentrations (see Appendix C) were measured from one target at the 6 MV AMS-facility of the ETH Zurich using the isotope dilution technique (Ivy-Ochs et al., 2004) and a gas-filled magnet to separate 36 S (Vockenhuber et al., 2019). For a detailed description of the sample preparation, Cl and 36 Cl measurements, and 5 surface exposure age calculation see Groos et al. (in revision).

Ground temperature measurements
For measuring hourly ground temperatures of the Bale Mountains, we installed high-quality UTL-3 Scientific Dataloggers (hereafter GT data loggers) in 2, 10, and 50 cm depth at five different locations with little vegetation between 3877 and 4377 m ( Fig. 1 and Table 2). The GT data loggers are developed by GEOTEST Ltd. in collaboration with the Swiss Institute for Snow 10 and Avalanche Research. They consist of a waterproof housing, a YSI 44005 thermistor for measuring temperature, a memory for up to 65.000 readings (>7 years by hourly interval), a replaceable 3.6 V lithium battery for power supply, and a USB interface for data transfer. According to the manufacturer, the measurement accuracy is <0.1°C at 0°C and the thermometric drift per 100 months is <0.01°C at 0°C. At each of the five measurement sites, the upper 50 cm of the ground were removed to install the GT data loggers (Fig. 2). We used data loggers with an external cable and thermistor for the measurements in 15 10 and 50 cm depth. A standard logger without external cable was placed just below the surface in 2 cm depth. After the installation, each hole was filled in the same order as during the excavation to ensure as little disturbance of the profile as possible. Additional low-cost tempmate.-B2 temperature data loggers (hereafter TM data loggers) in the size of a button cell ( Fig. 2) were distributed across the Bale Mountains between 3493 and 4377 m to increase the spatial coverage of near-surface (2 cm) hourly ground temperature measurements ( Fig. 1 and Table 2). The single-use TM data loggers consist of a splashproof 20 housing (we wrapped the loggers in thin tape for better protection), an unspecified thermistor, a memory for up to 8192 readings (341 days by hourly interval), and an irreplaceable 3.0 V battery. A logger-pan-to-USB cable is needed for connecting the TM loggers to a computer and retrieving the data. The measurement accuracy is ± 0.5°C at -10 to 65°C according to the manufacturer. Due to the much lower accuracy of the TM data loggers compared to the GT data loggers, we performed a comparative measurement indoor over several hours with logger GT04 as reference. Since the root-mean-square deviation of each TM data logger from the reference measurement was smaller than the stated accuracy of ± 0.5°C, a calibration was unnecessary. For a direct cross-comparison in the field, data logger TM08 was installed next to GT13 in 2 cm depth on top of Tullu Dimtu. Two TM data logger (16 and 17) were buried below Erica in 10 cm depth for comparison with sites (TM14 and TM15) with only little vegetation.
Several issues occurred during the measurement period from January 2017 to January 2020 and caused considerable data gaps 5 (Fig. 3). The data logger positions were orginially marked with four small plastic poles in 2 m distance. They were apparently too conspicuous and led to the loss of several items (GT01, GT04, GT18, TM01-03, TM08, and TM11). Therefore, we removed all poles and switched to natural markers (dwarf shrubs, stones, etc.). Data loggers GT01 and GT04 were replaced by GT16 and GT17. On Tullu Dimtu, data loggers GT13-15 were taken away in May 2017, but could be recovered and reinstalled in January 2018. Furthermore, we also noticed a relatively short battery life of just two years for some of the GT and TM data loggers 10 leading to a substantial data loss. Two years are much shorter than the battery life stated by both manufacturers for the respective sampling interval (GT~3-5 years, TM~5 years). After consultation, the manufacturer of the GT data loggers adjusted the internal handling of the lithium batteries to ensure the stated battery life. The two data loggers TM05 and TM06 temporarily recorded unrealistic low values (down to -40°C). Individual outliers and longer periods with implausible measurements were removed from the time series. After the first reading of data logger GT07, we realised that it was accidentally placed in 6 cm depth and not in 2 cm as intended. The relocation towards the surface in December 2017 led to an abrupt increase in the temperature amplitude. Therefore, we calculated hourly ground temperature gradients between 6 and 10 cm depth from GT07 and GT08 data by applying a simple linear regression to extrapolate the GT07 measurements from 6 to 2 cm in the period 21 st January to 10 th December 2017. A compilation of all ground temperature time series from the Bale Mountains is provided 5 in Table S1. Data gaps were filled using a simple linear regression and data from other GT or TM data loggers to generate a complete data set for the period 1 st of February 2017 to 20 th January 2020 (see Section 3.6). All data modifications made for each logger are listed in Table S2 (see data availability statement for supplementary Tables S1-4).

Meteorological measurements
Within the framework of the DFG Research Unit 2358 ten automatic weather stations (AWS) were installed in the Bale

10
Mountains national park between 1315 and 4377 m beginning of 2017 (Table 3). The AWS are manufactured by Campbell Scientific and consist of a three metre galvanised tubing tripod, a grounding kit, a weather-resistant enclosure, a measurement and control system (CR800), a solar module (SDT200), a 168 Wh battery, a charging regulator, a temperature and relative humidity probe (CS215) with radiation shield, a pyranometer (LI-200R), a two-dimensional ultrasonic anemometer from Gill Instruments, and a rain gauge from Texas Electronics (TR-525USW 8"). For protection, the AWS are wire-fenced by a 3 x 3 m 15 compound. Air temperature, relative humidity, and global radiation are measured in 2 m height, wind speed and wind direction in 2.6 m height, and precipitation in 1 m height. The measurement interval is 15 minutes. All measured variables are finally aggregated to hourly averages.

Statistical data interpolation and analysis
To obtain a complete and consistent data set of hourly ground temperatures for the Bale Mountains from 1 st February 2017 until 20 th January 2020 (Table S3), we applied a statistical gap-filling approach. Most of the ground temperature measurements from different locations or depths overlap for a certain period in time (see Fig. 3) and allow to establish a statistical relationship.
We applied a simple linear regression model to interpolate missing data points in the time series of a logger using data from 5 a nearby logger for which measurements were available. If multiple logger with a similar equidistance came into question for the interpolation, we chose the one that yielded the best fit (evaluated by the coefficient of determination R 2 ) and lowest root-mean-square error (RMSE). The overlapping measurement period between the predicting logger and dependent logger was split into a calibration and validation part. For the interpolation of incomplete time series in 10 or 50 cm depth, we drew on available data from 2 cm depth of the same location. We used a moving average of the data from 2 cm depth to account for 10 the time-lag response to atmospheric changes in greater depths. The number of preceding hours considered for the calculation of the moving average that yielded the best prediction (high R², low RMSE) of the ground temperatures in 10 or 50 cm depth was chosen. Details on the data gap-filling of each incomplete time series are provided in Table S4. The validation of the simple linear models applied for interpolation revealed an average R 2 of 0.86 ± 0.07 and RMSE of 1.9 ± 1.4°C. The time series of the data loggers TM05-06, TM08, and TM14-17 were not interpolated because the data served only for comparative 15 experiments (low-cost vs. high-quality loggers, vegetated vs. unvegetated locations, etc.) and were dispensable for the temporal and elevational analysis.
Most of the AWS installed in the Bale Mountains measured continuously, but some of the time series are interrupted due to issues with the power supply (Table 3). The hourly meteorological data from the different AWS are stored in an online database and gaps in the time series of all variables except wind speed and direction are interpolated statistically following a workflow 20 developed by Wöllauer et al. (in revision). Single missing values are interpolated linearly using the average of the adjacent data points. Longer gaps in a time series are filled using available data from several nearby AWS. A multiple linear regression model fitted with data from the overlapping measurement periods is applied to predict the missing values from data of those AWS that reveal a strong correlation and low RMSE. Predictor variables (AWS) with a high R 2 and low error are given a higher weight in the interpolation. 25 We evaluated the interpolated hourly meteorological and ground temperature data statistically to quantify frost occurrence and spatio-temporal ground temperature variations in the Bale Mountains. Twelve data loggers from 2 cm depth (excluding TM05-06, TM08, and TM14-17) and five loggers from 10 and 50 cm depth were considered for calculating mean annual ground temperatures, daily ground temperature cycles, thermal gradients, number of frost days, frost penetration depth, elevational gradients, etc. To study seasonal ground temperature variations related to changes in insolation, cloudiness and humidity, we 30 conducted the calculations separately for the entire study period, the dry season (Bega: November -February), and the two rainy seasons (Belg: March -June, Kiremt: July -October). Furthermore, comparative measurements were performed to investigate the differences in ground temperature between north-and south-facing slopes (GT16 and GT02-03 vs. GT17 and GT05-06), vegetated and unvegetated areas (TM16-17 vs. , and the performance of low-cost vs. high-quality data loggers (TM08 vs. GT13).

Ground temperature modelling and palaeoclimate reconstruction
The potential of periglacial landforms for paleoclimatic and environmental reconstructions has already been pointed out in pioneering studies from more than half a century ago (e.g. Galloway, 1965). Periglacial landforms are often more abundant 5 than glacial deposits, especially in dry regions, and can be a more reliable climate proxy than palaeo glacier extents as their formation is less sensitive to changes in precipitation. Here, we explore a novel and experimental approach to infer palaeoclimatic information from relict periglacial landforms and established ground temperature modelling (e.g. MacLean and Ayres, 1985) using on-site meteorological data and present ground temperature measurements.
We make the following main assumptions for our model experiment: First, the large sorted stone stripes on the Sanetti Plateau relationship, present-day hourly ground temperatures in 2 cm (T 2cm ) can be modelled using measured air temperature and incoming shortwave radiation:

25
where T air,i (i = 1,. . . ,n) is the hourly measured air temperature in°C, Q S,i is the hourly measured incoming shortwave radiation in W m -2 , β 0 is the intercept, β 1 is the coefficient for T air , and β 2 is the coefficient for Q S . The coefficients and goodness of fit for each of the three linear models are provided in Table 4. For simulating past ground temperatures, two additional parameters, ∆T air and ∆Q S , were introduced:

30
where ∆T air is the difference between the mean present-day and past air temperature in°C and ∆Q S is the difference between the mean present-day and past incoming shortwave radiation in W m -2 . For simplicity, we set ∆Q S to 30 W m -2 (the rough lowering of incoming shortwave radiation during the LGM at 15°N, see Groos et al., in revision). To infer the air temperature depression at the formation time of the periglacial landforms using Eq. 2, we increased ∆T air (starting with: ∆T air = 0°C) with every iteration until T 2cm became < -1°C. We tested all three developed multiple linear regression models (Tullu Dimtu, EWCP Station, and Tuluka) to quantify the uncertainty of the approach originating from differences in the model coefficients β (Table 4). Since the lowest-situated stone stripes on the Sanetti Plateau are located at an elevation of 3870-3890 m, we used 5 meteorological data (T air and Q S ) from the Tuluka AWS at 3848 m to run the three models. Alternatively, the meteorological data from the higher-situated AWS (Tullu Dimtu and ECWP Station) can be adjusted to the elevation of the stone stripes using a lapse rate of 0.7°C per 100 m (see Section 4.2). Running each model with the locally adjusted meteorological data led to the same calculated temperature depression as using the Tuluka AWS data. We rescaled the simulated ground temperatures in 2 cm depth to the maximum seasonal ground temperature variations in 10 and 50 cm depth observed today (see Section 4.2) to 10 model temperature variations in these depths: where (T 50cm,i ) are the simulated ground temperatures in 50 cm depth in°C (i = 1,. . . ,n), T 2cm is the mean air temperature in 2 cm depth in°C, a (= -1.25°C) is the predefined seasonal minimum, and b (= 1.25°C) the predefined maximum of T 50cm,i . For 10 cm depth (T 10cm ), a equals to -3°C and b to 3°C. The main drawback of the presented approach is the non-consideration 15 of ground moisture and thermal conductivity due to the lack of respective measurements. To further improve the method in the future, profile sensors measuring moisture, electrical conductivity, and temperature in 5 cm intervals between 0 and 50 cm depth have been installed at three AWS on the Sanetti Plateau in January 2020. The data are not yet available. (d) High-resolution orthophoto and DSM cross-section profile of the stone stripes derived from the aerial images.
peaks, even though ground temperatures below freezing can sporadically extend to much lower elevations (down to 2700-3000 m). We observed needle ice (3-5 cm long) mainly along water-saturated stream banks at places with cold air ponding. Needle ice is a typical small-scale example for diurnal freeze-thaw cycles in the Bale Mountains as it forms at clear nights throughout the dry season. Interestingly, we also found evidence for seasonal frost phenomena. Up to 10 m high frozen water falls evolve at shaded north-exposed cliffs in the Wasama Valley during the dry season and last until the onset of the following wet  . 5c). While the stone stripes are trough-shaped, the areas with finer material inbetween are more rampart-like (Fig. 5d).
The distance between the stone stripes equals in most cases to the width of the stripes. Typical for some of the stone stripes 15 is that they split up into two narrower branches in the upper part and merge to a single wider branch in the lower part. As the GPR survey suggests, the regolith layer between the stone stripes contains no larger rocks (exceeding several decimetres) and is more than 1.5 m deep (6b). The surface of the underlying solid rock was not detected. All larger rocks (up to 0.5 m wide and 2 m long) are located mainly in the troughs or on top of the regolith layer. On the slightly inclined (2 -9°) western Sanetti Plateau between 3950 and 4150 m, the stone stripes are 300 -1000 m long and mainly 5 -10 m wide (Fig. 5b). Most 20 of the stripes are connected to heavily eroded cliffs. In the upper part, some of the stripes split up into multiple branches and where the plateau flattens, a transition from sorted stone stripes to less developed stone circles is visible in the field, but hardly recognisable on satellite images.
The six dated rock samples from two different locations on the Sanetti Plateau originate from basaltic (BS01-04) and trachytic (BS05-06) lava flows as it is indicated by the varying alkali and silica contents (Table C1). We obtained very high 36 Cl con-25 centrations, especially for the two trachytic samples (>120 × 10 6 At g -1 ) from the western part of the plateau (Table C2). In these two samples (BS05 and BS05), 36 Cl has reached saturation. This means that the production and decay of 36 Cl average out. Since the resulting exposure ages (>1000 ka) are at the limit of the method, they are not explicitly stated in the figures and tables. Based on the remaining samples, we calculated non-erosion-corrected 36 Cl surface exposure ages of 84 ± 4, 281 ± 12, and 281 ± 13 ka for the southern and of 620 ± 13 ka for the western stone stripes (Table C2). However, due to the high 36 Cl 30 concentrations, an erosion rate of >1 mm ka -1 would lead to considerably older exposure ages for all samples except BS01.
The "old" ages conflict with a relatively young formation age (e.g. global LGM or postglacial) as suggested by the morphology and hardly weathered surface of the investigated angular and columnar boulders. Long-term exposure of the sampled rocks to 36 Cl-producing cosmic rays prior to the formation of the stone stripes could explain this mismatch. Despite the high 36 Cl concentrations, a temporary ice cover overlying the stone stripes for several thousand years during the last glacial cycle cannot  be entirely ruled out from the exposure dating alone. A meter-thick ice cover would reduce the production rate, but a period of several thousand years would not be sufficient to affect the 36 Cl concentrations noticeably or zero the inheritance. However, a temporary ice cover overlying the stripes seems unlikely in light of the absent field evidence for such a scenario.

Present frost occurrence and ground temperature variations
The observed present-day ground temperatures in the Bale Mountains show characteristic daily and seasonal variations, but are 5 in general way off from permafrost conditions (Fig. 7). At the location of the stone stripes on the southern Sanetti Plateau, the mean multiannual ground temperature between the surface and 50 cm depth is 11°C. On top of the highest peak Tullu Dimtu, the mean annual ground temperature is 7.5°C. The mean air temperature at the same location is 2°C and therefore about 5.5 18 https://doi.org/10.5194/esurf-2020-53 Preprint. Discussion started: 29 June 2020 c Author(s) 2020. CC BY 4.0 License. 19 https://doi.org/10.5194/esurf-2020-53 Preprint. Discussion started: 29 June 2020 c Author(s) 2020. CC BY 4.0 License.
K lower than the mean ground temperature. While the daily ground temperature range is largest near the surface and decreases with depth, the seasonal variations in all depths follow a similar cycle (Fig. 7). On the plateau, the ground cools down during the dry season (Bega) and warms up during the wet seasons (Belg and Kiremt). The difference between the seasonal minimum and maximum of daily mean temperatures over a year is about 10 K near the surface, 6 K in 10 cm, and 2.5 K in 50 cm depth (in Fig. 7, daily mean ground temperatures are only presented for 50 cm depth). This shows that seasonal temperature variations 5 can also be of relevance for tropical mountains with a pronounced diurnal climate. The time series is far too short for deriving any long-term trends, but the interannual ground temperature variability observed during the three-year period (2017 -2020) was rather low (<0.5°C).
Near the surface, the diurnal ground temperature amplitude is well pronounced and varies on average between 10-20°C during the wet and 20-30°C during the dry season (Fig. 8a). Extreme temperatures of up to 45-50°C during cloudless days m is similar to the annual lapse rate obtained for the plateau (0.70°C per 100 m) and northern declivity from measured air temperatures (Fig. 9). Interestingly, the lapse rate obtained for the Harenna Escparment is less steep (0.62°C per 100 m) and might represent wetter conditions and pronounced cloud formation at the southern declivity between~1500-3800 m compared 20 to the drier plateau and northern declivity. However, the number of AWS below the afro-alpine belt is not sufficient to determine unequivocally elevations where the mean annual lapse rate changes from a more or less dry adiabatic to a moist adiabatic and vice versa. Distinct elevational changes in the lapse rate could indicate the mean annual condensation level as well as the upper atmospheric storey where dry north-easterly trade winds dominate.
The comparative experiment on Mount Wasama shows clear differences between the thermal regime of the southern and 25 northern slope (Fig. 10a). The southern slope is on average more than 2°C warmer and reveals a more pronounced seasonality and larger diurnal amplitude which favours freezing and thawing and might explain the presence of solifluction lobes. While the mean daily temperature at the southern slope peaks towards the end of the dry season (January to February) when the sun is in its zenith, it reaches its maximum at the northern slope a few month later when the sun approaches its northernmost position.
The ground temperature differences between vegetated and unvegetated areas on the Sanetti Plateau are less obvious (Fig. 10b).

30
Small Erica trees and bushes buffer the diurnal temperature amplitudes of the shaded ground, but have only little impact on the seasonality. Like at Mount Wasama, both south exposed locations at Tullu Dimtu (vegetated and unvegetated) have their temperature maxima at the end of the dry season, whereas the vegetated and unvegetated locations in the plain heat up during June/July. The cross-comparison between low-cost and high-quality data loggers on top of Tullu Dimtu revealed a promising relationship (R 2 = 0.98) and proved that the tested low-cost loggers, which have not been developed explicitly for scientific 35 20 https://doi.org/10.5194/esurf-2020-53 Preprint. Discussion started: 29 June 2020 c Author(s) 2020. CC BY 4.0 License.   applications, are suitable for short-term (< 1 year) ground temperature measurements and experiments at high elevations. Both loggers measured nearly the same mean ground temperature (8.46 vs. 8.48°C). Only the standard deviation of the low-cost logger was a bit larger (9.1 vs. 7.3°C) since it was installed minimal closer to the surface.

Modelled palaeo ground temperatures
At the three locations on the Sanetti Plateau (Tullu Dimtu, EWCP Station, and Tuluka), where ground temperatures and a set of meteorological variables were measured concurrently, the ground temperature is mainly controlled by air temperature and global radiation (Fig. 11). The two variables can explain together about 75 ± 3 % of the ground temperature variance ( is not surprising in view of the non-consideration of ground moisture. Precipitation, relative humidity, and wind speed affect ground moisture as well as evaporation and therefore alter the energy balance at the surface and energy transfer into the ground. Measuring and considering ground moisture (directly at the AWS) would likely help to reduce the uncertainties of the applied multiple linear regression models (RMSE of 3-4°C). Nevertheless, the established relationship between ground temperature and air temperature/global radiation is strong enough to use the models for a first palaeoclimatic reconstruction experiment.

10
To obtain mean annual ground temperatures associated with deep seasonal ground frost or sporadic permafrost (T ground < -1°C ) at the elevation of the lowermost stone stripes on the southern Sanetti Plateau during their time of formation, the three tested linear models require a mean ground temperature depression of~12°C. This would translate into a mean air temperature depression of 7.6 ± 1.3°C (the error is the standard deviation of the three model outputs) which would imply a mean annual air temperature on the southern plateau of -1.9 ± 1.3°C. The deduced stronger decrease of the ground temperature over the air 15 temperature is due to the observed modern relationship. A cooling/warming of the air of 1°C relates to a decrease/increase of the ground of 1.6-1.9°C and vice versa (see Table 4 and Fig. 11). The geophysical reasons for this statistical relationship can be manifold. They are associated with the radiative forcing and energy exchange between the atmosphere and ground, which in turn is affected by many factors ranging from insolation, air pressure, relative humidity to the thermal conductivity, specific heat capacity, density, humidity, albedo, etc. of the ground. 20 Over a year, the seasonal mean daily ground temperature fluctuations between the surface and 50 cm depth were theoretically large enough to freeze and thaw the upper half metre of the ground (Fig. 12). However, seasonal freezing and thawing below 50 cm depth seems unlikely in the past if the ground properties and seasonal temperature fluctuations were similar like today.
A mean annual ground temperature in the order of -1°C seems critical for the formation of deep seasonal frost on the Sanetti Plateau since much warmer temperatures would prevent seasonal freezing and lower temperatures seasonal thawing.  annual present-day ground temperature between the southern stone stripes (~11°C) is way off from permafrost conditions.

5
Experimental modelling suggests that a distinct ground temperature depression of~12°C and air temperature depression of 7.6 ± 1.3°C would be necessary for the formation of deep seasonal frost at the elevation of the stone stripes. The main aim of the following discussion is to elaborate when and how the large structures may have formed and what their presence implies for the palaeoclimate and palaeogeoecology of the tropical Ethiopian Highlands.
In contrast to small-scale features which do not necessarily require mean annual air temperatures below freezing, large sorted 5 patterned grounds like stone circles and polygons that are well-documented for the High Arctic (e.g. Kessler and Werner, 2003;Hallet, 2013) occur commonly in permafrost areas with mean annual air temperatures of -4 to -6°C (Goldthwait, 1976). The only other location worldwide where stone stripes in the same order of magnitude as on the Sanetti Plateau and even larger have been described are the non-volcanic Falkland Islands in the South Atlantic (André et al., 2008).
The vernacular term for extensive blockstreams and stone stripes in the Falkland Islands is "stone runs". Stone runs cover is the coexistence of coarse and fine-grained material (large angular blocks and regolith) and the evidence for glaciations and colder conditions during the Pleistocene (Clapperton, 1971;Clapperton and Sudgen, 1976;Groos et al., in revision).
The origin and genesis of the stone runs in the Falkland Islands has been discussed controversially over the last one hundred Pleistocene. Nevertheless, the understanding of the physical processes underlying the frost-related sorting of such large blocks is still fragmentary (Aldiss and Edwards, 1999). The limited process understanding, the small number of analogies worldwide, and the lack of a cross-section profile complicate the interpretation of the stone stripes on the Sanetti Plateau in the Bale Mountains, but the following observations and findings suggest that deep seasonal frost and periglacial processes played a major role in the formation of this landform. 20 As aforementionend, the Bale Mountains were covered by an extensive ice cap and experienced a pronounced cooling of 4-6°C between 50-30 ka (Ossendorf et al., 2019;Groos et al., in revision). It is important to note that the stone stripes on the western and southern Sanetti Plateau are located beyond the glacial remains and the assumed maximum extent of the former ice cap. The obtained 36 Cl surface exposure ages of >600 ka ( 36 Cl has reached saturation in two samples) for the western and of 84 ± 4, 281 ± 12, and 281 ± 13 ka for the southern stone stripes predate the local last glacial maximum. However, the ex-25 posure ages probably do not represent the timing of formation or stagnation of the features. Since the sampled trachytic blocks and columnar basalt originate from eroded cliffs and volcanic plugs, it is likely that they were exposed to cosmic radiation during and prior to the formation of the stone stripes. The well-preserved morphology, the absence of erratic boulders in the surrounding area, and the high 36 Cl concentrations indicate that the stone stripes have not been eroded and deformed by a dynamic and warm-based glacier. However, we cannot compeletely rule out that stagnant, cold-based, and relatively shallow 30 ice covered the stripes for several thousand years. Such an ice cover would not have been sufficient to zero the inheritence (high 36 Cl concentrations). We interpret the hardly weathered surface of the angular blocks and little reoccupation by vegetation as an indication for a younger formation stage of the stone stripes coinciding probably with the coldest and driest phase in Africa (30-15 ka) during the last glacial period (e.g. Tierney et al., 2008). The ice extent in the Bale Mountains at~18 ka was slightly smaller compared to the local LGM despite the general cooling trend in Eastern Africa. The lack of any evidence for a major 35 glacier advance between the local LGM and the~18 ka stage might be indicative of a cold and dry climate which provides ideal conditions for periglacial processes.
A precondition for the formation of patterned grounds is cyclic freezing and thawing of a decimetre-to meter-thick layer and the coexistence of fine-grained material and larger stones or blocks (Kessler et al., 2001;Kessler and Werner, 2003). Both large blocks and fine-grained material are present on the Sanetti Plateau. A decimetre-thick regolith layer rich in silt and loam 5 covers the underlying bedrock of the plateau (Lemma et al., 2019). Whether the regolith has developed during the Pleistocene or during warmer periods before, as suggested for the Falkland Islands, remains unclear. The accumulation of trachytic blocks and columnar basalt at some places on the plateau is probably related to intensive frost wedging at cliffs and volcanic plugs during the last glacial cycle. As our ground temperature modelling suggests, a half-meter-thick frozen layer could have formed seasonally on the plateau outside the glaciated area as the result of a pronounced air temperature cooling of 7.6 ± 1.3°C during 10 the coldest period of the last glacial cycle. Deep frozen grounds and sporadic permafrost still exist at some of the highest tropical and subtropical mountains in Africa (Kaser et al., 2004;Vieira et al., 2017). of individual narrower branches to wider single stripes is remarkably similar to patterned grounds at gentle slopes predicted by numerical models after several hundred freeze-thaw cycles (Werner and Hallet, 1993;Mulheran, 1994;Kessler et al., 2001;Kessler and Werner, 2003). Such numerical models can reproduce the self-organization of different sorted grounds by varying just a few parameters (mainly stone concentration, hillslope, and degree of lateral confinement) and need about 500 to 5000 freeze-thaw cycles to form similar stripe patterns as found on the southern Sanetti Plateau (Fig. 5). Less cycles would lead to a 25 more random configuration and more cycles would eliminate the smaller branches and lead to a "perfect" sorting of the stripes (see Fig. 2 in Werner and Hallet, 1993). Assuming downslope displacement rates of 10-50 cm per year (or cycle) for clasts as it is observed for small-scale periglacial features in the tropics (Francou and Bertran, 1997)  climate records from Eastern Africa and worldwide (e.g. Jouzel et al., 2007;Tierney et al., 2008). Moreover, the reconstructed temperature decrease from the Ethiopian Highlands in comparison with the reconstructions from the lower-situated Congo Basin and Lake Tanganyika in the order of 4 to 4.5°C (Weijers et al., 2007;Tierney et al., 2008) reveals an amplified tropical 15 cooling at high elevations in Eastern Africa during the global LGM. Strong evidence for such a cooling is also provided by sea surface and air temperature reconstructions from different lakes along an elevational transect in Eastern Africa. Loomis et al.
(2017) explain the observed elevational trend with a steeper tropical lapse rate (Γ LGM = 6.7 ± 0.3°C km -1 vs. Γ modern = 5.8 ± 0.1°C km -1 ) related to a drier atmosphere during the global LGM. The present-day mean annual lapse rate and ground temperature gradient on the Sanetti Plateau of 0.7°C per 100 m ( Fig. 8 and 9) is larger than the palaeo (Γ LGM ) and modern (Γ modern ) East

20
African lapse rate between 474 and 3081 m. This emphasises that attention should be drawn to temporal as well as vertical changes in the lapse rate when reconstructing or simulating the palaeoclimate of the tropics. Despite improvements, global climate models still tend to underestimate the cooling at high elevations in the tropics during the last glacial cycle (Loomis et al., 2017).
The palaeoclimatic and environmental findings presented here have direct implications for the settlement history and ecology 25 of the Bale Mountains. Latest archaeological excavations at 3469 m in the northwestern part of the Bale Mountains along with biogeochemical, zoogeographical, and glacial chronological investigations reveal that Middle Stone Age foragers resided in the highlands already between 47 to 31 ka and made use of the available alpine resources (Ossendorf et al., 2019). Why the residential site was abandoned after 31 ka is unclear, but might be related to a gradual cooling and dessication of the highlands until the global LGM as suggested by the large-scale periglacial landforms and lack of major glacier advances between 50-30 30 and~18 ka (Groos et al., in revision). Moreover, the temperature depression, ice cover, and periglacial conditions must have also affected the habitat of endemic mammal species like the Ethiopian wolf, giant mole-rat, and mountain nyala that currently populate the Sanetti Plateau and upper valleys (Miehe and Miehe, 1994). Due to the absence of any evidence for glacial extinction events in the region, we conclude that endemic plants and mammals as well as Middle Stone Age foragers coped with the harsh climatic and environmental conditions in the Ethiopian Highlands during the Pleistocene. To reduce the uncertainty of the statistical model applied for ground temperature simulations and air temperature reconstructions, considering the impact of moisture on the thermal conductivity and heat capacity of the ground as well as energy fluxes into the ground is necessary. It is possible that ground moisture and stagnant water played a more important role during the Late Pleistocene than today due to (perma)frost-induced waterlogging and perennial melting of snow and ice. Why the relict 10 patterned grounds are restricted to the southern and western Sanetti Plateau can be explained by the assumed preconditions for their formation: a relatively flat and unglaciated terrain, presence of coarse and fine grained material, deep ground frost, and absence of a thick snow layer.

Conclusions
This contribution provides further evidence that the tropical Bale Mountains in the southern Ethiopian Highlands were subject 15 to severe climatic and environmental changes during the Late Pleistocene. Both glacial and periglacial processes have shaped the afro-alpine environment. Compared to the modern nocturnal and seasonal frost phenomena, relict periglacial landforms like blockfields along the Harenna Escarpment and sorted stone stripes on the Sanetti Plateau are much larger and more developed.
The large sorted stone stripes are exceptional for the tropics and probably formed under periglacial conditions in proximity of the palaeo ice cap on Tullu Dimtu during the coldest period(s) of the last glacial cycle. We hypothesise that the slightly 20 inclined and unglaciated areas of the Sanetti Plateau, the coexistence of regolith and large blocks, the occurrence of deep seasonal frost, as well as relatively dry conditions beyond the ice cap provided ideal conditions for frost heave and sorting and the formation of large patterned grounds. Based on our ground temperature measurements and modelling experiment, we propose a distinct ground temperature depression of~12°C and air temperature depression of 7.6 ± 1.3°C as precondition for the formation of deep ground frost on the Sanetti Plateau. The novel idea of using a statistical relationship between measured 25 present-day ground temperatures and meteorological variables to assess past climatic changes through the simulation of palaeo ground temperatures has also potential for other regions where relict frost-related periglacial landforms exist. Comparing the reconstructed air temperature depression from the Bale Mountains with climate records from lower elevations in Eastern Africa emphasises a strongly amplified cooling at high elevations that has already been outlined for many other tropical mountains.
Such a cooling in tandem with the extensive glaciation and frost action must have dramatically affected the habitat of endemic 30 mammal species like the Ethiopian wolf, giant mole-rat, and mountain nyala that currently populate the Sanetti Plateau and upper valleys of the Bale Mountains. Attention should therefore be given to the amplified middle troposphere cooling when reconstructing and modelling climatic and geoecological changes in the tropical mountains of Eastern Africa.
Data availability. The ground temperature data (Tables S1-4) can be downloaded from the open access library PANGAEA (provisional link until data are published : https://filesender.switch.ch/filesender/?vid=06bf00d3-d750-e748-82f8-00006e12ef78). Raw data, aerial images, additional field photos, etc. are available upon request by email to the corresponding author.
Author contributions. ARG, NA, and HV designed the research concept, conducted the geomorphological mapping, sampled the stone stripes for exposure dating, and installed the ground temperature loggers. ARG and NA processed the rock samples in the laboratory. FH set up the weather stations. LW conducted the GPR measurements and serviced the weather stations. FH, LW, and TN processed and provided the meteorological data. ARG and JN processed the ground temperature data, conducted the statistical analysis, and performed the ground temperature simulations. ARG drafted the manuscript and figures with contributions from all authors.   Groos et al. (in revision). *Erosion rate = 0 mm ka -1 . **Erosion rate = 1 mm ka -1 . ***Erosion rate = 2 mm ka -1