Changing flood frequencies under opposing late Pleistocene eastern Mediterranean climates

Floods comprise a dominant hydroclimatic phenomenon in aridlands with significant implications for humans, infrastructure, and landscape evolution worldwide. The study of short-term hydroclimatic variability, such as floods, and its forecasting for episodes of changing climate therefore poses a dominant challenge for the scientific community, and predominantly relies on modeling. Testing the capabilities of climate models to properly describe past and forecast future short-term hydroclimatic phenomena such as floods requires verification against suitable geological archives. However, determining flood frequency during changing climate is rarely achieved, because modern and paleoflood records, especially in arid regions, are often too short or discontinuous. Thus, coeval independent climate reconstructions and paleoflood records are required to further understand the impact of climate change on flood generation. Dead Sea lake levels reflect the mean centennial-millennial hydrological budget in the eastern Mediterranean. In contrast, floods in the large watersheds draining directly into the Dead Sea, are linked to short-term synoptic circulation patterns reflecting hydroclimatic variability. These two very different records are combined in this study to resolve flood frequency during opposing mean climates. Two 700-year-long, seasonally-resolved flood time series constructed from late Pleistocene Dead Sea varved sediments, coeval with significant Dead Sea lake level variations are reported. These series demonstrate that episodes of rising lake levels are characterized by higher frequency of floods, shorter intervals between years of multiple floods, and asignificantly larger number of years that experienced multiple floods. In addition, floods cluster into intervals of intense flooding, characterized by 75% and 20% increased frequency above their respective background frequencies during rising and falling lake-levels, respectively. Mean centennial precipitation in the eastern Mediterranean is therefore coupled with drastic changes in flood frequencies. These drastic changes in flood frequencies are linked to changes in the track, depth, and frequency of mid-latitude eastern Mediterranean cyclones, determining mean climatology resulting in wetter and drier regional climatic episodes.

In the eastern Mediterranean, mean hydrological changes in the Dead Sea watershed are recorded by changing lake levels, reflecting decadal to centennial hydroclimatic trends 1,15 (Fig. 1b,c). Dead Sea level rises (falls) were shown to reflect increased (decreased) annual precipitation, and are strongly correlated with increased (decreased) frequency of mid-latitude eastern Mediterranean cyclones (EM cyclones) [15][16][17] . The sedimentary record of the Dead Sea therefore provides an opportunity to test whether changes in mean precipitation are also associated with changes in flood frequency. In this paper, we present evidence corroborating the notion that flood occurrence in the EM is strongly correlated with mean climatology and that flood occurrences are non-stationary and comprise intervals of intense flooding due to changes in the track, frequency and depth of mid-latitude EM cyclones [18][19][20] (Fig. 2).

Flood climatology, lake levels and the geological record of the Dead Sea
The mean hydrological budget of the Dead Sea and its level are controlled by EM cyclones delivering ca. 90% of annual precipitation to the northern watershed, currently reaching the lake through its main water source, the perennial Jordan River 15,17,19,21,22 . EM cyclones also induce rainstorms that affect tributaries draining directly into the Dead Sea, capable of producing floods 23 (Fig. 1d). Under modern conditions, 15-20 EM cyclones affect the northern wetter parts of the eastern Mediterranean during 45-60 days of the rainy season (October-May) 21,24,25 , whereas the arid southern and eastern parts of the watershed experience up to seven precipitating storms on average, each lasting up to two days 26 . Other synoptic-scale systems capable of generating floods are Active Red Sea troughs Figure 1. (a) Regional satellite image of the study area (extracted from The Blue Marble Next Generation, NASA's Earth Observatory 11 ) (b) Reconstructed lake level curve of late Pleistocene Lake Lisan (upper panel) 12 , and age depth model for the Inernational Continental Scientific Drilling Program Dead Sea Deep Drilling Project (ICDP-DSDDP; lower panel) 13 . (c), Schematic map with ICDP-DSDDP coring location, modern mean annual rainfall, perennial streams, and the maximum extent of Lake Lisan (at 25 ka). (d) Satellite image of the Dead Sea and its major tributaries (created using ArcGIS ® software by Esri, and includes World Imagery basemap 14 . Sources: Esri, DigitalGlobe, GeoEye, Earthstar Geographics, CNES/Airbus DS, USDA, USGS, AeroGRID, IGN, and the GIS User Community). Also depicted are the margins of Lake Lisan during the studied intervals at lake levels of 190 and 280 (m below mean sea level), depicted by blue and red lines respectively. Note the large regression of the northern shoreline during the studied time interval, whereas the eastern and western shorelines only slightly differ from the modern Dead Sea (lake level 430 mbmsl).

Results
In this study, micro-facies analysis of two sections of the DSDDP core (ca. 27 and 18 ka) reveal for the first time that these clastic laminae are comprised of several graded sub-laminae, indicating discrete and countable floods within a single wet season (Fig. 3). Significant differences in flood frequency characterize the opposing climates, reflected in lake level rises and falls during the late Pleistocene. Episodes of rising lake levels are characterized by higher frequency of floods, shorter intervals between years of multiple floods, and a significantly larger number of years that experienced multiple floods ( Fig. 4; Tables S1, S2). The two studied intervals (core 5017-1-A, sections 48-2 and 51-1, 18 and 27 ka, respectively; Fig. S2) and their respective flood time series demonstrate pronounced non-stationarity, dividing the studied intervals into flood-rich clusters and background intervals (Figs 4, S3). Clusters are characterized by pronounced increased flood frequency compared with their respective background intervals, with significantly higher flood frequencies during rising lake level (Table S2). Several key differences are noted: (a) significantly more detrital laminae with ≥2 floods per year during lake level rises (27% vs. 12%; p-value < 0.01; Table S1), representing a 30% surplus in recorded floods during lake level rise (1104 vs. 836; Table S1). (b) Years with ≥2 floods per year cluster into intervals of similar mean duration (ca. 65 yr) during both lake level falls and rises. (c) Flood clusters are marked by +100% and +40% increased flood frequency relative to background levels, during level rise and level fall, respectively. (d) The average time interval (∆ ) between consecutive years characterized by ≥2, ≥ 3, and ≥5 floods per year are 1.5-2 times longer during background periods compared with flood clusters, and more importantly, they are 2-4 times longer during lake level fall than during lake level rise (Figs 4, S3; Table S2).

Discussion
In this study, microfacies analyses of Dead Sea varves disentangle the complex interplay between global climate change, synoptic circulation patterns and past flood hydrometeorological variability. Intra-seasonal detrital sub-laminae reveal individual floods, deposited within a single wet season during the late Pleistocene for the first time (Fig. 3). Consideration of the strong seasonal Mediterranean climate, in which rainfall is limited to the wet-season 17,61 , point to the following hydroclimatic-lacustrine sedimentary sequence: 1. Intense rainstorms in large tributaries east and west of the Dead Sea produced floods of sufficient intensity and volume, capable of carrying suspended load into the ICDP-DSDDP coring location at the lake center (Figs 1d, S1) 20,27,62 . Establishing the threshold of these floods remains beyond the scope of this paper, but it is suggested that they were larger than common gauged floods in the modern record (Fig. S4). 2. Coarse bedload was deposited in the sub-lacustrine canyons and tributary mouths, forming coarse-clastic fan-deltas and shore deposits 63,64 , whereas the silt-dominated suspended load was carried as sediment plumes towards the center of the lake, where it was deposited as thin (ca. 20-500 μm) graded sub-laminae 55,57,65 (Figs 3; S5-6). 3. Most sub-laminae are capped by organic-rich laminae of few μm thickness, resulting from the deposition of suspended terrestrial plant remains, coupled with proliferation of epilimnion algae in response to incoming nutrient-rich floodwaters, diluting the saline lake waters, as reported in rare modern observations [66][67][68] (Fig. 3). 4. Incoming floodwaters replenished the lake's epilimnion with bi-carbonate and other salts, supporting the lake's primary productivity and enabling aragonite precipitation during the subsequent dry season 53,54,69,70 .
Among the potential flood-producing synoptic circulation patterns, EM cyclones are the most likely candidates for generating high-sediment yield floods, and are limited between October and May 23,27,62 (Fig. 2). During the Last Glacial Maximum (LGM), lake-level reached a maximum of 160 mbmsl 1,12 (ca. 250 above modern level), and the northern and southern shores of Lake Lisan were ca. 150 and 90 km away from the coring site, respectively (Figs 1, S2). Furthermore, the epilimnion of Lake Lisan was significantly fresher (0-100 TDS) 4 than modern Dead Sea brine (240 TDS) 71 . Thus, fine-grained flood-borne detritus could have reached the coring site only by floods from the larger tributaries entering the lake from the steep eastern and western escarpments 37 (Figs 1d, S5-6).
Because mean hydroclimatic conditions and large floods in the eastern and western tributaries are generated by EM cyclones 15,16 , mean climatic conditions can now be associated with changes in the frequency of flood-producing EM cyclones during the LGM 23,27 . More specifically, EM cyclones have distinguished dry (divergence) and precipitating (convergence) regions, separated by a belt of maximal wind speed 18 (Fig. 2a). Major precipitation events over the Dead Sea watershed occur when: (a) the cyclone is characterized by a southern track (i.e. the cyclone center passes through the EM towards the Syrian desert; Fig. 2a) 20 ; or (b) the cyclone is deep enough so its effect extends south into the arid regions of the Negev desert (Fig. 2b) 19,21 . Shallow cyclones and cyclones characterized by a northern track, on the other hand, only affect the northwestern EM, with negligible impact on the Dead Sea watershed and its tributaries (Fig. 2c,d). Under modern conditions, these headwaters (at altitude of 600-800 m) are occasionally covered by snow, and since late Pleistocene winters were colder 72,73 , with more frequent snowfall in lower altitudes, EM cyclones were prone to generate snow and rain over snow at the southern mountainous areas, east and west of the Dead Sea, thus facilitating sediment transport into the deep lake (Figs S5-6).  During wetter conditions, characterized by increased frequencies of deep and southern-track EM cyclones, frequencies of years with ≥2 floods per year in both background and cluster intervals significantly increased, meaning that more floods reached the coring site in the lake's center within a single season (Tables S1-2). In contrast, when the frequency of EM cyclones decreased and lake levels dropped accordingly, frequency of years with ≥2 floods per year decreased as well. Regionally, the well-established association of increased flood frequency with lake level rise during the last glacial supports earlier assertions that level rises in this basin are primarily the product of increased precipitation.
The increased (decreased) frequency of recorded floods per year and their causative synoptic-scale atmospheric patterns during rising (falling) lake levels illustrate that during the LGM, flood hydrometeorology is embedded in-and coupled with mean hydrological conditions and climatology of the eastern Mediterranean. Hence, significantly more years experienced multiple floods during lake level rise (191 years out of 701) than during lake level drop (82 years out of 708; p-value < 0.01; Table S1). Finally, some of the studied wet winters present significantly larger number of floods per year (>10) relative to modern observations (Fig. 4), demonstrating the essential role of paleoflood records in complementing modern hydrometeorological records for improving our understanding of hydrometeorological variability. Clustering of wet seasons with ≥2 floods demonstrates that the effect of climate change on in situ hydrometeorological variability and flood occurrence is non-stationary and non-linear.
The study of weather, synoptic circulation patterns and the hydrological cycle in situ during global climate change is predominantly limited to modeling. Testing climate models capabilities in capturing hydroclimatic variability such as short-term flood clustering should involve verification against suitable records. Analyses of the seasonal to multi-annual record of Lake Lisan reveals the underlying processes comprising centennial scale climatology. This information can be introduced into climatic models to improve understanding and quantification of natural weather and climate variability. Because high-resolution archives of hydroclimatic variability are scarce, while reconstructed lake levels often smooth high-resolution hydroclimatic variability, this record can be further exploited to interpret and to model short-term hydroclimatic variability through the entire watershed during episodes of climate change.

Materials and Methods
Age-Depth model of the ICDP DSDDP Core 5017-1. Site 5017-1 of the ICDP DSDDP is located in the deepest floor of the Dead Sea ( Fig. S1; 31°30′28.98″N\35°28′15.60″E). The core was extracted by the U.S. Drilling, Observation, and Sampling of the Earth's Continental Crust (DOSECC) using the Deep Lake Drilling System on a barge from November 2010 to January 2011. The composite 5017-1 profile comprises authigenic halite, gypsum and aragonite as well as clastic material 42 (Fig. S2). Some muddy segments of the core, primarily during MIS2-4, comprise alternating aragonite (authigenic) and detritus (alochtonous) forming annual varves 50,74 , in places disturbed by slump sediments 44 . Aragonite precipitates from lake waters during dry season evaporation, and detritus is deposited by floods during the rainy season 45,55,57 . Age-depth model for the ICDP 5017-1 core is based on 54 calibrated 14 C ages using Markov chain Monte-Carlo simulation, adapted to provide a monotonic smooth spline curve (95.4% confidence interval) 13 .
Thin sections of the ICDP-DSDDP cores. Two core sections (51-1 and 48-2) of the ICDP Dead Sea Deep Drilling Project (ICDP-DSDDP) core (Fig. S2), coeval with significant lake level rise and fall, as determined by radiocarbon chronology 13 , were selected for this study. Core sections selection was based on the availability of non-disturbed varved segments of sufficient length that also had good age constrains, with 14 C ages extracted from within the studied core sections (Fig. S1). The sections represent similar time spans of approximately 700 years (708 and 701 years for lake level drop and rise, respectively) and were continuously sampled for thin sections using the standard procedure 75 adjusted for salty sediments. In total, 33 large-scale thin-section samples (10 × 2 cm) were prepared; 18 sections representing lake level rise, and 15 lake level fall. The thin sections were scanned and examined using a petrographic polarized microscope (Leica DMLP) and fluorescence microscopy operated with violet and polarized light conditions (Nikon AZ100M). Images were taken with a digital camera (Olympus DP72) and processed with Nikon photo software (NIS Elements AR 4.3). These segments contain both laminated varves of alternating aragonite and detritus and mass transport layers, related to slope instability. We avoided the mass transport layers and focused on annually laminated segments to properly consider the annual sedimentary cycle and establish flood frequencies (Figs 3, S2).

Varve counting and microfacies analyses.
Varve counting was carried out under plane-and crosspolarized microscope. The common varve comprises a graded sequence capped by a thin lamina of organic-rich material. However, some detrital laminae are contain more than one graded sequence and show a multiple sub-laminae structure, which represents repeated individual floods. To avoid over-identification of sub-laminae, all sections were examined twice, and the counting of sub-laminae was conducted in a conservative fashion, so when the number of sub-laminae was questionable (within ±1 sub-laminae), the lower number of sub-laminae was used. Due to the varved nature of the studied laminated segments, a minimum number of one flood per wet season is assigned to each varve. Although we assume that some years probably experienced floods too small to reach the coring location, the nature of the sediments, does not allow an objective identification of years without any floods. This limitation results in inherited bias of the data were possible zero-floods laminae were assigned a minimum value of 1. We therefore adapted our statistical analyses and focused on years with ≥2 floods per year.
Statistical analyses. In total, 708 and 701 varves were counted in the segments representing lake level drop and rise, respectively. The populations of the episodes representing rising and dropping lake levels were compared with Matlab© 76 and R 77 using Kolmogorov-Smirnov 78 , Mann-Whitney (Wilcoxon) 79 , Ansari-Bradley 80 , chi-squared 81 and logistic regression 82,83 tests to determine the significance of the difference between the two segments. The difference between the populations is significant in all of the above-mentioned tests. A negative regression coefficient of −0.51 in the logistic regression model further emphasizes that the probability for a detritus laminae to be associated with period of rising lake level increases with the number of sub-laminae. Identification of clusters was carried out in a semi-objective approach with the consideration of running Mann-Whitney and Ansari-Bradley tests with a window width of 75 yrs (Fig. S3), along with the moving average and its first and second derivatives.

Chi2 (on lumped contingency
Analysis of Deviance