ASSESSMENT OF CONTRIBUTING FACTORS OF THE MAY 2021 DISASTERS IN TAJIKISTAN A forensic study under the Strengthening Critical Infrastructure against Natural Hazards Project (SCINHP, P158298) Assessment of Contributing Factors of the May 2021 Disasters in Tajikistan A forensic study under the Strengthening Critical Infrastructure against Natural Hazards Project (SCINHP, P158298) Authors: Joel Fiddes WSL Institute for Snow and Avalanche Research SLF, Davos, Switzerland Simon Allen Department of Geography, University of Zurich, Switzerland Holger Frey Department of Geography, University of Zurich, Switzerland Photo on front page: Damages in Kulob on May 11, 2021. Photo: Radioli Ozodi (RFE/RL) This assessment was developed with support from the Japan-World Bank Program for Mainstreaming DRM in Developing Countries, which is financed by the Government of Japan and managed by the Global Facility for Disaster Reduction and Recovery (GFDRR) through the Tokyo Disaster Risk Management Hub. 2 Headlines This document reports the findings of a forensic assessment of the flood, mudflow and mass movement disaster which struck Tajikistan in the period of May 7- 13, 2021. The study has been conducted by the University of Zurich (UZH), Switzerland and the WSL Institute for Snow and Avalanche Research (SLF), Davos, Switzerland, under the Strengthening Critical Infrastructure against Natural Hazards Project of the World Bank Group, with support from the Japan-World Bank Program for Mainstreaming DRM in Developing Countries. This report is based on a desktop, forensic assessment, using satellite information, global hydrological and topographic data, ground station data and model results. The list below summarizes the headlines of this report and indicates in which sections the related explanations and findings are presented in the report: Satellite imagery analysis and exploration of glacier and permafrost data revealed that => there is no evidence for mass movements originating from within starting zones in glacial or periglacial environments during this event (except snowmelt discharge and related higher soil moisture). [Section 2] Analysis of pre- and post-event satellite imagery and landuse/landcover data suggest that => most events originated from reactivated, existing mudflow channels. Damages seem to have been enhanced by increased exposure over the past decades. [Section 2] Remote sensing based land-use change analysis indicate => an increase in bare ground over the past 3 decades in many of the affected districts (potentially driven by overgrazing and/or deforestation), which is likely to have contributed to enhanced runoff and erosion during the heavy rainfall event in May 2021. [Section 2] By comparing combined satellite/ gauge precipitation products (from GPM IMERG) and data from Tajik Met stations with climate reanalysis data (from ERA5-land), it was found that => the precipitation event leading to the disaster was extraordinary, i.e. above the 95/99 percentiles of the 1979-2020 precipitation rates. [Section 3] By combining satellite data (MODIS) and a dedicated snow model, the contribution of snowmelt to the event was assessed and found that => snow cover and snow melt contribution to runoff was likely above the climatic average, but not extreme. [Section 3] Regional climate model results from WCRP CORDEX for Central Asia indicate => an increase of intense precipitation days (> 20mm/day) with time and greenhouse gas concentration, and therefore an increased likelihood for events similar to May 2021 over the 21st century. [Section 4] Previous studies on glacier evolution and related impacts on runoff and glacial hazards suggest that => peak river runoff can shift by up to 1 month earlier in the year. This could increase the likelihood that extreme rainfall events coincide with high base flow in the rivers. At the same time, droughts during the summer months can increase. Further, increasing risks from glacial lake outburst floods are expected for the future. [Section 4] 3 1. Introduction Background and aim During the period of May 7-13, 2021, Tajikistan experienced a series of floods, mudflows and mass movements, which led to deaths and damages to buildings and infrastructure. While the Government of Tajikistan has requested international assistance to respond to and recover from these disasters, they were not considered to be significant enough to declare a national emergency. They do, however, reflect the kind of local and recurrent events that are common and likely increasingly frequent in the country, undermining sustainable development. The World Bank is currently implementing the Strengthening Critical Infrastructure against Natural Hazards Project (SCINHP) and is preparing the Tajikistan Resilient Landscape Restoration Project (TRELLIS), Tajikistan Resilient Irrigation Project (TRIP), and the Fifth Phase of the Central Asia Regional Links Program (CARs-5). In addition, the potential of a 2nd phase of SCINHP is being explored. All of these projects will seek, through varying approaches and with different sectoral investment targets, to increase the resilience of Tajikistan against hydrometeorological hazards such as those experienced in May 2021. Within this framework, a forensic assessment of the May 7- 13 disaster has been undertaken by the University of Zurich (UZH), Switzerland and the WSL Institute for Snow and Avalanche Research (SLF), Davos, Switzerland, that is presented in this report. The overarching objective of this study is to perform a primarily remote-based assessment of the processes that led to the May 2021 disaster; assess how extreme the processes that led to this disasters are in the current climatic context; and evaluate the possible impacts of future climatic changes for the frequency and magnitude of such events in the future. Area of Interest Reported damages occurred mainly in the central and southwestern parts of Tajikistan. Based on a combination of different reports from Tajikistan’s Rapid Emergency Assessment and Coordination Team (REACT), 17 districts were identified, where damages as a consequence of the heavy early May rainfalls have been reported (cf. Table 1). These districts defined the initial area of interest for the analyses presented below. Population density is relatively high in the affected districts in the south west of the country, with some particularly densely populated districts, including Dushanbe and Kulob (Fig. 1) A special focus has been drawn on the Kulob district, where the impact of the disaster was an order of magnitude greater than in any other district (see section 2). Topographically, the affected districts can be categorized into three groups: The lowland districts in the south west of the country (Danghara, Farkhor, Jomi, Khurson, Kulob, Muminobod, Qumsangir, Rudaki, Shurobod, Vakhsh, Vose and Yovon), which do not contain relevant mountain zones. The Hissor, Darvoz, Rasht and Varzob districts can be considered as districts with mountain influence (minor glaciation), while Tavildara is the only high mountain district that has been affected by the May 2021 disaster. All the affected districts with mountain areas have low population densities (<50/km2, Tavildara 5.8/km2) (Fig. 1). 4 Figure 1: Population density of Tajikistan based on the Gridded Population of the World, Version 4 (GPWv4) and glacier areas (white) from the Randolph Glacier Inventory (RGI, Pfeffer et al., 2014), Version 6.0, available from GLIMS/NSIDC (http://www.glims.org/RGI/index.html). Districts affected by the May 21 disaster are shown in red, with numbers indicating the number of affected households as reported by REACT (normal font) or the WASH report (italic font). 5 2. Impacts of the May 2021 events and catchment characteristics Key messages: ● Damages from the May event were concentrated in the city of Kulob, where up to 1500 households were affected (next highest was Vakhsh with 217 households, cf. Figure 1), with significant damage visible to transport infrastructure, and partial destruction of flood defense structures. ● There is no evidence of floods or mass movements originating out of the glacial environment during this event. On the contrary, damages are greatest in lower elevation districts, far downstream of glacial catchments. ● Most events originate from reactivated, existing mudflow channels, with no significant new erosion channels identified in Sentinel imagery using either automated or manual mapping. ● In Kulob, there is minimal erosion or damage evident along the main Yakhsu river from the May event. Rather the flooding entered the city along three smaller rivers originating from the east. ● The small capacity of the streams draining into Kulob suggests they would have responded quickly and were overwhelmed by the heavy rain of May 7 - 13 which was concentrated just to the west of Kulob. ● An increase in bare ground and/or urban landscape over the past 3 decades has been detected in many of the affected districts, which is likely to have contributed to enhanced runoff and erosion. ● Rapid urban expansion was clearly a key factor driving the magnitude of the disaster in Kulob. Primary information on damages caused during the May 7 - 13 disaster has been compiled from the series of REACT reports, and complemented with information from the WASH Sector Rapid Needs Assessment. Both sources are consistent in focussing on the Khatlon province, where affected households and the number of affected people were particularly concentrated in the city of Kulob (Table 1). Flooding in northern districts of Rasht, and around Dushanbe were widely reported in initial media articles and in the first of the REACT reports, but damages in these areas seem limited to the inundation of streets, and some damage to the periphery of several houses. To put the extent of the May 2021 event into a broader historical context, the impacts have been compared with flood, landslide, and other mass movement disasters affecting Tajikistan over the past 30 years (Figure 2). Officially, the EMDAT disaster database currently records only 200 people as having been affected by the May 2021 event, which would make it a comparatively small disaster. However, this number is at odds with the value of nearly 17,500 established from the REACT reports which would make it the largest disaster of the past decade. This discrepancy may be simply due to differences in the definition of “affected people”, or the way in which the records are compiled. In either case, the impacts associated with disaster are certainly not exceptional in the historical context. 6 Figure 2: Total affected population from flood and landslide disasters recorded for Tajikistan in the EMDAT inventory (https://www.emdat.be). The May 2021 disaster is currently recorded in EMDAT as affecting only 200 people, whereas the affected population given in the REACT situation report is up to 17,500 people. Using 10m Sentinel 2B imagery, efforts were made using automated techniques focusing on vegetation and soil changes to identify areas of erosion, inundation and deposition across all affected districts. However, due to the fact that flooding and mudflows primarily travelled within well-established and regularly active flood channels, the methods proved unsuccessful. Only in Kulob was some evidence of channel widening evident in the 10 metre imagery. Subsequent higher resolution post-disaster imagery from Google Earth was then used to confirm areas affected by the flood and mudflows, including the identification of inundated residential areas, destroyed bridges, and manmade river embankments (Figure 3). 7 Figure 3: Post-disaster (A) Sentinel 2 (16.05.2021) and B-D) Google Earth (14.05.2021) imagery showing inundation of residential or commercial areas, destruction of bridges, and damage to manmade river embankments occurring in the city of Kulob. The floods and mudflows primarily entered the city from three side-valleys flowing from east to west (pink arrows). Across the different initial reports, the terminology describing the involved processes is mixed, referring to floods, landslides and mudflows. From video footage observed, more typical flood conditions appear to have predominated in the northern districts (e.g. in Dushanbe) and there is little evidence from satellite imagery of any significant erosion or deposition in these catchments. In contrast, imagery from Kulob reveals areas of upstream erosion, and sediment deposition within affected areas of the city (Figure 3). Levee formation is not evident, suggesting relatively thin depositions, indicative of hyperconcentrated flows, or what may more generally be referred to in the region as mudflows. Regardless of the actual sediment volume of the flows, which could be ascertained only with field studies, high discharges and flood conditions are a necessary driver of the impacts that occurred. Table 1: Overview of impacts across the affected districts/cities, based on immediate reports available post-disaster. District locations are shown in Figure 4(A). 8 Affected Province Affected Affected Affected Other city/district households households people damage (REACT) (WASH) (REACT) (REACT) Dangara (3) Khatlon 134 100 600 Damaged bridge Darvoz (2) GBAO Temporary closure of road Farkhor (4) Khatlon 15 50 105 Hissor (14) Tadzhikistan 50 Loss of cattle Territories Jomi (5) Khatlon 167 187 1500 Khuroson (6) Khatlon 50 Roads Kulob (7) Khatlon 1500 504 10500 Road, bridges Muminobod (8) Khatlon 120 120 960 5 bridges, loss of cattle Qumsangir (9) Khatlon 120 50 900 Rasht (15) Tadzhikistan Auxiliary Territories premises of 10 houses Rudaki/Dushanbe (1) Dushanbe 10 Flooding of streets Shurobod (10) Khatlon 31 50 201 Roads, bridges, loss of cattle Tavildara/Sangvor Tadzhikistan Temporary (16) Territories closure of road Vakhsh (11) Khatlon 217 52 1520 Varzob (17) Tadzhikistan Loss of cattle and Territories vehicles Vose (12) Khatlon 160 27 880 Yovon (13) Khatlon 52 52 288 As with many disasters occuring in mountainous countries, there was some initial suggestion of glacial lake outburst floods (GLOFs) or other processes originating out of the glacial environment having played a role in the events of May 2021. However, it rapidly became clear that with the disaster primarily impacting Khatlon province, the likelihood of such processes being involved was minimal as the affected districts are far downstream and beyond the reasonable reach of flood or mass movement processes originating from the glacial or permafrost environments (Figure 4). For other mountainous districts, such as Rasht or Sangvor, Sentinel imagery was closely inspected, but no evidence of recent instabilities was detected, consistent with the lack of observed impacts in these areas (Table 1), and the fact that the heavy rainfall was not centred over these regions (Figure 4A). An analysis of key topographic and hydrological characteristics defining the basins of the affected districts was undertaken, drawing on basin-level data (level 07) provided by hydroSHEDs (https://www.hydrosheds.org/), and extracted from the Shuttle Radar Topography Mission (SRTM) 90m DEM (Table 2). Notably, the two districts in which damages were greatest (Kulob and Vakhsh), are both characterised by small river areas (channel width at bankfull flow multiplied by length), low average streamflows, and low stream powers. The typically small hydrological capacity of such streams would have led to a situation in which the streams were rapidly overwhelmed by the heavy rain of May 7 - 13, in comparison with other less-affected districts where the steam capacities are larger. 9 This hypothesis is strongly evidenced in Kulob by the fact that there is no indication of erosion seen along the main Yakhsu river, which has a massive floodplain and related capacity to accommodate large flood events. Rather, the flooding and mudflows have clearly originated out of the basins draining from the east into Kulob, and these basins are characterised by lower order, smaller capacity streams (Figure 3). Table 2: Overview of topographic and hydrological characteristics for the hydrological basins of the flood-affected districts. Values are an average of all hydrological basins that intersect with the district boundary. For comparison, average values for the entire Khatlon province are indicated. Hydrological information is taken from HydroATLAS and HydroRIVERS (https://www.hydrosheds.org/). Affected River area1 Mean Mean basin Stream Mean stream Stream city/district (hectares) elevation slope (°) gradient discharge power (m) (decimetre (m3) (kW/m2) per km) Dangara (3) 6815 998 9.3 222 549 64 Darvoz (2) 9360 3150 25.7 916 1381 915 Farkhor (4) 5535 601 7.2 109 779 203 Hissor (14) 3329 1489 15.3 328 206 24 Jomi (5) 5831 1003 6.1 260 1379 138 Khuroson (6) 5831 1003 7.6 260 1379 51 Kulob (7) 1082 1278 5.7 317 127 9 Muminobod 7197 2026 16.6 592 760 5 (8) Qumsangir (9) 5530 552 2.2 83 1531 16 Rasht (15) 14398 2795 25.6 730 738 102 Rudaki/Dushan 765 1467 3.6 284 250 4 be (1) Shurobod (10) 9750 1829 17.5 561 1418 796 Tavildara/Sang 17454 3205 25.5 856 813 114 vor (16) Vakhsh (11) 756 717 4.4 122 60 1 Varzob (17) 2640 2014 23.6 441 181 9 Vose (12) 2535 1178 6.2 265 200 24 Yovon (13) 5085 1178 9.3 332 1044 124 Khatlon 3390 994 9.2 238 1378 6 Province Average 1 In the HydroSHEDS database, the surface area of every river reach is calculated by multiplying channel width (at bankfull flow) by length. Long-term (1971-2000) monthly maximum discharge is used as a proxy to represent bankfull flow (see Lehner and Grill, 2013) . 10 Figure 4: Districts affected by the May 2021 flood and mudflow events (red outline, numbers refer to Tables 1-3). A) Accumulated precipitation from May 6-13 from the GPM IMERG multi-satellite estimate with climatological gauge calibration (GPM_3IMERGHHL v06). B) March average snow cover from MODIS, with glaciers (from RGI 6.0) in white. C) Permafrost zonation index, after Gruber et al. (2012). See Table 1 for district names. Further analyses focused on landuse/landcover (LULC) characteristics within the flood-affected districts. To achieve this, supervised classification of Landsat 5 (1993) and Landsat 8 (2020) scenes covering the study region were undertaken. In total, 6 scenes from each year were acquired, with dates ranging between late August - late September. Preprocessing steps included cloud, snow, and shadow masking. In total 100 samples of different LULC classes were extracted, and the spectral signatures used as a basis for the supervised classification using a maximum likelihood algorithm. 11 Results were rather mixed at the larger-scale, with few noteworthy differences between Kulob, other flood affected districts, and Khatlon province as a whole in terms of the coverage of cropland, pasture land or urban area (Table 3). However, the higher road density within Kulob stands out, and is testament to the relatively dense urban infrastructure of the city. Attempts to assess temporal changes in landcover (particularly changes in crop and forest cover) between 1993 and 2020 were mostly unsuccessful, due to the large size of the study region, and huge diversity in vegetation characteristics across regions and between years that cannot be easily distinguished using the spectral signatures extracted from Landsat imagery. However, there is higher confidence in the results obtained for bare ground, showing an increase in bare ground within many of the flood affected districts from 1993 - 2020, particularly within Kulob (35% increase in bare ground). Whether this is due to overgrazing, deforestation, or urbanisation (difficult to distinguish spectrally from bare ground) cannot be determined without more comprehensive inter-annual analyses. Nonetheless, an increase in bare ground, regardless of the driving process, will have facilitated more rapid runoff and erosion during times of heavy rainfall. Table 3: Overview of key landuse/landcover characteristics for the hydrological basins of the flood- affected districts. Basin-level data is extracted from HydroATLASand HydroRIVERS (https://www.hydrosheds.org/), while change in bare ground was assessed from landsat-based landuse/landcover classification. Affected Population Cropland Pasture land Urban area Road density % change in city/district density (per (%) (%) (%) (m per km2) bare ground km2) (1990 - 2020) Dangara (3) 125 17 29 3 198 -5 Darvoz (2) 17 10 38 0 138 16 Farkhor (4) 124 27 43 3 215 -14 Hissor (14) 292 17 36 9 125 -14 Jomi (5) 186 17 26 3 144 1 Khuroson (6) 186 17 26 3 144 1 Kulob (7) 171 19 36 3 273 35 Muminobod 97 12 45 2 216 10 (8) Qumsangir (9) 165 19 46 3 217 -8 Rasht (15) 40 10 21 1 109 47 Rudaki/Dushan 388 16 22 19 127 -35 be (1) Shurobod (10) 106 12 58 2 241 18 Tavildara/Sang 23 8 24 0 121 22 vor (16) Vakhsh (11) 170 17 32 4 196 -25 Varzob (17) 222 15 22 11 124 -1 Vose (12) 140 19 34 3 248 34 Yovon (13) 133 16 22 2 132 3 Khatlon 126 16 44 2 200 -4 Province average To better distinguish urban infrastructure, from areas of erosion, rock and bare ground, LULC mapping for 1993 and 2020 was repeated for a smaller area of interest, focussing on Kulob. With a smaller area of interest, there is less variation in the spectral signatures for a 12 given LULC class and results are typically more robust. At this scale, the truly significant and rapid expansion of urban infrastructure both upstream and within the flood-inundated areas of Kulob becomes evident (Figure 5). Replacement of croplands/grass with an expanding urban landscape has likely increased the susceptibility of the land to rapid runoff and erosion, particularly on hilly areas (e.g. Figure 5c). At the same time, expansion of residential and commercial property adjacent to the streams has significantly increased exposure levels, and has undoubtedly been a key factor in the high number of households and people affected by the May 2021 disaster in Kulob. Figure 5: (A): Automated landuse/landcover classification results based on Landsat imagery from 1993 and 2020, showing expansion of urban infrastructure within and upstream of the areas affected by the May 2021 flooding in Kulob. Much of this urban expansion occurred 13 over the past two decades, as shown in the enlarged views in google earth from 2004 (B,C) and post-disaster May 2021 (B’,C’). 14 3. The May 2021 event and its climatic context Key messages: ● The May 2021 event was primarily an extreme precipitation event (99th percentile) with no evidence of any significant influence of the cryosphere. ● The event period 6-13 May had accumulated precipitation totals of 50-100mm in the worst hit regions. ● The most significant event (May 11) centred on Kulob is largely missing from the Synop station network, confirming the fact that Tajikistan's monitoring network does not provide adequate coverage. ● Satellite data reveals higher than average snow cover and strong melt in days prior to the event, entailing heightened river discharges, confirmed by snow model runs. In this analysis we draw on three data sources with different uncertainty profiles; meteorological stations, combined satellite/gauge products and meteorological reanalysis. Precipitation is often highly variable in space and time and considerable uncertainties are associated with all data products. A joint analysis considering all available data is often useful to build confidence in results. Meteorological stations: Tajikistan's station network is relatively sparse given the topographic complexity and climate regimes that characterise Tajikistan. Additionally, station data often is incomplete due to operational reasons and in some cases only available in recent years. Constructing climate norms is often problematic and even on event timescales precipitation undercatch (especially in the case of solid precipitation) leads to measurement biases and additionally, the representivity of single point measurements of larger weather systems is often questionable. Combined satellite/ gauge products: We use the GPM IMERG (late run) combined product that is available at 0.1 degrees (~11km). This combines multiple satellite retrievals with the ground gauge network. Spatial patterns are well represented due to use of optical imagery however satellite products often underestimate solid precipitation and assimilation of ground based gauges is less effective where data is sparse. Data archive length is 20 years. Climate reanalysis: We use ERA5-land the high resolution versions of the latest reanalysis product from ECMWF available at 10km. Precipitation is purely analytical (model based) but overall the model is heavily constrained by a sophisticated assimilation framework that makes use of earth, air and space borne monitoring capabilities and provides a consistent climate record of the past 40 years. It is the best choice for constructing climate norms, yet reanalysis can be characterised by biases particularly in regions where observations are sparse meaning the NWP model is less well constrained. Reanalyses have been shown to be wet-biased over High Mountain Asia, however as discussed, mountain precipitation is difficult to quantify with high associated uncertainty. IMERG accumulated precipitation totals show up to 100 mm during the period 6-13 May 2021 (Fig. 6) with the main event focused on Khatlon (Afghanistan was also affected during this event). A second event is focused over the Fan mountains at the western end of the Zarafshan range on the border with Uzbekistan (39.3°N 68°E). However no damage was recorded in the Panjakent (Zerafshan) area. In Figure 7 we look at the time-series of averaged values over the Kulob catchment (Fig. 6, red polygon). This indicates 4 peaks during the period May 6-13 with the most intense period 13-18h local time on May 11 with 27mm registered. Single grid cells in the upper part of the catchment recorded up to 100mm over this period. District average accumulations are given in Table 4, however these need to be interpreted with caution as they often do not follow hydrological boundaries. For example the highest intensity 15 precipitation occurred in the upper Kulob catchment, yet this is registered to Shurobad district, explaining Kulobs 4th place ranking. Meteo stations from Tajikistan are shown in Figure 8 with stations in or near affected districts shown in red and further analysed. Figure 9 indicates that the main event was largely absent from station data suggesting the density of the station network is not sufficient to capture such events. Interestingly the largest peak of 35 mm was recorded at station Hovaling on the 4 May, before the main event period and in a district with no reports of damage. Figure 6: Accumulated precipitation 6-13 May 2021, from GPM IMERG. Mean accumulated precipitation given for each affected district in black. Red polygon indicates the worst affected Kulob catchment with mean accumulated precipitation (57mm). The catchment spans the two worst affected districts (Table 3) explaining why impacts were strongest in Kulob. 16 Figure 7: GPM IMERG half-hourly precipitation intensities averaged over the Kulob catchment. Average accumulated totals over the catchment were 57 mm with as much as 98 mm in upper parts of the catchment. The main precipitation phase occurred on 11 May from 13h-18h local time with accumulated totals of 27mm. Table 4: District average accumulated precipitation 6-13 May. These administrative boundaries do not necessarily correspond to hydrological boundaries. For example the catchment draining to Kulob town received the highest precipitation intensities, yet belongs to Shurobod and Muminobod districts. 17 Figure 8: Location of Tajik Met stations (black) and stations in or near affected districts (red). Affected district outlines (black). 18 Figure 9: Daily sums of precipitation from stations in or near affected districts. This data is a quality controlled dataset made available from NOAA archive “Global Surface Summary of the Day” based on WMO synop stations. In order to put this event in a historical or “climate” context we use the ERA5-Land archive to compute the distribution of precipitation events over 41 years (1979-2020). We then use the 95th and 99th percentile to define what constitutes a “severe” and “extreme” event respectively. Importantly this is a wet-day percentile i.e. computed only for the subset of non-zero precipitation days defined by a threshold of 1mm/day. This threshold is based on observational constraints (Schär et al. 2016). Figure 10 shows mean annual precipitation according to both GPM IMERG and ERA5- Land. While spatial patterns are similar, ERA5-Land has around a factor two higher magnitude. Although, as discussed, GPM IMERG is likely a low estimate due to uncertainties with solid precipitation detection. In reality there is a great variation with altitude in the humid regions with the ERA5-Land values quite probable at high elevation and much drier valley bottoms. However, both products are too coarse at c. 10 km to capture this slope-scale variability. Both datasets show that the main affected region of Khatlon is located in a transition region of Tajikistan between humid and arid regimes. 19 We computed the 95th/99th percentile precipitation event as a measure of what constitutes extreme precipitation in this region and how that varies in space. Figure 11 presents these results computed from ERA5-Land (the only climate length record). For the Kulob catchment this is computed as 22 and 31 mm/day for the 95/99th percentile, respectively. We then computed percentiles for the (shorter) IMERG dataset so we could compare the event scale totals given in Figure 6 & 7, directly. Precipitation percentiles computed as an area average in the Kulob region (2001-2020) give 95% as 14.9 mm/day and 99% as 24.8 mm/day, as expected, somewhat lower than percentiles computed according to ERA5-Land. These results confirm the event as extreme with precipitation on May 11 exceeding 27mm in the Kulob catchment. However, it was likely double this at higher elevations in the catchment. Figure 10: Mean annual precipitation from IMERG (Satellite/gauge) archive 2001-2020 and ERA5- Land (reanalysis) 1979-2020. Spatial patterns are similar but magnitudes differ significantly. Figure 11: Precipitation rates 95/99th percentile (mm/day) computed according to the ERA5-Land record (1979-2020). This indicates the magnitude of what constitutes a severe/ extreme precipitation event across the country. In the Kulob region this corresponds to approximately 22 mm/day. In order to assess the contribution of snowmelt to the event we use a combination of optical satellite data (MODIS) and modelling using a dedicated snow model. Figure 12 shows a snow 20 covered area (SCA) curve computed from the MODIS 8-day SCA product for both the current season (2020- 2021) and the climate norm computed as an average over the entire MODIS archive (2001- 2021). This indicates whether snow cover is above or below average in terms of coverage and also how quickly the snowpack is melting (and therefore generating runoff) through interpretation of the steepness of the curve. Figure 12 indicates that snowmelt may have had a role to play as snow cover was above average in the days prior to the May event and importantly, melting rapidly. The first order effect is higher soil moisture contributing to reduced soil stability and secondary enhanced runoff, further exacerbating the effect of the strong precipitation event. Figure 13 shows the snow cover depletion in the month leading up to the event and indicates that while snowmelt may have contributed to events in the affected mountainous districts and likely the Dushanbe floods, it was likely not a factor in the more significant Khatlon events, particularly Kulob where melt-out of the snowpack occurred at least 1 month earlier. Figure 12: Snow covered area curves computed from the entire MODIS 8-day maximum snow cover product (MODA1) 2001-2021 for the entire area of Tajikistan. Comparison with the current season indicates above average snow cover and strong snow melt (steep curve) in late April 2021 preceded the May disaster and may have contributed through higher soil moisture and river discharge. 21 Figure 13: Evolution of snow cover prior to disaster as observed from the MODIS satellite. Affected districts given in red. We additionally employed a snow model FSM (Essery 2015) in combination with the downscaling scheme TopoSCALE (Fiddes & Gruber 2014) to compute snow water equivalent and runoff due to snowmelt in the upstream regions of Khatlon province. TopoSCALE downscales 3D meteorological fields from ERA5 reanalysis using high resolution DEMs that enable dynamic scaling with elevation and correction of radiative fluxes according to surface geometry (Fiddes et al. 2014). The modelling domain is divided into computational units using a multivariate clustering scheme that enables efficient application of numerical models (Fiddes et al. 2012). The snow model FSM (Essery 2015) is then forced by the downscaled meteo timeseries for a given computational unit and surface variables (SWE, runoff etc) are computed, with results mapped to the 2D simulation domain. Again these results confirm that snow cover was unlikely an important factor in terms of soil moisture (Fig. 14) however we do simulate higher than average snow melt peaks in the main channel of the Yahksu River (Fig. 15). However, as can be seen in Figure 16, this is a wide heavily braided channel with capacity for large flood peaks. There was no evidence of damage in this main channel (see Section 2), and indeed very little infrastructure is exposed in the channel flood plain. 22 Figure 14: Modelled SWE upstream of Kulob on 2 May 2021. Runoff shown in Figure 15 is computed from the catchment indicated in red. Figure 15: Modelled mean snowmelt runoff (mm) from upper catchment of Khatlon in days leading up to the disaster. Two above average snowmelt peaks coincide with the disaster suggesting higher than average discharges likely in the main channel of the Yakhsu River. However little damage was observed in the main channel probably due to it being a broad braided river bed with a large capacity. Note these peaks correspond to the rapid decrease in SCA observed in Fig 12, signalling rapid snowmelt. 23 Figure 16: The braided Yakhsu River channel that drains the upper Khatlon catchment. It has a large flood capacity due to broad braided nature and few exposed infrastructure. 24 4. Future perspectives under climate change Key messages: ● Precipitation is expected to increase, most significantly under high emission scenarios by the late 21st century (2071-2099) in the humid parts of Tajikistan (Central regions, Western Pamir, Fan mountains). ● The number of heavy precipitation days (R20) is projected to increase by 2-4 times, with upper estimates being reached under RCP8.5 in the late 21st century. ● The likelihood of similar disasters (or at least the climate drivers) is therefore projected to also increase in frequency under all future RCP climate scenarios. We analysed the latest regional climate model results from WCRP CORDEX region Central Asia (CAS) at 0.22 degrees (https://cordex.org/). Four model chains were available (MOHC-HadGEM2- ES/GERICS-REMO2015_v1, CNRM-CERFACS-CNRM-CM5/RMIB-UGent-ALARO, MPI-M-MPI-ESM- LR/GERICS-REMO2015, NCC-NorESM1-M/GERICS-REMO2015) from the ESGF data nodes (http://esgf.llnl.gov/) and analysed together as an ensemble mean. We considered RCP emission scenarios 2.6 and 8.5 at both “near future” (2031-2060) and “far future” (2071-2099) periods with respect to the reference period (1981-2010). Evaluation of the CORDEX-CAS-22 domain runs ALARO and REMO2015 is given by Top et al. (2021) which provides important insights relevant to impact modelling, specifically quantification of biases with respect to observational datasets. However, it should also be kept in mind that observational datasets are highly uncertain in the CAS study region due to low density of ground measurements and rugged topography and therefore evaluation of variables such as precipitation remains problematic. With this in mind we present the trends without a further bias adjustment and suggest these are interpreted as trends only and not absolute values . CAS-22 reveals increasing precipitation throughout Tajikistan's humid regions, with more stationary or even drying conditions in the arid far-east, which given low absolute amounts is likely insignificant (Fig. 17). We used the index “R20” which refers to the number of heavy precipitation days per year (defined as > 20mm / day). This compares well to the presented 95th percentile precipitation event in most humid parts of Tajikistan (20-30mm/day Fig. 11) as well as the May 11 peak phase total in Kulob region (27 mm). We see an increase in this metric in all scenarios and time periods by a factor of 2-4, with higher values under RCP8.5 in the far future period (Fig. 18). 25 Figure 17: Mean annual precipitation anomaly with respect to the reference period. Wettening is seen in all humid regions (central regions, western Pamir, Fan mountains) with stationary or even drying conditions in arid regions (Far-east GBAO, south-west Khatlon, Sugd). 26 Figure 18: CORDEX CAS-22 multimodal means of number of heavy precipitation days defined by the index R20 (20mm/day). Results are presented for the reference period, near (2031-2061) and far future (2071-2099) periods under RCP 2.6 and 8.5. In parallel to the expected increase of extreme precipitation with time, also river runoff will change in the future under climate change. Excess glacier melt can contribute to stream flow in the ablation period, but at some point in time, the so-called peak water will be reached, after which glacial melt contribution will decrease due to reduced glacier mass. The latest glacier modeling study from Rounce et al. (2020) shows heterogeneous results for the Tajik region, but under most RCPs peak water is expected to take place in the first half of the 21st century. This is in line with findings from Huss and Hock (2018). Due to limited glacier coverage, snow distribution is more relevant for future river runoff in the districts analyzed here. A recent hydrological modeling study from Didovets et al. (2021) found that the peak of river discharge will shift by up to one month towards earlier in the year, mainly due to the shortening of the snow accumulation period. No clear trend was found in overall river discharge, but reduced runoff during summer months is expected, confirming the results of earlier studies. This shift in peak river runoff bears the potential to increase the probability for heavy rainfall events to coincide with high base flow in the rivers and thus increasing the probability for disasters similar 27 to the May 21 event. Such probabilities, however, are hard to quantify given the considerable uncertainties in the underlying GCMs and related process models. Although not relevant for the May 21 disaster, mass movements in general have the potential to cause damages to people and infrastructure in Tajikistan. Recent GLOF events in Tajikstan1 and neighbouring Kyrgyzstan2 in late July and early August were striking reminders of this. With raising temperatures, the formation of numerous glacial lakes is expected in the course of continued glacier retreat. Figure 19 shows simulated trajectories of potential outbursts from future glacial lakes, expected to form during the 21st century. The data is taken from Zheng et al. (2021), who extracted sites of future glacier lakes based on modeled glacier thicknesses from Farinotti et al. (2019) and modeled potential outburst trajectories. Of the districts affected by the May 21 disaster, only Tavildara and, to a lesser extent, Darvoz, Rasht and Varzob districts currently contain glacial lakes. In these districts new critical lakes will form during the current century, leading to new lake outburst hazards. Notably there is little variation between different RCPs regarding future glacial lake formation. In other, mountainous parts of the country, however, a significant increase in glacial lake outburst flood hazard has to be expected in the future. Figure 19: Trajectories of potential outburst floods of current glacial lakes (green trajectories) and of lakes expected to form during the 21st century according to different RCPs (blue, orange and purple). Data from Zheng et al. (2021). 1 https://asiaplustj.info/en/news/tajikistan/incidents/20210729/mudflow-triggered-by-high-temperature-and- glacier-melting-blocks-surkhob-river-in-eastern-tajikistan 2 https://kaktus.media/doc/443541_v_chyyskoy_oblasti_prorvalo_vysokogornoe_ozero_akpay_v_belogorke.ht ml 28 5. Conclusions and implications Based on the analyses presented above, the following conclusions can be drawn on the May 2021 disaster in Tajikistan: ● The precipitation event leading to the disaster was extraordinary (above the 95/99 percentiles of the 1979-2020 precipitation rates computed from the ERA5-Land record). ● Snow cover and snow melt contribution to runoff was likely above the climatic average as well, but not extreme. ● There is no evidence for mass movements initiating from within glacial or periglacial environments during this event, and as such, the cryosphere did not play a role beyond snowmelt discharge in channels and higher soil moisture during the snowmelt season. ● Most events originated from reactivated, existing mudflow channels (-> no new erosion channels identified in Sentinel imagery). Damages appear to have been enhanced by increased exposure over the past decades (as shown by significant expansion of urban infrastructure evident in Kulob). ● An increase in bare ground over the past 3 decades has been detected in many of the affected districts (potentially driven by overgrazing and/or deforestation), which is likely to have contributed to enhanced runoff and erosion during the heavy rainfall event in May 2021. ● Climate models (CORDEX CAS-22 multimodal means) indicate an increase of intense precipitation days (> 20mm/day) with time and RCP, and therefore an increase of the likelihood for events similar to May 2021 in the future can be anticipated. ● According to modeling studies, peak river runoff is expected to shift by up to 1 month earlier in the year. This could increase the likelihood that extreme rainfall events coincide with high base flow in the rivers. At the same time, droughts during the summer months can increase. In high-mountain areas of Tajikistan, increasing risks from glacial lake outburst floods are expected for the future. Most of the damages of the May 2021 disaster occurred where there has been increased exposure of assets within zones which show clear evidence of having been previously affected by mudflows and floods. The anticipated increase of both frequency of heavy precipitation events as well as further expansion of settlements and infrastructure therefore calls for efforts on future-oriented, scenario-based hazard mapping of related processes. Spatial planning based on hazard maps could prevent an increase of people and infrastructure that will be exposed to such processes. Hazard zonation as a risk reduction measure could be complemented by Early Warning Systems (EWS). EWS are currently highly promoted on the international policy agenda (e.g. in the Sendai Framework for Disaster Risk Reduction) and have proven to efficiently reduce exposure and vulnerabilities towards hydro-meteorological risks. Basic requirements for such systems are the availability of consistent and dense observation data, combined with efficient and reliable forecast models. Efforts towards improving hydrometeorological station networks, in particular in the mountainous regions of Tajikistan, are needed to improve the feasibility of such multi-hazard EWS. 29 References Alexander, L. V., and Coauthors, 2019: On the use of indices to study extreme precipitation on sub- daily and daily timescales. Environ. Res. Lett., 14, 125008. Didovets, I., and Coauthors, 2021: Central Asian rivers under climate change: Impacts assessment in eight representative catchments. J Hydrology Regional Stud, 34, 100779, https://doi.org/10.1016/j.ejrh.2021.100779. Essery, R., 2015: A factorial snowpack model (FSM 1.0). Geosci. Model Dev., 8, 3867–3876. Farinotti, D., M. Huss, J. J. Fürst, J. Landmann, H. Machguth, F. Maussion, and A. Pandit, 2019: A consensus estimate for the ice thickness distribution of all glaciers on Earth. Nat Geosci, 12, 168– 173, https://doi.org/10.1038/s41561-019-0300-3. Fiddes, J., and S. Gruber, 2012: TopoSUB: a tool for efficient large area numerical modelling in complex topography at sub-grid scales. Geosci. Model Dev., 5, 1245–1257. Fiddes, J., and S. Gruber, 2014: TopoSCALE v.1.0: downscaling gridded climate data in complex terrain. Geosci. Model Dev., 7, 387–405. Gruber, S., 2012: Derivation and analysis of a high-resolution estimate of global permafrost zonation. Cryosph., 6, 221–233. Huss, M., and R. Hock, 2018: Global-scale hydrological response to future glacier mass loss. Nat Clim Change, 8, 135–140, https://doi.org/10.1038/s41558-017-0049-x. Lehner, B. and G. Grill, 2013: Global river hydrography and network routing: baseline data and new approaches to study the world’s large river systems. Hydrological Processes, 27, 15, 2171-2186. Pfeffer, W. T., and Coauthors, 2014: The Randolph Glacier Inventory: a globally complete inventory of glaciers. J Glaciol, 60, 537–552, https://doi.org/10.3189/2014jog13j176. Rounce, D. R., R. Hock, and D. E. Shean, 2020: Glacier Mass Change in High Mountain Asia Through 2100 Using the Open-Source Python Glacier Evolution Model (PyGEM). Frontiers Earth Sci, 7, 331, https://doi.org/10.3389/feart.2019.00331. Schär, C., and Coauthors, 2016: Percentile indices for assessing changes in heavy precipitation events. Clim. Change, 137, 201–216. Top, S., and Coauthors, 2021: Evaluation of regional climate models ALARO-0 and REMO2015 at 0.22° resolution over the CORDEX Central Asia domain. Geosci. Model Dev., 14, 1267–1293. Zheng, G., and Coauthors, 2021: Increasing risk of glacial lake outburst floods from future Third Pole deglaciation. Nat Clim Change, 11, 411–417, https://doi.org/10.1038/s41558-021-01028-3. 30