LamaH | Large-Sample Data for Hydrology and Environmental Sciences for Central Europe

Very large and comprehensive datasets are increasingly used in the field of hydrology. Large-sample studies provide insights into the hydrological cycle that might not be available with small-scale studies. LamaH (Large-Sample Data for Hydrology) is a new dataset for large-sample studies and comparative hydrology in Central Europe. It covers the entire upper Danube to the state border Austria / Slovakia, as well as all other Austrian catchments including their foreign upstream areas. LamaH covers an area of 170 000 km2 in 9 different countries, ranging from lowland regions characterized by a continental 10 climate to high alpine zones dominated by snow and ice. Consequently, a wide diversity of properties is present in the individual catchments. We represent this variability in 859 observed catchments with over 60 catchment attributes, covering topography, climatology, hydrology, land cover, vegetation, soil and geological properties. LamaH further contains a collection of runoff time series as well as meteorological time series. These time series are provided with daily and also hourly resolution. All meteorological and the majority of runoff time series cover a span of over 35 years, which enables long-term analyses, also 15 with a high temporal resolution. The runoff time series are classified by over 20 different attributes including information about human impacts and indicators for data quality and completeness. The structure of LamaH is based on the well-known CAMELS datasets. In contrast, however, LamaH does not only consider headwater basins. Intermediate catchments are also covered, allowing, for the first time within a hydrological large sample dataset, to consider the hydrological network and river topology in applications. We discuss not only the data basis and the methodology of data preparation, but also focus on possible 20 limitations and uncertainties. Potential applications of LamaH are also outlined, since it is intended to serve as a uniform basis for further research. LamaH is available at https://doi.org/10.5281/zenodo.4525244 (Klingler et al., 2021).


Introduction
Hydrology and hydrological processes are characterized by high spatiotemporal variability. Runoff generation in small-scale, alpine catchments with steep and complex topography is dominated by different processes compared to lowland rivers with 25 flat topography. The water balance in an energy-limited, humid catchment in Europe is completely different than, for example, in a water-limited catchment in dry (semi-) arid regions in Africa or Australia. A water droplet flowing via the Russian Lena into the Arctic Sea has a completely different biography than a water droplet from Rwanda in Central Africa, which reaches the Mediterranean Sea via the Nile after more than 6 600 km. Boundary conditions and major drivers for the differences are the catchment properties, which can be described by characteristics regarding topography, hydro-climate, land cover, geology 30 and soil conditions.
In order to deepen our understanding of the hydrological process and further increase the reliability of (hydrological) models, it is necessary to account for this spatiotemporal variability in our approaches. A number of international initiatives (e.g. Distributed Model Intercomparison Project (DMIP; Smith et al., 2004); Inter-Sectoral Impact Model Intercomparison Project (ISI-MIP; Warszawski et al., 2014); Model Parameter Estimation Project (MOPEX; Duan et al., 2006) or Hydrological 35 Ensemble Prediction Experiment (HEPEX; Schaake et al., 2007)) have been launched in recent decades with the aim to advance the prediction of hydrologic variables through comprehensive model benchmarking in different regions of the world.
New efforts strive for creating homogeneous and consistent datasets, which serve as a solid basis towards the development of new modelling approaches.
In this context, a trend towards more complete and extensive datasets is apparent: 1) Remote sensing has enabled consistent 40 and global mapping of Earth's atmosphere and surface. The Sentinel (ESA, 2021;Malenovsky et al., 2012), Landsat (NASAa. 2021;Irons et al., 2012) and MODIS (NASAb, 2021; Barnes et al., 2003) missions are probably known to a broader public. 2) New software platforms or applications for obtaining and processing these mostly very data-intense (e.g. regarding data volumes) remote sensing products also facilitate the applicability. Examples of such platforms are "Google Earth Engine" (GEEa, 2021;GEEb, 2021;Gorelik et al., 2017;Klingler et al., 2020), the "Copernicus Open Access Hub" (COPa, 2021) or 45 the "Copernicus Climate Data Store" (COPb, 2021). 3) There is growing awareness that our understanding of the complex hydrological processes can be deepened through "large-sample" studies (Gupta et al., 2014). Large-sample hydrology (LSH) includes information from a broad range of different watersheds in order to derive robust conclusions (Addor et al., 2019).
Several research groups in different areas of hydrology have already focused on LSH for this reason (e.g. Berghuijs et al., 2014;Blöschl et al., 2019a;Döll et al., 2016;Gudmundsson et al., 2019;Luke et al., 2017;Kuentz et al., 2017;Singh et al., 50 2014;Van Lanen et al., 2013). 4) Finally, data-driven models and deep learning approaches have recently gained significant attention in hydrology (Sit et al., 2020). Independent from the fact that these developments are controversially discussed (Nearing et al., 2020), their excellent performance in time series prediction, also in an ungauged setting (e.g. Kratzert et al., 2019a), is related to the ability of machine learning to identify patterns and relationships in data (Kratzert et al., 2019b). These 18 different river basins (Table 1). An overview of the domain covered in LamaH with the river regions and the 882 runoff gauges with their elevations is illustrated in Fig. 1. The difference to the 859 catchments defined in basin delineation A (see chapter 3) can be explained by the fact that 23 gauges, which mostly do not have a clearly definable catchment area (e.g. gauges at artificial channels or below large karst springs, see also chapter 5.8), were not considered in basin delineation. All river regions in the project except regions 1 and 11 are part of the Danube´s catchment. River region 1 covers the upper 100 catchment of the Rhine from its sources to Lake Constance and region 11 covers the Austrian catchment area of the Vltava, which is the largest tributary of the Elbe.

Basin and sub-basin delineations 110
Meteorological time series and catchment properties are usually based on global datasets, which are provided either in raster or vector form. In LamaH, a catchment property or time step of a meteorological time series usually represents the mean of the orographic catchment of a gauge. Consequently, aggregation of spatially distributed information of the corresponding source datasets was required. As aggregation areas are catchment boundaries (Fig. 2) from the Digital Hydrological Atlas of Austria (HAO, 2007) and from HydroATLAS (Linke et al., 2019) used. The sub-basins of both data sources were combined 115 and, if necessary, adjusted to reflect the gauges catchment area at their downstream end. In a next step, all (smaller) subcatchments, which belong to the respective upstream catchment area of an individual gauge, were combined. Each gauge is therefore assigned an aggregation area representing the overall orographic catchment area. We refer to this method of catchment / basin delineation in the further text and the dataset as "basin delineation A". Plausibility of this type of basin delineation was checked by calculating the ratio between the aggregation area and the officially, e.g. in the metadata of the 120 gauges, declared catchment area (attribute "area_ratio" in Table A1). The median basin size over all 859 basins of LamaH applying basin delineation A is 178 km², with a range of 3.9 km² to 131 000 km². Basin delineation A is known from the CAMELS datasets. The advantage of basin delineation A is the independency between the basins, since the aggregation area fully represents the orographic catchment area of a gauge. However, for gauges with larger catchments, aggregation with basin delineation A leads to a significant loss of information, as variability as well as small-scale characteristics are lost. Another 125 drawback is the multiple, but differently weighted, mapping of local features (see overlap in Fig. 2a). Therefore, basin delineation A is supplemented by a form of delineation (catchment delineation B, also 859 catchments) where the orographic catchment area of the next upstream gauge (may be none, one, or more) is subtracted from that of the current gauge (Fig. 2b).
This results in the representation of intermediate catchments, which become part of a large connected river network. The dependency among these intermediate catchments requires a catchment or gauge hierarchy (Fig. 2b,attribute "HIERARCHY" 130 in Table A1), as well as information regarding the upstream-downstream relationship ("NEXTUPID" or "NEXTDOWNID" in Table A1). The median basin size applying basin delineation B is 114 km², with a range of 1.3 km² to 2 496 km². The third basin delineation provided in LamaH (further referred as basin delineation C in the text and dataset) is similar to basin delineation B, but only includes catchments with no or only low anthropogenic influence (454 catchments) (Fig. 2c). The intention is to provide modelers in subsequent studies with an easy to use data basis of hydrological patterns that are near to 135 natural conditions. Anthropogenic influences in the catchments and runoff data is described more detailed in section 5.8. LamaH contains daily and hourly runoff time series for 882 different gauges, located in 4 different countries (Austria, Germany, Switzerland and Czech Republic). The main provider of runoff time series was the Hydrographic Central Bureau of Austria (HZB, 2020), which contributed data for 609 gauges located in Austria. The hydrographical services of the German federal states Bavaria (GKD, 2020) and Baden-Württemberg (LUBW, 2020) provided 125, respectively 61 runoff time series. 25 runoff time series came from the hydrological office of Switzerland (BAFU, 2020), while time series for 61 gauges were 160 provided by the Czech Hydrometeorological Institute (CHMI, 2020). The format of all obtained time series was unified, enabling much easier data processing. The various gauge attributes and metadata are listed and described in Table A1. The unit of discharge is m 3 s -1 for both daily and hourly resolution. Conversion to runoff heights can be performed using the catchment area provided (attribute "area_gov" in Table A1).
Runoff time series are in most cases derived by water level -discharge relationships (rating curves). Changes in channel profile, 165 e.g. after floods with strong bedload transport, can lead to an incorrect runoff calculation. However, attempts are usually made to minimize this source of error by periodically adjusting the rating curve. The adjustment frequency of these rating curves is not publicly available, but only gauges in the highest quality class (quality classes are declared in Bavaria and the Czech Republic) were included in LamaH. Runoff time series with daily resolution are often provided with longer observation periods than those with hourly resolution. Therefore, daily and hourly runoff time series can be obtained separately from the listed 170 hydrological offices. However, we normally requested only the time series with hourly resolution and derived the daily time series from them. This approach was chosen for the runoff data from Austria, Germany and Switzerland, since those time series with hourly resolution mostly include quite long recording periods. Fig. 3a  Although the exact scope of data verification by the staff of the various hydrological services is not further specified, we have added an attribute describing the check status (attribute "checked" in Table A1) to each time step of the runoff time series. The 180 Austrian, Czech and Swiss runoff data are provided exclusively checked, while the runoff data from the Bavarian hydrographic service is in most cases quality controlled until years 2014 -2016. Data from the German federal state Baden-Württemberg is often checked only from the year 2010 onwards. Some time series contain gaps, also after checking by the hydrological services. In order to reduce the number of gaps, up to 6 consecutive gaps (6 hours) were filled with linear interpolation during our processing. Any remaining gaps in the time series were marked with the number -999. The fraction of remaining gaps in 185 the continuous runoff time series is declared by the attribute "gaps" (Table A1)  gauges with very few gaps (< 0.1‰) are mostly located in Austria, Czech Republic and Switzerland. About 80% of the 882 gauges have no gaps in their continuous time series after our processing. The time steps with gaps before our processing are listed in separate files, attached to the dataset. The spatial distribution of the gauge hierarchies from basin delineation B (see Fig. 2b) is mapped in Fig. 3c., where 50% of all gauges have a hierarchy of 1 and thus represent headwater catchments. The 190 highest hierarchy (26) is found for the very last downstream gauge of the Austrian Danube (ID 399).

Meteorological data 195
Given the extent of the ECMWF (European Centre for Medium-Range Weather Forecasts) ERA5-Land dataset with global coverage (Muñoz Sabater et al., 2021), it was possible to obtain gap-free time series with daily and hourly resolution for 15 meteorological variables and 39 years (Table A2). ERA5-Land is a derivative of the ERA5 climate reanalysis (Hersbach et al., 2020), but only covers the terrestrial components. Further developments compared to ERA5 include an interpolation package for a finer temporal resolution, an additional sea level adjustment of the meteorological fields, as well as more efficient 200 possibilities for the import of updates (Muñoz Sabater et al., 2021;Yang and Giusti, 2020). ERA5-Land has a spatial resolution of 0.1 arc degrees (about 9 x 11 km at the latitudes of the project area) compared to the grid size of ERA5 of 0.25 arc degrees.
The temporal resolution of ERA5-Land is 1 hour, while ERA5 only has a 3-hour resolution. There is no data assimilation (fitting to observations) applied for ERA5-Land, but observations are indirectly implemented via the assimilated atmospheric fields of ERA5 (Hennermann and Guillory, 2020;Yang and Giusti, 2020). In accordance to ECMWF regulations, an 205 uncertainty estimate for ERA5-Land will be released (Muñoz Sabater, 2019b;Muñoz Sabater, 2017), but was not available at the time of writing (January 2021).
The meteorological time series were determined for all 3 forms of basin delineation (A/B/C in chapter 3). In order to do so, we intersected the aggregation areas of the various basin delineations with the gridded data of ERA5-Land in order to get area fractions for an area-weighted aggregation (arithmetic averaging). As already mentioned in the introduction, we would also 210 like to point out possible uncertainties of the published data. We therefore determined the components of the water balance for the period 01. 10.1989 to 30.09.2009 and plotted them (Fig. 4a). Values of catchments influenced by cross-basin water transfers, water withdrawals or intakes, large karstic springs or high infiltration (see chapter 5.8) are not shown in Fig more objective interpretation. In case of long-term water balances, it is usually feasible to neglect artificial storage in the catchment. The difference between long-term mean precipitation (P) and runoff height (Q) should be equal to the total 215 evapotranspiration (ETA) in a fulfilled water balance. This would be shown by having all points in Fig. 4a on the 1:1 line, which is not the case. Reasons for the rather strong scatter (Pearson correlation R = 0.30) may be an insufficient representation of precipitation or total evapotranspiration by ERA5-Land, an inaccurate recording of runoff (e.g. strong, unrecorded groundwater flow or change in river profile at the gauging station and thus inadequate water level -discharge relationship), a significant deviation between orographic and hydrographic catchment area (subsurface inflows and outflows, especially in 220 karstic areas), or lastly, in case of existing glaciers, a negative mass balance (Lambrecht and Kuhn, 2007;Kuhn, 2004;Kobolschnig and Schöner, 2011;Oerlemans et al., 1998;WGMS, 2005). Using other precipitation datasets for the same evaluation does not result in a significantly more compliant long-term water balance. CHIRPS Daily v2 (Funk et al., 2015) resulted in a correlation R between (P -Q) and ETA of 0.34 or MSWEP v2.2 (Beck et al., 2017; in even a lower R-value of 0.26. Even if we cannot resolve the issue at hand, the total evapotranspiration from ERA5-Land and its dependence on 225 elevation seems quite plausible compared to other studies ( Fig. 4a; HAO, 2007, Map 3.3;Herrnegger et al., 2012, Fig. 20).
Negative differences of mean precipitation and runoff height (Fig. 4a) and thus runoff coefficients > 1.0 (Fig. 4c, 32 of 594 catchments) are mainly present in higher terrain (negative mass balance of glaciers, Fig. 8d) and in catchments with a high fraction of carbonate sedimentary rocks (indicator of karst, Fig. 11c). Since ERA5-Land indirectly incorporates in-situ observational data via the assimilated atmospheric fields of ERA5 (Yang and Giusti, 2020), systematic measuring error of a 230 terrestrial station being used could also explain insufficient mean precipitation . The individual components of the water balance are attached to the dataset for every catchment, since this evaluation might be useful for explaining any deviations in a later modelling. The Budyko curve (Fig. 4b;Budyko 1974) describes the relationship of the ratio current evapotranspiration / precipitation (ETA/P) to the ratio potential evapotranspiration / precipitation (PET/P) and indicates whether evapotranspiration of a 235 catchment is limited by energy or water. Ideally all points should lie in the proximity of the Budyko curve. The deviation from this ideal case can primarily be explained by a significantly too high PET of ERA5-Land over nearly the entire range of elevation. For example, 98% of all 859 watersheds show mean annual PET sums above 1000 mm. As these PET sums are not realistic at the latitudes of the project area (HAO, 2007, Map 3.2;Herrnegger et al., 2012, Fig. 17), we did not include the potential evapotranspiration of ERA5-Land in the LamaH dataset. The runoff coefficient (Q/P) as a function of the ratio mean 240 precipitation / total evapotranspiration (P/ETA) is shown in Fig. 4c. The altitudinal dependency can be clearly seen in Fig. 4c, while catchments with lower mean elevation show less scatter. Fig. 4d shows the contrast of the long-term precipitation of ERA5-Land with those of ERA5 (Pearson correlation R = 0.936). ERA5 indicates systematic surplus of precipitation at catchments with mean altitudes above 2000 m a.s.l. compared to ERA5-Land, while at catchments with mean altitudes between 800 and 1200 m a.s.l. the opposite is more likely to be the case. The correlation between the long-term precipitation of  Land and those of the dataset CHIRPS daily v2 (Funk et al., 2015) is 0.916 (Fig. 4e). Further, the mean precipitation sums of CHIRPS daily v2 tend to be lower than those of ERA5-Land over the whole range of altitudes. More scatter (R = 0.841)

Catchment attributes
The various physio-geographical characteristics of a catchment, as well as their interactions, are essential for water storage and transport on and below earth's surface (Blöschl et al., 2013). The spectrum of influencing catchment characteristics 265 includes topography, climate, hydrology, land cover, vegetation, soil, geology, as well as the type and degree of (anthropogenic) impact on runoff processes. Furthermore, catchment attributes are crucial to determine interrelationships among different watersheds along several gradients (Addor et al., 2017a;Falkenmark and Chapman, 1989;Fan et al., 2019).
In most cases, we used freely available datasets with global or at least European coverage for deriving the different catchment attributes. Aggregation of the spatially distributed information of the basic datasets is performed for each of the 3 different basin delineations (A/B/C, according to chapter 3) and by default by calculating the area-weighted arithmetic mean (otherwise indexed). Furthermore, the upscaling is performed by two different approaches: 1) In the first approach (referred as "upscaling approach: 1") the aggregation is based on all the raster cells, which centroids are located inside the catchment, whether the catchment completely covers them or not. Especially at small catchments it is possible that no or only one raster centroid intersects the basin delineation. In this case the aggregation is calculated from all contributing raster cells using an area-275 weighted mean. Upscaling approach 1 is mainly used for relatively fine-gridded data sources (< 1 km), since it is not that computing-intensive and potential inaccuracies are negligible. 2) The second approach of upscaling (referred as "upscaling approach: 2") exclusively performs the aggregation area-weighted. Upscaling approach 2 is used for coarser gridded and vectoral data sources. The applied approach is indicated in the corresponding tables in the appendix.
The data basis for LamaH, methods of processing, possible uncertainties as well as the spatial distribution of catchment 280 properties (Addor et al., 2017b) are discussed in more detail in the following subsections. It is clear that due to the large dataset this account is far from complete. The individual attributes are listed in tabular form in the appendix with a more detailed description, units and with reference to the data sources.

Topographic indices
We calculated 10 topographic attributes, which are listened in Table A3. The attribute "area_calc" describes the calculated 285 aggregation (catchment) area, depending on the applied method of basin delineation (see chapter 3). Basin delineation A shows that about 34% of all 859 basins (aggregation areas) are smaller than 100 km 2 , 50% are between 100 and 1 000 km 2 , 14% are between 1 000 and 10 000 km 2 , and about 2.8% are larger than 10 000 km 2 . Large catchment areas are especially present for the gauges at the Danube and its larger tributaries (Fig. 5a). One reason for using multiple basin delineations is the reduction of aggregation areas and thus providing a more representative representation of local conditions and maintain natural 290 variability. When applying basin delineation B, about 45% of all 859 aggregation areas have an area of less than 100 km 2 , 52% between 100 and 1 000 km 2 and only 2.3% have an area above 1 000 km 2 .
A key factor for hydrologic processes is elevation, as it affects numerous other catchment characteristics including climate, land cover, vegetation, or soil development (Addor et al., 2017a). We derived the mean catchment elevation (Fig. 5b, "elev_mean" in Table A3), the median elevation ("elev_med"), standard deviation within a catchment ("elev_std"), the 295 elevation range (maximum-minimum elevation in the catchment, Fig. 5c, "elev_ran"), as well as the mean catchment slope ( Fig. 5d, "slope_mean") from NASA's SRTM dataset (Farr et al., 2007). SRTM features a grid size of 30 m and provides a maximum global absolute error of 16 m at a 90% confidence interval, while accuracy decreases with increasing elevation and slope (Farr et al., 2007). The slope was calculated with the algorithm of Horn (1981) using the terrain elevation from SRTM.
High mean catchment elevations and slopes are most apparent in the Eastern Alps, which extend from the southwest to the 300 central east of the project area. This high elevated area is mainly surrounded by the flatter Alpine foothills and regions with older geological zones (Fig. 5b). The shape of the catchment area and the stream network also influence runoff formation. The direction of precipitation in relation to the longitudinal axis of the catchment is of major interest in case of flood situations, especially in larger catchments.
For this reason, we also specified the angle between the north direction and the longitudinal axis ("mvert_ang") in addition to 305 the distance of the longitudinal axis of a catchment ("mvert_dist"). In combination with the two wind components of ERA5-Land ("10m_wind_u", "10m_wind_v" in Table A2) it is possible to derive the relative rainfall trajectory. The shape of the catchment is also relevant for the rise of the flood wave. The attribute of length elongation according to Schumm (1956) (Fig.   5e, "elon_ratio") is an indicator regarding the "roundness" (the higher, the rounder) of the catchment. But irregularities like large notches in catchment's shape may reduce the significance of this attribute. Stream density (Fig. 5f, "strm_dens") is a 310 function of several characteristics (e.g. climate, relief, soil properties, geology, vegetation, land use, glaciation or karstification) and can therefore be an informative indicator for comparing watersheds (Olden and Poff, 2003). The EU-Hydro-River Network Database (EEA, 2019) is used for calculating the stream density, since it is consistent and a fine-resolved dataset, which is important in this respect.

Climatic indices
LamaH includes 12 different attributes reflecting aspects of climatic characteristics (Table A4). These attributes were 320 calculated mainly from the meteorological time series of ERA5-Land for the period 01. 10.198910. to 30.09.200910. (Addor et al., 2017b. As an alternative to ERA5-Land´s potential evapotranspiration, which shows unrealistic high values (see section 4.2), the Reference Evapotranspiration (ET0) from the "Global Aridity Index and Potential Evapotranspiration (ET0) Climate Database v2" (Trabucco and Zomer, 2019), which was computed for the period 1970 to 2000, is provided. ET0 describes the atmosphere's capacity for evapotranspiration given defined vegetation characteristics. Potential evapotranspiration (PET) can 325 be derived from ET0 using correction factors for vegetation and soil properties (Allen et al., 1998;Hargreaves, 1994).
Long-term climatic characteristics are described by long-term daily precipitation (Fig. 6a, "p_mean" in Table A4), reference evapotranspiration (Fig. 6b, "et0_mean"), total evapotranspiration ("eta_mean"), and the aridity index (Fig. 6c, "aridity"), which is defined as the ratio between ET0 and precipitation. The spatial pattern of long-term precipitation sums (Fig. 6a) clearly shows an elevation gradient and blocking effects along the northern Alps. The west of the project area is characterized 330 by higher mean precipitation due to the stronger influence of oceanic climate. The relationship between mean catchment elevation ( Fig. 5b) and long-term reference evapotranspiration (

Hydrological indices
The runoff time series are characterized by 14 different attributes (Table A5), which were calculated for the period 01. 10.198910. 355 to 30.09.200910. (Addor et al., 2017b. The indices were computed for those gauges, which cover the whole period of investigation (717 gauges). However, the evaluations for the entire period of record (first 01. October to 30. September 2017) are additionally made available within the dataset. Hydrological attributes can be divided into those describing long-term characteristics, seasonality, and more short-term situations such as high and low flow.
Aridity by itself can be a good predictor for runoff occurrence in a catchment (Arora, 2002;Blöschl et al., 2013;Budyko, 360 1974). This is also shown by the similar spatial pattern of long-term runoff height (Fig. 7a, "q_mean" in Table A5, R = -0.70), and runoff coefficient (Fig. 7b, "runoff_ratio", R = -0.65) compared to those of aridity (Fig. 6c). The runoff coefficient (Q/P) is the fraction of precipitation that drains a surface after deducting evapotranspiration, groundwater flow or change in storage in the long-term. Explanations for runoff coefficients greater than 1 are given in chapter 4.2. The ratio of baseflow to total runoff can be a useful indicator for watershed classification (Sawicz et al., 2011;Fan, 2015), and is further referred as the 365 baseflow index ("baseflow_index"). It should be noted that this index is highly dependent on the method used to separate the hydrograph (Beck et al., 2013;Chapman, 1999;Eckhardt, 2008). For this reason, we used the Ladson filter (Landson et al., 2013) and the approach of Tallaksen and Van Lanen (2004) for hydrograph separation. The runoff-precipitation elasticity ("stream_elas") characterizes the inertia of change in mean runoff given a change in mean precipitation (Sankarasubramanian et al., 2001). For example, a value of 3 would indicate a change in runoff of 3% given a change in precipitation of 1%. High 370 runoff-precipitation elasticity is especially present in the eastern part of the project area (Fig. 7f). The fraction of days without discharge (not shown, "zero_q_freq") may indicate strong infiltration (e.g. Danube Sinkhole; Hötzl, 1996), artificial water withdrawal, or ceasing baseflow.
The seasonality of runoff is expressed by the attribute "hfd_mean", which shows the number of days from the beginning of the hydrologic year (01. October) to the date when half annual of the runoff volume is reached (Court, 1962). Higher number of 375 days in Fig. 7c can be explained primarily by water storage in form of snow (Fig. 6g) or glaciers (Fig. 8d). Variability in runoff ( Fig. 7d, "slope_fdc") is expressed within LamaH by the slope of the flow duration curve between the log-transformed 33 rd and 66 th runoff percentiles (Sawicz et al., 2011). High values are indicative for high runoff variability over the year, which can be caused by seasonal water storage in the form of snow (Fig. 6g) or a strong response of runoff to precipitation (Yokoo and Sivapalan, 2017). Extreme runoff events such as high or low flow are described by indices representing mean frequency (Fig.  380 7g, "high_q_freq" / Fig. 7j, "low_q_freq"), duration (Fig. 7h, "high_q_dur" / Fig. 7k, "low_q_dur") and magnitude. The threshold for high flow (at least 9 times median daily discharge) is chosen according to Clausen and Biggs (2000), and that for low flow (max. 0.2 times median daily discharge) according to Olden and Poff (2003). The magnitudes of extreme flows are expressed by the 95 th (high flow) and the 5 th (low flow) runoff percentiles. Both Q95 (Fig. 7i) and Q5 (Fig. 7l) percentiles exhibit a spatial distribution similar to that of aridity (Fig. 6c, RQ95 = -0.62 / RQ5 = -0.63). The hydrological indices (Fig. 7) are spatially 385 less smoothly distributed compared to the climatic indices (Fig. 6). The reasons might be the influence of the (non-) linear hydrological processes by locally heterogeneous catchment characteristics or uncertainties in runoff measurement (Addor et al., 2017a;Westerberg et al., 2016).

Land cover characteristics
All attributes concerning land class (Table A6)  and pixel-based classification (Bossard et al., 2000). However, the total reliability of the predecessor dataset CLC 2000 is 87.0 ± 0.8% according to a reinterpretation approach. The worst class-level reliability (< 70%) was found for sparse vegetation (CLC class 333) (Büttner and Maucha, 2006). The dominant land class within a basin delineation is derived by the majority of the intersecting raster centroids, while the fractions are derived by area share of the specific raster cells.
Agricultural land (Fig. 8a, "agr_fra" in Table A6) has high fractions in catchments with low mean slope (Fig. 5d, R = -0.89). 405 The opposite occurs for fraction of bare areas (Fig. 8b, "bare_fra"), since the vegetation period is very short at high elevated terrain and a high terrain slope fosters gravitational erosion processes. Following the CAMELS datasets, no differentiation was made between deciduous and coniferous forests when calculating the forest share. The proportion of forest is highest in the central-eastern region of the project area (Fig. 8c, "forest_fra"), where agriculture and settlement are less prevalent and the mountains are often lower than the forest line. Catchments with a relatively high proportion of glaciers (Fig. 8d, "glac_fra") 410 are mainly located in the western Eastern Alps. The influence of glaciers upon the hydrological regime is primarily apparent in the upper parts of the river regions Inn (region 3 in Fig. 1 and Table 1), Salzach (region 4) and Drava (region 18). High proportions of water surface (Fig. 8e, "lake_fra") can be explained by large lakes, which were mostly formed at the end of the last great ice age about 10 000 years ago (mainly in the Alpine foothills), or by large artificial water reservoirs (mainly in the Czech Republic). Catchments in the Vienna metropolitan area (eastern part of river region 10), as well as in the lower Rhine 415 valley (northern part of river region 1) show quite high fractions of urban area (Fig. 8f, "urban_fra"). However, most catchments (about 74%) have less than 5% urban area.

Vegetation indices
We calculated 6 different catchment attributes describing vegetation indices, which are based on Leaf Area Index (LAI), Normalized Difference Vegetation Index (NDVI), and Green Vegetation Fraction (GVF) ( Table A7). All vegetation indices are based on long-term monthly means, using either the maximum, minimum, or difference between the maximum and 425 minimum monthly means (based on 12 monthly means). Processing of the remote sensing datasets was done using the Google Earth Engine platform (GEEa, 2021;GEEb, 2021;Gorelick et al., 2017).
LAI represents vertical vegetation density and is defined as the sum of one-sided green leaf area per unit area for deciduous forests and half of the total needle area per unit area for coniferous forests. LAI was derived from the "MODIS MCD15A3H" dataset, which is a 4-day composition with 500 m grid resolution (Myneni et al., 2015). The maximum and minimum monthly 430 means were calculated for the period 01.08.2002 to 01.01.2020 using a cloud filter. The maximum monthly mean of LAI (Fig.   9a, "lai_max" in Table A7) and also the difference between maximum and minimum (Fig. 9d, "lai_diff") show a spatial correlation with the forest fraction (Fig. 8c,  NDVI is based on the "MODIS MOD09Q1" dataset with a temporal resolution of 8 days and a spatial resolution of 250 m (Vermote, 2015). The calculation was performed for the period 01.04.2000 to 01.01.2020, also applying a filter on cloudy 440 satellite images. A negative correlation is apparent between the NDVImax (Fig. 9b, "ndvi_max", R = -0.78) or the NDVImin (Fig. 9e, "ndvi_min", R = -0.84) and the mean catchment elevation (Fig. 5b).

GVF (Green Vegetation Fraction) indicates the fraction of soil that is covered by green vegetation and can
where NDVI represents the (maximum or minimum) monthly mean of NDVI, NDVIs the annual maximum NDVI of bare 445 ground, and NDVIc,v the annual maximum of vegetated ground surface as a function of IGBP land class ( Table 1 in Broxton et al., 2014). NDVIs was set to 0.09 in accordance to Broxton et al. (2014), while the spatial distribution of the IGBP land classes was obtained from the "MODIS MCD12Q1" dataset of the year 2012 (Friedl and Sulla-Menashe, 2019). As the values for NDVIs and NDVIc,v were derived for a global scale and thus do not necessarily correspond to conditions in the project area, it is possible for GVF values to exceed the normal range between 0 and 1. In order to maintain consistency, we did not constrain 450 the GVF to the normal range, however. The spatial distribution of GVFmax (Fig. 9c, "gvf_max") shows similar spatial patterns to those of LAImax (R = 0.79) as well as NDVImax (R = 0.94), while GVFdiff (Fig. 9f, "gfv_diff") tends to be higher in regions with a higher fraction of precipitation falling as snow (Fig. 6g).

Soil characteristics
LamaH includes 10 different attributes to characterize soil properties (Table A8), where 8 of them are derived from the 1 km grid sized "European Soil Database Derived Data" (ESDD; Hiederer, 2013a;2013b). ESDD is based on the "European Soil 460 Database" (ESD; Panagos et al., 2012;Panagos, 2006), while the maximum available soil water content (TAWC) in ESDD was calculated using pedotransfer functions (Hiederer, 2013a). ESDD provides soil attributes for a topsoil layer and a subsoil layer having the boundary at 30 cm soil depth. Values from these two layers were therefore aggregated, weighted by the available root depth ("root_dep" in Table A8), or in the case of TAWC also summed up. The attribute describing the depth to bedrock "bdrk_dep" is based on the layer "average soil and sedimentary-deposit thickness" of the dataset "Global 1-km Gridded 465 Thickness of Soil, Regolith, and Sedimentary Deposit Layers" (GGT; Pelletier et al., 2016). GGT has a spatial resolution of 30 arc seconds (approximately 1 km) and is derived from landform-specific models (for upland, lowland, slope, and valley floor) considering geomorphological principles, and incorporating data for topography, climate, and geology. Calibration and validation in GGT were performed using independent borehole profiles (Pelletier et al., 2016). The 3D Soil Hydraulic Database of Europe (3DSHD, Toth et al., 2017) dataset with a grid size of 250 m served as source for extracting the saturated hydraulic 470 soil conductivity ("soil_condu"). 3DSHD was derived using pedotransfer functions (Toth et al., 2015) incorporating attributes from the SoilGrids250m dataset (SG250; Hengl et al., 2017), while SG250 is based on machine learning techniques including data from about 150 000 soil profiles as well as remote sensing data for climate, vegetation, geomorphology, and lithology . Data within 3DSHD is provided for 7 different soil layers, so a depth-weighted harmonic averaging was applied. 475 The provided soil attributes in LamaH may include large uncertainties and should therefore be considered with caution for several reasons. First, the soil attributes from ESD are mainly based on extrapolated observations of soil profiles and expert estimates (ESDB, 2004). Especially in the case of heterogeneous soil conditions and large distances between soil profiles the reliability of ESD dataset must be cautioned. Data from soil profiles are also integrated in 3DSHD Toth et al., 2017)  Depth to bedrock (Fig. 10a, "bedrk_dep" in Table A8) shows similar spatial patterns as mean catchment slope (Fig. 5d, R = -485 0.56), and mean elevation (Fig. 5b, R = -0.46). About 37 % of all 859 catchments have a mean depth to bedrock of more than 1.5 m. This depth represents the maximum root-available depth in ESDD (Fig. 10b, "root_dep"). The depth available for roots tends to be higher in Germany and the Czech Republic than in other regions. If this is an indication of different measurement methods across the countries is unclear. Low available rooting depths in Austria are, according to Fig. 10b, mainly present where the fraction of carbonate sedimentary rocks (Fig. 11c) or glaciers (Fig. 8d)  organic soil content below 1%, while the highest organic contents are located in the southern German region (Fig. 10c, "oc_fra"). Further interrelationships of the various grain size fractions and the dominating bedrock are recognizable: 1) A high proportion of sand (Fig. 10d, "sand_fra") is especially prevalent where the fraction of metamorphic bedrock is also high (Fig.   11b, R = 0.47). 2) Moreover, the fraction of silt (Fig. 10e, "silt_fra") tends to be high on catchment level where a high fraction of carbonate sedimentary rock (Fig. 11c, R = 0.52) is present. 3) Finally, we can observe an increase in clay content (Fig. 10f,  495 "clay_fra") with increasing proportion of mixed sedimentary rock (Fig. 11d, R = 0.47). Soil porosity (Fig. 10g, "soil_poros") shows similar spatial patterns compared to sand fraction (R = -0.79), while saturated hydraulic conductivity (Fig. 10h, "soil_condu") tends to increase with decreasing mean catchment elevation (Fig. 5b, R = -0.63). The available soil water depth (TAWC, "soil_tawc") was determined in ESDD by including water content at field capacity, gravel content and root-available depth (Hiederer, 2013a). That also explains the high correlation of TAWC ( Fig. 10i) with the root-available soil depth (Fig.  500 10b, R = 0.94).

Geologic characteristics
We used the datasets GLiM (Hartmann and Moosdorf, 2012;Global Lithological Map) and GLHYMPS (Gleeson et al., 2014; Global Hydrogeology Maps) for deriving 16 different geologic attributes (Table A9). GLiM summarizes 92 regional geological maps in vector form and was used to extract the fractions of the different geological classes. GLiM offers 3 levels of detail, while the 1 st level species the dominant lithologic class. The optional 2 nd as well as 3 rd level further specify, for example, the 510 structure of the rock or local conditions (Hartmann and Moosdorf, 2012). For LamaH only the 1 st level of GLiM was used, which contains 16 different geological classes. The classes "evaporites", "no data" and " intermediate volcanic rocks" do not occur within the project area. The 3 most common dominant geologic classes (Fig. 11a, "gc_dom" in Table A9)  Alps) and the north-eastern border (Swabian Alb) (Fig. 11c, "gc_sc_fra"). The flysch and molasse zone (Alpine foothills and 520 central parts of the German project area) is basically characterized by a high fraction of mixed sedimentary rocks (Fig. 11d, "gc_sm_fra").
Attributes concerning permeability and porosity of the lithologic bedrock were extracted from GLHYMPS. There is a high spatial correlation between GLHYMPS and GLiM, as geologic classes of GLiM served as a starting point for assigning hydraulic properties in GLHYMPS. Huscroft et al. (2018) declares that permeability in GLHYMPS is determined only for 525 saturated conditions. GLHYMPS is only intended for regional-scale applications (i.e. spatial resolution greater than 5 km), as the influence of local heterogeneities such as fault zones can be neglected above this scale (Gleeson et al., 2014).
A high proportion of metamorphites or plutonites (mt, pa, pi in Fig. 11a) is commonly associated by low bedrock porosity (Fig. 11e, "geol_poros"). Catchments within the flysch and molasse zones in contrast exhibit relatively high porosity. High bedrock porosity is not necessarily followed by high subsurface permeability ("geol_perme"), yielding a much more 530 inhomogeneous spatial pattern in Fig. 11f than in Fig. 11e. The reason may be rock structure (2 nd stage of GLiM), which can have different impacts on permeability and porosity ( Table 1 in Gleeson et al., 2014).

(Anthropogenic) impact on runoff process and measurement
We provide 4 attributes (Table 2) in order to simplify filtering and evaluation of runoff gauges regarding any (anthropogenic) 540 impact on runoff processes or its measurement. We have tried to represent the diversity of (human) impact by 13 different types of impact ("typimpact" in Table 2). The type of (human) impact on runoff or measurement was determined primarily from gauge-metadata declared by hydrographic services (BAFU, 2020;GKD, 2020;HZB, 2020;LUBW, 2020). Additionally, publicly available information, as well as manual aerial photo evaluations were used for determination. Typical types of human impact in the project area are large water reservoirs often associated with hydropower plants and cross-basin water transfers. 545 The following types of influence were not classified because the necessary information is not consistently available, or only with great effort: 1) icing, especially at smaller rivers in winter; 2) variable channel profiles leading to inaccurate rating curves; 3) high groundwater flow in the area around the gauge; and 4) subsurface in-or outflows especially in highly karstified areas.
The hydrographs with hourly resolution in the months of January and July for the years 1990, 2005 and 2017 were additionally manually evaluated regarding systematic diurnal variations ("diur_art" / "diur_glac"). Systematic fluctuations were further 550 subdivided into those caused artificially (e.g. by storage power plants, power plants with swell operation or sewage treatment plants) and those caused naturally (snow or glacier melt). The degree of gauge impact ("degimpact") is determined using the classes 1) u -"no influence", 2) l -" low influence", 3) m -"moderate influence", 4) s -"strong influence", and 5) x -"not considered further", mostly based on the type of impact and any systematic diurnal variations. A low degree (l) was assigned to those gauges characterized by impact type D (lake with unaffected outlet) or J (herding / vegetation at gauge). Larger 555 (artificial) lakes are in case of flooding effective retention areas, and can therefore attenuate and delay the flood peak. A medium degree of impact (m) was assigned in case of impact type B (flood retention reservoir), C (lake with controllable outlet), F (emergency outlet of water reservoir), G (extreme events are influenced / not properly measured), K (fishing ponds), and L (high infiltration). An exception was made for 3 gauges at the upper Danube, which can be strongly (s) affected by full seepage during the summer months (Hötzl, 1996). Gauges with impact type A (water reservoir with all-season water filling) 560 were assigned in most cases a strong (s) degree of impact, in case of very large catchment area (selected Danube gauges) also a medium degree. An extent-based assignment of the degree of impact was finally made in case of impact type E (water withdrawals) or I (water intake). The hydrographic yearbook of Austria declares anthropogenic cross-basin water transfers by increasing or decreasing the natural catchment area of a gauge (BMLFUW, 2013). Given this information, it was possible to calculate the relative anthropogenic change in catchment size for Austrian gauges, which allows a more comprehensive and 565 objective classification. No influence (u) was assigned, if the relative artificial change of catchment size was less than 1%, while a low degree (l) of impact was noted in case of a change between 1 and 3%. Furthermore, a medium degree of impact (m) was attributed for changes up to 10%, and beyond a strong degree (s). There is no information regarding artificial changes in catchment size for gauges outside Austria. For this reason, the degree of influence was additionally determined there, and also at Austrian gauges influenced by other kinds of water withdrawal (river branches, diversions or irrigation), on the basis 570 of publicly available information, as well as aerial photo analyses. We thereby assigned mostly a strong level of impact, but in a few cases (e.g. withdrawals for drinking water) also a medium degree. Systematic diurnal variations of artificial origin (e.g. swell operation) were classified as "strongly influenced" ("diur_art" in Table 2). Obviously, a gauge or catchment area can be characterized by several types of impact. In such cases, the highest degree of impact was chosen. For gauges located in or downstream of urban areas and without any type of influence, we applied a low degree (l) of impact. The reason is a potential 575 influence by undetected water withdrawals or stormwater drains. If there is no obvious type of influence and the gauge is located above populated areas, the grade of impact was declared as unaffected (u). Finally, the degree of influence (x) was attributed for all those gauges, which 1) do not have a clearly assignable catchment area (e.g. gauges at artificial channels (impact type H) or below karstic springs), 2) are characterised by several time series (e.g. with or without consideration of mill channels) and 3) have too many gaps (> 50%) in the time series. These gauges were subsequently assigned no meteorological 580 time series or catchment characteristics due to lack of catchment delineation.
The spatial, as well as the frequency distribution of the degree of impact is shown in Fig. 12. 3.5% of 882 gauges are not influenced (u), 48% show a low influence (l), 18.9% are moderately influenced (m) and 27% are strongly influenced (s), while 2.6% belong to class (x). Low-influenced gauges are predominant in the northwest of the German project area, in the north of the Austrian central region (river region 5, 6, 7, 8, 9, 10 in Fig. 1), but also in the east (river region 16), as well as in the south 585 of Austria (east of river region 18). Strongly influenced gauges are in contrast mainly prevalent, where large water reservoirs are in operation for hydropower generation (primarily in the Alpine region), but also for seasonal water balancing or flood protection (primarily in the Czech Republic and in the north of the German project area). It should be noted that gauges located far downstream of large reservoirs may still be strongly influenced by them.   overcome the mentioned barriers. Apart from the complete territory of Austria, LamaH also includes all neighboring upstream 600 areas of the rivers flowing through Austria. LamaH contains runoff time series as well as 15 different meteorological time series (daily and hourly resolution) and over 60 attributes for 859 catchments. The 3 different catchment delineations allow investigations for individual (headwater) catchments, but also within an interconnected river network considering intermediate catchments. It is clear that LamaH also contains deficits and uncertainties, also due to the large number of data sources included. These uncertainties have been addressed. 605 Blöschl et al. (2019b) highlighted numerous open hydrological challenges, such as runoff prediction in unobserved basins (PUB). Methods based on machine learning show promising results for time series prediction (e.g., Kratzert et al., 2019a;Kratzert et al., 2018). However, uniformly structured "large-sample" datasets are helpful when applying these datadriven methods, because on the one hand the necessary preparatory work is drastically reduced and on the other hand the exchange or comparability of the modelling results is considerably facilitated. Given the scope of LamaH, we hope that this 610 dataset will serve as a solid data base for further investigations in various fields of hydrology and adjacent fields of environmental science. The high variability in the data in combination with the interconnected river network as well as the high temporal resolution of the time series could grant an improved understanding of processes in water transfer and storage, if appropriate modelling methods are used.

615
Data availability. LamaH is freely available at https://doi.org/10.5281/zenodo.4525244 (Klingler et al., 2021). The dataset is basically divided into 6 parts including basin delineation A/B/C, gauges, stream network, as well as an appendix. The first 4 parts mentioned contain shapefiles, various text files regarding the attributes as well as time series. The stream network is available with shapefiles, which contain numerous attributes, while the nomenclature of the CORINE Land Cover dataset (chapter 5.4) is deposited in the appendix. The entire folder structure, supplementary information regarding the time series, 620 and recommended citations to use are located in the folder "Info". The runoff time series of the German federal states Bavaria and Baden-Württemberg are retrospective checked and updated by the hydrographic services. Therefore, it might be appropriate to obtain more up-to-date runoff data from GKD (2020) or LUBW (2020).