Water level at the outlet (Fig. ) was measured using a VEGAPULS WL-61 optical height gauge (VEGA, Schiltach, Germany) over a geophone weir, and discharge was determined using an established rating curve . A network of four distributed meteorological stations (at 1253, 1500, 1780, and 2100 m a.s.l.; Fig. ) included precipitation and air temperature, mainly measured with a Lufft-WS 300/400 (Fellbach, Germany), and solar radiation, measured with an Apogee incoming radiometer (SP230, Logan, Utah, USA), mounted in an energetically autonomous wireless network with a GPRS connection Sensorscope Sarl, Lausanne, Switzerland;. At a single point (1253 m a.s.l.), a MADD tipping bucket rain gauge provided back-up measurements Yverdon, Switzerland,. Data from the Swiss automatic meteorological station network are used to fill gaps of the local observations (12 missing days in 2017). Water temperature was recorded every 15 min or more frequently at all but one sampling site (TT; see Fig. ) with Hobo temperature/light pendants and one temperature/conductivity logger for parts of the study period (Supplement, Table S2; Onset Computer Corporation, Bourne, MA, USA). One WTW-Tetracon 325 (Xylem Analytics, Germany) was installed, immersed at the outlet for the entire year, and logged temperature every minute.

At each point, temperature and EC were measured simultaneously with water sampling using a WTW probe (multi-3510 with a IDS-tetracon-925, Xylem Analytics, Germany). When temperature was not measured during sampling, it was substituted by an observation from the continuous water temperature loggers. When EC was not measured in the field, it was post-analyzed using a glass 6 mm probe in the laboratory (Jenway 4510, Staffordshire, UK).

A total of 135 composite rain samples were collected using funnels flowing into insulated bags at three locations corresponding to the rain gauges (1253, 1500, and 2100 m a.s.l.; see Fig. ) and emptied approximately weekly or biweekly during the summer seasons between June 2016 and November 2018 (3). In total between February 2016 and April 2018, 199 snow samples were collected in a distributed manner across the catchment, with between 3 and 26 samples per month in the winter, and included the entire snowpack depth when possible (see additional details in Table S3).

The snow-covered area was calculated as a ratio from 0 to 1 over the non-forested part of the catchment using 21 , images and 24 Sentinel-2 images between August 2016 and December 2017 . Time steps between the images ranged from 0 to 43 d, averaging 11 d. A linear interpolation of the combined time series was used to provide a fractional area covered with snow per sampling date. Satellite images were taken on average 7 d before or after sampling, though some were on sampling days and others were up to 20 d prior or after.

Both baseflow and dq/dt were calculated from the discharge data at the outlet at the highest recorded streamflow resolution (10 min). Baseflow was defined as the lowest value in a 10 d moving window. On the other hand, the change in discharge over time, dq/dt, was determined as the slope of the linear regression fitted on a 48 h window preceding the moment that each sample was collected. This 2 d window was chosen to reduce sensitivity to the noise of streamflow records, but it also corresponds to the persistence of eDNA in the environment while also preventing the domination of the diurnal streamflow cycle. For this study, dq/dt is a proxy for how the stream network contracts and expands in combination with overland flow. Rather than assuming a particular scaling relationship, we took the dq/dt at the outlet to be a good proxy for the upstream network contraction, which is justifiable as significant, and rapid increases in discharge correspond to precipitation events covering large parts of the catchment (i.e., are not concentrated on some small sub-catchment) in the Vallon de Nant . Detailed results of a sensitivity analysis regarding the optimal time window length are reported in the Supplement (Fig. S1).

Water was sampled at the 10 stream sites directly following the sampling of eDNA using 12 mL amber screw vials with a solid polypropylene cap and a silicone rubber/PTFE septa (BGB Analytik, Boeckten, Switzerland) at the sampling points (Fig.  and Supplement Table S3). Water was analyzed for stable isotope compositions of oxygen and hydrogen using a Picarro Wavelength-Scanned Cavity Ring-Down Spectrometer (WS-CRDS) 2140-i (Santa Clara, California, USA). Samples were injected a minimum of six times, and the last three measurements were averaged to calculate a raw value. Values were corrected according to a standard curve determined with internal standards, which are regularly calibrated against international standards VSMOW (Vienna Standard Mean Ocean Water) and SLAP (Standard Light Antarctic Precipitation) of the IAEA (International Atomic Energy Agency). Delta units of isotope compositions are reported .

All available samples of precipitation were used to determine a local meteoric water line (LMWL; Fig. S2), which was used to determine the line-conditioned excess (lc-ex, L), or offset from the LMWL: 1L=δ2H-7.82⋅δ18O-10.47. A positive value of lc-ex falls above the LMWL and a negative one below . The value of lc-ex reflects a combination of the source water and the physical processes that have occurred since precipitation, with more negative values indicating more evaporation and condensation .

eDNA was sampled by filtering 250 mL of stream water on each of four filter replicates per site using GF/F filters (25 mm diameter, 0.7 µm, Whatman International Ltd., Maidstone, UK) directly in the field at the 10 sites immediately before stable isotopes were sampled. Following filtration, we dried the filter by squeezing air through it, rolled the filter with tweezers, and stored the filter in a 1.5 mL tube following the description of . If it was not possible to collect 250 mL on a single filter, we used more filters until we reached the goal of 1 L filtered water per site. The tubes were transported on ice and were then stored at -20 ∘C. We implemented a negative filter control consisting of 1 L Milli-Q water that was previously treated with UVC light, before beginning sample collection in the field, to verify a DNA-free status of materials used. Sites were sampled in all cases except during extreme avalanche conditions, snow cover of the sampling site, or when the sampling site was dry.

We extracted the DNA from the filters after all samples had been collected. We used the DNeasy^® Blood and Tissue kit (Qiagen, Hilden, Germany) following the protocol for animal tissue besides a few changes (see Supplement, Sect. S3.1). In order to reduce biases due to lab procedure, we extracted the samples in a random order. During extraction, we combined two of the four filter replicates, resulting in two extractions per sampling event at the individual sites. We included a negative extraction control, containing a filter previously treated with UVC light, for each batch of extractions, resulting in eight extraction controls. The extracted DNA was stored at -20 ∘C until further processing.

Generally, eDNA exists in low concentrations in the environment. Thus, to sequence the DNA region of interest, it first needs to be amplified. In our study, we targeted a subsection of a previously identified barcoding gene region for eukaryote species , a 313 base pair fragment of the cytochrome oxidase I gene COI;Table S4. We used the Illumina dual-barcoded two-step protocol to amplify and sequence our samples consisting of polymerase chain reactions (PCR), replicates to minimize stochasticity effects, implementation of additional controls, indexing, and pooling (see Supplement, Sect. S3.2, for individual steps).

eDNA laboratory methods are optimized for the detection of low amounts of DNA and thus very sensitive to contamination. In order to reduce false positives, we followed the previously described laboratory precaution for work with eDNA . Reused field material, such as filter housings and syringes, was soaked 40 min in 2.5 % sodium hypochlorite (i.e., bleach), rinsed with deionized water, and treated with UVC light prior to reuse in the field. Furthermore, four types of controls were sampled during the field (filtration) and laboratory procedure (extraction, PCR negative and PCR positive; see Supplement, Sect. S3.3), which were run alongside the eDNA samples during the laboratory workflow and are later used for data cleaning.

The main goal of the bioinformatic analysis is to remove errors due to sequencing techniques and regroup sequences into operational taxonomic units (OTUs, operational taxonomic units). High-throughput techniques produce errors during sequencing, and in order to increase appropriate sequence identification, the resulting reads are cleaned and undergo an initial check for quality (see Supplement, Sect. S3.4). We determined variations of the amplified DNA sequences and reduced the influence of sequencing errors with UNOISE3 usearch v10.0.240;. The detected sequence variants were clustered additionally at 99 % sequence identity to reduce diversity and to account for possible amplification errors. This clustering of the sequences results in so-called ZOTUs (i.e., zero-radius operational taxonomic units). In this sense, a ZOTU is a cluster of DNA sequences that are very similar and can be seen as a rough proxy for a species. As a final step, the ZOTUs were assigned to a taxonomic name if possible by comparing it to taxonomic databases (blast 2.3.0 and usearch v10.0.240, tax filter=0.9). However, databases are still highly incomplete , and thus matches cannot be expected at a high rate.

The three types of negative controls produced during the field and laboratory procedure were used to remove noise caused by possible contamination. In a first step, we addressed possible contamination following : for each ZOTU found in one of our negative controls, we calculated a relative frequency by dividing the sum of reads for an individual ZOTU by the sum of all ZOTUs in the negative controls, which was then used as a threshold to clean the field samples (see Supplement Sect. S5). Any ZOTU with a frequency below the threshold was removed from further analysis. We then merged the two eDNA replicates for each sampling point and removed any samples that were below 20 000 reads (2 out of 107 samples), likely originating from errors in the field or laboratory work.

The number of species, also called richness, is a simple measure of biodiversity in ecology. However, richness is strongly affected by the sampling effort (i.e., the higher the sampling effort, the more species that will be found). Therefore, in order to compare diversity measures between sites, data were rarefied, a method to standardize effort . The number of reads per sample is used as a proxy for sampling effort. Consequently, we rarefied the data to the sequencing depth, or number of reads, of the lowest field sample (26 759 reads). The Illumina MiSeq run resulted in 12.5 million reads and 9858 ZOTUs after the bioinformatics pipeline, of which 2.9 million reads and 9635 ZOTUs remained after removal of contamination and rarefaction (see Table S5 in the Supplement for detailed information).

In order to test the capacity of eDNA to provide insight into the flow path of the water, we assessed whether the eDNA composition varied according to sampled water type. In the study catchment, we identified three characteristic aquatic environments (tributaries, springs, and the main channel) representing unique habitats, each with corresponding eukaryotic communities. We first used a Kruskal–Wallis test to compare the mean ranks of six variables (δ18O, δ2H, lc-ex, electrical conductivity, water temperature, and rarefied ZOTU richness) to determine if there were statistically significant differences in richness between the seasonally spread sampling days of the different water types: the main channel (M), springs (S), and tributaries (T) . Second, we tested whether we can distinguish eDNA samples from the three water types based on the identity of the detected eDNA sequences (i.e., ZOTUs' sequence). For this analysis, we used a non-metric multidimensional scaling NMDS; approach, an ordination method that represents pairwise comparisons, or rank orders, of sampling sites in a two-dimensional space. Sites clustered close together are more similar in terms of their ZOTUs than sites further apart, which enabled us to identify if communities detected with eDNA cluster according to water type or hydrologic parameters. We used weak ties (stress type 1), allowing ties to be broken where equal observed dissimilarities exist . We calculated the stress of the NMDS according to as a measure of how well the configuration matches the observed data. The NMDS was plotted for the individual sampling days separately, to observe the development of differences over the sampling season.

Dissimilarity measures used for the NMDS are based on the sequence identities of the ZOTUs and were calculated to do a pairwise comparison between sites. We used an unweighted UniFrac as a dissimilarity input to the NMDS, which is based on presence/absence and the phylogenetic tree (i.e., the evolutionary relationships among ZOTUs based on their sequence similarity) of the ZOTUs. We further fitted variables of interest (elevation, EC, water temperature, total daily solar radiation, baseflow, daily snow cover area, dq/dt, δ18O, and lc-ex) onto the NMDS ordination two-dimensional space with a commonly used method that maximized correlation between the ordination and the corresponding environmental variables function envfit, R package vegan;version 2.5-4, to identify the directions towards which the vectors change most rapidly in the presented ordination space. Next, we focused on ZOTUs that have a taxonomic assignment and used the R package indicspecies to identify candidate taxa that are statistical significantly associated with a particular water type or combinations of those by using a permutation test .

Due to the spatiotemporal variability in water sources, Alpine catchments are a mosaic of different water types and habitats, indicated by a high range of physical and chemical processes . We explored the influence of water source contributions on eDNA sampled over a season associated to two variables: the change in discharge at the catchment outlet, dq/dt and subsurface exchange as revealed by changes in electrical conductivity (EC) in springs . We used rarefied ZOTU richness, calculated for each individual sampling site and time point as a simple measure of eDNA diversity. We utilized a generalized linear mixed-effect model GLMM; function glmer in the R package lme4; with ZOTU richness (R) as the response variable, change in discharge (D=dq/dt) and the water type (W) as fixed effects, and the sampling site (S) as a random effect. The univariate response variable Y=R was modeled with vectors X=[S] and Z=[D,W,DW] of the explanatory variables associated with the fixed effect (Z) and the random effect (X). We assumed a Poisson distribution and the conditional mean E(Y|X,Z)=μ=g-1(η) related to the linear predictor by a log-link function with inverse h=g-1 . The identity of the sampling sites was defined as a random effect in an intercept model, to take sampling site variability into account.

We compared the models with and without interaction of the two explanatory variables (D, W). The testing of the interaction allows us to identify if the intercept and slope of richness vary according to the water type, whereas a purely additive approach would only allow us to test whether the intercepts were different, while assuming the slopes were the same. We selected the best model based on a χ2 test, classically used to compare model fits. Since the change in discharge, dq/dt, was negatively correlated with ZOTU richness in springs, we performed a second analysis where we replaced dq/dt with EC (E), in the above model. EC is an indicator of the subsurface exchange that defines the water type of springs likely better than dq/dt. EC was scaled and centered for the model due to large eigenvalue ratios.

In order to examine the contribution of ZOTUs from upstream springs and tributaries to the richness in the main channel downstream, we calculated overlap in ZOTUs. We identified ZOTUs that were unique to samples from one water source, shared by only two, or all three water sources. If a ZOTU was only found in spring samples or in spring and main channel samples, but never in any tributary samples over the whole data set, the ZOTU origin was assigned to springs. If only found in tributaries or in tributaries and main channel samples but never in any spring sites, its origin was assigned to tributaries. The ZOTU's presence in the main channel did not influence this decision, since we assumed the main channel contains eDNA of both tributaries and springs, either because it may be shared due to the similar habitat or due to downstream transport of eDNA .

Next, we partitioned the ZOTUs in main channel samples into those originating from springs and tributaries, providing a count of how much springs and tributaries, respectively, are contributing to the main channel. We then plotted the relative contributions separately against dq/dt and EC, as these were identified as major drivers in our prior analysis of ZOTU richness (see Sects.  and ). To confirm this relationship, we computed a generalized linear model (GLM) with the relative contribution (C) as a response variable, dq/dt(D) and the origin of the ZOTU (O) as explanatory variables: 2C=β0+β1⋅D+β2⋅O+ϵ, where ϵ comes from a quasi-binomial error distribution. This choice results from the fact that (i) binomial error distributions are appropriate for proportional variables; and (ii) our data set shows overdispersion, i.e., higher variability than what can be explained by the standard binomial error distribution. We compared the models with and without interaction of the two explanatory variables and selected the best model with a F test. As before, we also tested a model with EC (E) as n explanatory variable replacing D, as it might be more important to recover the input of springs: 3C=β0+β1⋅E+β2⋅O+ϵ.

Despite all variability, there was a positive relationship between the change in streamflow, dq/dt, and detected eDNA richness in the tributaries and main channel. High-discharge events resulted from either rainfall or snowmelt and cause expansion of the surface channel network, or in other words, increased overland and subsurface flow , a well-established consequence . Likely, this expansion put water in contact with additional surface areas, likely resulting in the wash-in of terrestrial-origin eDNA and causing an increase in the diversity signal. However, due to the lack of taxonomic information, the exact origin of these ZOTUs (i.e., terrestrial origin or not) cannot completely be identified. Conversely, when streamflow decreased, corresponding to an overall contraction of the network, reduced overbank flow, less exchange with the hyporheic zones, and diminution of channel bank erosion likely led to more characteristic DNA of the sampled water types in our record (Fig. ). The tributaries can be considered to be the suppliers of the ecohydrologic system, gathering genetic diversity when they contain more water and hydrologic connectivity increases. The richness at spring sites decreased with increasing dq/dt, perhaps because contributions from precipitation diluted the eDNA signal. This is further supported by the observed positive relationship between dq/dt and an increased number of ZOTUs originating from tributaries but not from springs.

Detected richness in springs was relatively stable through the seasons (as were stable isotopes and temperature); we even observed a slight negative relationship between richness and dq/dt. This is not surprising, as the springs have a relatively constant flow and stable surface area. However, EC increased over the season in springs, indicating elevated cumulative contact with subsurface substrate , which also positively correlated with richness. Our models suggest that, similarly to surface water, richness in springs could be increased by interaction with other habitats, in this case potentially subsurface areas (indicated by the higher EC). However, we found a relatively small contribution of eDNA from springs to the main channel, at least during our sampling season, and no relationship between EC and the contribution of ZOTUs to the main channel from springs. The quantity of the contribution from springs to the main channel, in terms of both eDNA and water, is most likely greater and thus more visible during winter baseflow conditions , which were omitted from our study.