A river carries eDNA away from its source. Experiments with caged fish in small streams showed that the signal can be detected downstream of the animals and that the distance and strength of detection change with discharge (Jane et al., 2015). In a larger system, DNA of lake-dwelling invertebrates was recovered several kilometres down the outflowing river (Deiner and Altermatt, 2014). A positive sample therefore indicates that the species is somewhere upstream within a distance that depends on flow, and a sampling layout must be read with that in mind. Most fish eDNA is not free molecules but sits in cells and fragments of tissue: size fractionation of carp eDNA found the bulk of it in particles between about one and ten micrometres (Turner et al., 2014), which is why filtration, and the choice of pore size, matters so much.

Suspended sediment affects every stage of the method. Fine filters block after a fraction of a litre, so the volume screened per sample falls and with it the chance of capturing a rare molecule. Work in turbid tropical waters found that coarser filters allowed much larger volumes to be processed before clogging (Robson et al., 2016), and alternatives to conventional filtration have been compared for turbid water specifically (Williams et al., 2017). Clay and organic matter retained on the filter are extracted along with the DNA and can inhibit PCR, producing false negatives unless inhibition is tested for and removed (McKee et al., 2015). Finally, DNA bound to particles settles. Carp eDNA was found to be more concentrated in bed sediment than in the overlying water and to persist there longer (Turner et al., 2015), so that a flood or a careless boot can return old DNA to the water column.

A species that is present may be missed at three levels: its DNA may be absent from the water at the site on the day, absent from the sample taken, or unamplified in the PCR replicate. Site occupancy models estimate the probability of presence while accounting for detection below one (MacKenzie et al., 2002). Schmidt et al. (2013) showed how the approach applies to eDNA, and hierarchical versions with separate parameters for sites, samples and PCR replicates are implemented in freely available software (Dorazio and Erickson, 2018). These models require replication at each level, which is the main reason the design below specifies it.

A species-specific assay is preferable to metabarcoding when one rare fish is sought, because all the sequencing effort goes to the target. The primers and probe must be tested in silico and against tissue DNA of every related species in the basin, since an assay with too few mismatches to a common relative will produce false positives (Wilcox et al., 2013). The limit of detection and limit of quantification should be determined and reported in the standard way (Klymus et al., 2020), and assay details recorded in line with the MIQE guidelines (Bustin et al., 2009). Reference sequences for poorly known river fish are often missing, and obtaining vouchered tissue may be the first task.

Before the survey proper, the full protocol is run at a small number of places where the species is known to be present, such as a fishing ground with recent catches, and at one where it is certainly absent. The pilot shows what volume can be filtered at the prevailing turbidity, whether extracts are inhibited, and roughly what proportion of samples and PCR replicates are positive where the fish exists. Those proportions feed the replication plan in Step 4. A pilot that yields no positives at a known site signals a problem with capture or amplification that must be solved before any absence elsewhere can be interpreted.

Sites are placed in relation to habitat and to hydrology. Because the signal travels downstream, sites are spaced so that each can be interpreted as integrating the reach above it, and are positioned below likely habitat such as deep pools, confluences and rocky runs. Sampling immediately below a tributary junction is avoided unless the tributary is itself of interest, since incomplete mixing can hide or misplace a signal. Where the channel is wide, samples are taken from more than one point across it. Turbidity, temperature, pH and an estimate of flow are recorded at each visit as covariates of detection.

At each site several independent water samples are taken from flowing water, upstream of the boat or the operator and clear of the bed. Each is passed through a filter of relatively coarse pore size, or through a pre-filter followed by a finer one, until flow stops, and the volume passed is recorded. Using two or more filters per sample and pooling them at extraction raises the volume screened. The number of samples per site is set from the pilot using the reasoning of Furlan et al. (2016), who give a framework for calculating the sensitivity of an eDNA survey from detection probabilities at each level. Where pilot detection is low, more samples per site are worth more than more PCR replicates per sample.

Filters are preserved at once, in ethanol or a lysis buffer, or kept on ice and frozen the same day. Equipment that touches water is single-use or decontaminated with dilute bleach between sites. A field blank of clean water is filtered at every site with the same equipment and carried through the whole process. These measures are standard (Goldberg et al., 2016), but they deserve more care on a muddy river bank, where sediment on hands, ropes and boat decks is a ready vehicle for DNA from the previous site.

Extraction should use a method shown to cope with sediment-laden filters, and extraction blanks are included in each batch. Every reaction carries an internal positive control so that inhibition shows as a delay or failure of the control. Inhibited extracts are cleaned with an inhibitor-removal column or diluted and run again; both remedies have costs, since columns lose some DNA and dilution lowers the target concentration (McKee et al., 2015). Each extract is run in several PCR replicates with no-template controls on every plate. A subset of positive amplicons is sequenced to confirm identity. The criterion for calling a sample positive is fixed before the survey and stated in the report.

The detection histories, sites by samples by PCR replicates, are analysed with a multiscale occupancy model (Dorazio and Erickson, 2018). Turbidity and filtered volume enter as covariates of the probability that a sample contains target DNA, and the inhibition result can enter at the PCR level. The output of the survey is then an estimated probability of occupancy for each site with its uncertainty, together with the cumulative probability of detection achieved by the effort applied. A site with no detections is reported with the probability that the species was present and missed, not as an absence.