The importance of measuring streamflow is underpinned by the reality that it is the only truly integrated representation of the entire catchment that we can plainly observe (McCulloch, 1996). Traditional streamflow measurement approaches relying on sophisticated sensors (e.g., pressure transducers and acoustic Doppler devices), site improvements (e.g., installation of weirs or stable cross sections), and discharge measurements performed by specialists are often necessary at key observation points. However, these approaches require significant funding, equipment, and expertise and are often difficult to maintain, and even more so to scale (Davids et al., 2017). Consequently, despite growing demand, the amount of streamflow data being collected continues to decline in several parts of the world, especially in Africa, Latin America, Asia, and even North America (Hannah et al., 2011; Van de Giesen et al., 2014; Feki et al., 2017; Tauro et al., 2018). Specifically, there is an acute shortage of streamflow data in headwater catchments (Kirchner, 2006) and developing regions (Mulligan, 2013). This data gap is perpetuated by a lack of understanding among policy makers and citizens alike regarding the importance of streamflow data, which leads to persistent funding challenges (Kundzewicz, 1997; Pearson, 1998). This is further compounded by the reality that the hydrological sciences research community has focused much of its efforts in recent decades on advancing modeling techniques, while innovation in methods for generating the data these models depend on has been relegated to a lower priority (Mishra and Coulibaly, 2009; Burt and McDonnell, 2015), even though these data form the foundation of hydrology (Tetzlaff et al., 2017).

Considering these challenges, alternative methods for generating streamflow and other hydrological data are being explored (Tauro et al., 2018). For example, developments in using remote sensing to estimate streamflow are being made (Tourian et al., 2013; Durand et al., 2014), but applications in small headwater streams are expected to remain problematic (Tauro et al., 2018). Utilizing cameras for measuring streamflow is also a growing field of research (Muste et al., 2008; Le Coz et al., 2010; Dramais et al., 2011; Le Boursicaud et al., 2016), but it is doubtful that these methods will be broadly applied in headwater catchments in developing regions soon because of high costs, a lack of technical capacity, and the potential for vandalism. In these cases, however, involving citizen scientists to generate hydrologic data can potentially help fill the growing global hydrological data gap (Fienen and Lowry, 2012; Buytaert et al., 2014; Sanz et al., 2014; Davids et al., 2017; van Meerveld et al., 2017; Assumpção et al., 2018).

Kruger and Shannon (2000) define citizen science as the process of involving citizens in the scientific process as researchers. Citizen science often uses mobile technology (e.g., smartphones) to obtain georeferenced digital data at many sites, in a manner that has the potential to be easily scaled (O'Grady et al., 2016). Turner and Richter (2011) partnered with citizen scientists to map the presence or absence of flow in ephemeral streams. Fienen and Lowry (2012) showed that water level measurements from fixed staff gauges reported by passing citizens via a text message system can have acceptable errors. Mazzoleni et al. (2017) showed that flood predictions can be improved by assimilating citizen science water level observations into hydrological models. Le Coz et al. (2016) used citizen scientist photographs to improve the understanding and modeling of flood hazards. Davids et al. (2017) showed that lower frequency observations of water level and discharge like those produced by citizen scientists can provide meaningful hydrologic information. Van Meerveld et al. (2017) showed that citizen science observations of stream level class can be informative for deriving model-based streamflow time series of ungauged basins.

While the previously referenced studies focus mainly on involving citizen scientists for observing stream levels, we were primarily concerned with the possibility of enabling citizen scientists to take direct measurements of streamflow. Using keyword searches with combinations of “citizen science”, “citizen hydrology”, “community monitoring”, “streamflow monitoring”, “streamflow measurements”, “smartphone streamflow measurement”, and “discharge measurements”, we found that research on using smartphone video processing methods for streamflow measurement has been ongoing for nearly 5 years (Lüthi et al., 2014; Peña-Haro et al., 2018). Despite the promising nature of these technologies, we could not find any specific studies evaluating the strengths and weaknesses of citizen scientists applying these technologies directly in the field themselves.

Etter et al. (2018) evaluated the error structure of simple “stick method” streamflow estimates (similar to what we later refer to as the float method) from 136 participants from four streams in Switzerland. Participants estimated cross-sectional area with visual estimates of stream width and depth. Floating sticks were used to measure surface velocity, which was scaled by 0.8 to estimate average velocity. Besides this study, we could not find other evaluations of simple streamflow measurement techniques that citizen scientists could possibly use. Therefore, in addition to the stick method, we turned to the vast body of general knowledge about observing streamflow to develop a list of potential simple citizen science streamflow measurement methods to evaluate further (see Sect. 2.1 for details).

Our aims in this paper were to (1) perform an initial evaluation of selected potential simple streamflow measurement methods, (2) evaluate these potential methods with actual citizen scientists, and (3) apply the preferred method at a larger scale. Our research questions are listed as follows.

Which simple streamflow measurement method provides the most accurate results when performed by “experts”?

This research was performed in the context of a larger citizen science project called SmartPhones4Water or S4W (Davids et al., 2017, 2018; https://www.smartphones4water.org/, 15 July 2018). S4W leverages young researchers, citizen science, and mobile technology to improve lives by strengthening our understanding and management of water. S4W focuses on developing simple field data collection methods and low-cost sensors that young researchers and citizen scientists can use to fill data gaps in data-scarce regions. Our aim is to partner with young researchers, local schools, and communities to use these openly available data to improve the quality and applicability of their water-related research. S4W's first pilot project, S4W-Nepal, initially concentrated on the Kathmandu Valley and is now expanding into other regions of the country. S4W-Nepal facilitates ongoing monitoring of precipitation, stream and groundwater levels and quality, freshwater biodiversity, and several short-term measurement campaigns focused on monsoon precipitation, land use changes, stone spout (Nepali: dhunge dhara) flow and quality, and now streamflow. One immediate application in the Kathmandu Valley is to improve estimates of water balance fluxes, including net groundwater pumping.

While identifying and refining methods for citizen scientists to measure streamflow may be an important step towards generating more streamflow data, these types of citizen science applications are not without challenges of their own. For example, citizen science often struggles with the perception (and possible reality) of poor data quality (Dickinson et al., 2010) and the intermittent nature of data collection (Lukyanenko et al., 2016). Additionally, there are other non-citizen-science-based streamflow measurement methods (e.g., permanently installed cameras) that may undergo rapid development and transfer of technology and thus make a significant contribution towards closing the streamflow data gap.

Additionally, the use of “citizen scientist” herein is restricted to only student citizen scientists, which are a narrow but important subset of potential citizen scientists. Our vision was to partner with student citizen scientists first to develop and evaluate streamflow measurement methodologies. Once methodologies are refined in coordination with students, we aim to partner with community members and students in the rural hills of Nepal to improve the availability of quantitative streamflow and spring flow data.

Streamflow measurement techniques suggested in the United States Bureau of Reclamation Water Measurement Manual (USBR, 2001) that seemed potentially applicable for citizen scientists included deflection velocity meters, the Manning–Strickler slope area method, and pitot tubes for measuring velocity heads. The float, current meter, and salt dilution methods described by several authors also seemed applicable (British Standards Institute, 1964; Day, 1976; Rantz, 1982; Fleming and Henkel, 2001; Escurra, 2004; Moore, 2004a, b, 2005; Herschy, 2009). Finally, Church and Kellerhals (1970) introduced the velocity head rod, or what we later refer to as the Bernoulli run-up (or just Bernoulli) method. Table 1 provides a summary of these eight simple measurement methods. For the categories of (1) inapplicability in Nepal (specifically to headwater catchments), (2) cost, (3) required training, and (4) complexity of the measurement procedure, a rank of either 1, 2, or 3 was given by the authors, with 1 being most favorable and 3 being least favorable. Theses ranks were then summed, and the three methods with the lowest ranks (i.e., Bernoulli; float; and salt dilution, or slug) were selected for additional evaluation in the field.

This investigation was carried out in three distinct phases including phase 1 – initial evaluation, phase 2 – citizen scientist evaluation, and phase 3 – citizen scientist application (Table 2).

Reference flows evaluated in phase 1 ranged from 6.4 to 240 L s-1 (Table 4; sorted in ascending order by reference flow). Elevations of measurements ranged from 1313 to 1905 m a.s.l. (meters above sea level). Salt dilution calibration coefficients (k) averaged 2.79×10-6 cm µS-1 and ranged from 2.57 to 3.02×10-6 cm µS-1. Absolute errors with respect to reference flows averaged 23 %, 15 %, and 37 %, while biases for all methods were positive, averaging 8 %, 6 %, and 26 % for float, salt dilution, and Bernoulli methods, respectively.  Standard deviations of errors were 29 %, 19 %, and 62 % for float, salt dilution, and Bernoulli methods, respectively. The largest salt dilution errors occurred for reference flows of 21 L s-1 or less (i.e., sites 1 through 7), while float and Bernoulli errors were more evenly distributed throughout the range of observed flows. Field notes from Bernoulli flow measurements for two measurements (site IDs 9 and 19) were destroyed by water damage, so Bernoulli flow and percent difference data were not available for these sites. Detailed reports for reference flow measurements along with calculations for each simplified streamflow measurement method are included in the Supplement.

Reference flows evaluated in phase 2 ranged from 4.2 to 896 L s-1 (Table 5). Absolute errors for expert measurements averaged 41 %, 21 %, and 43 %, while biases for all methods were positive, averaging 41 %, 19 %, and 40 % for float, salt dilution, and Bernoulli methods, respectively (Table 5 and Fig. 4). Standard deviations of expert errors were 34 %, 26 %, and 51 % for float, salt dilution, and Bernoulli methods, respectively. Salt dilution calibration coefficients (k) averaged 2.95×10-6 cm µS-1 and ranged from 2.62 to 3.42×10-6 cm µS-1. Measurement sites in the Dhobi watershed were pool and drop stream types, with slopes ranging from 0.076 to 0.148 m m-1. Streambeds for these sites were predominantly cobles, gravels, and sands. Smaller tributaries measured in the Nakkhu watershed (N2, N4, and N6) were also pool and drop stream types with slopes of 0.105, 0.091, and 0.055 m m-1, respectively. The remainder of the sites in the Nakkhu watershed were pool and riffle stream types with slopes ranging from 0.020 to 0.075 m m-1.

Box plots of CS Flow group errors combined with expert measurement errors for float (a), salt dilution (b), and Bernoulli (c) methods show that errors, for both expert and CS Flow groups, are smallest for the salt dilution method (Fig. 4). The number of CS Flow group measurements used to develop individual box plots ranged from 6 to 12 for each site and totalled 117 for all 15 sites. Two groups measured site D3 twice, so even though there were only 10 groups, there were 12 measurements available for comparison for this site. For the remainder of sites (except N5), problems with either capturing, compressing, uploading, or interpreting the video of EC used for determining salt dilution flow limited the number of usable measurements to less than the number of groups (i.e., 10). Absolute errors for CS Flow group measurements averaged 63 %, 28 %, and 131 %, while biases for all methods were positive, averaging 52 %, 7 %, and 127 % for float, salt dilution, and Bernoulli methods, respectively. Standard deviations of CS Flow group errors were 82 %, 36 %, and 225 % for float, salt dilution, and Bernoulli methods, respectively.

For the float method (Fig. 4a), 13 median CS Flow group errors were positive, while two sites (i.e., D3 and N7) were negative. Float expert errors (i.e., red circles) were within the interquartile range (IQR; blue boxes between the first and third quartile) of CS Flow group errors for 10 out of 15 sites. One float expert error and 21 CS Flow group errors were over 100 %. Float error medians and distributions were more variable in the Dhobi watershed than the Nakkhu watershed. For the salt dilution method (Fig. 4b), seven median CS Flow group errors were positive, while eight were negative. Salt dilution expert errors (i.e., red circles) were within the IQR of CS Flow group errors for 7 out of 15 sites. Zero salt dilution expert errors and two CS Flow group errors were over 100 %. Salt dilution error distributions were more compact for the Dhobi watershed compared to the Nakkhu watershed. For the Bernoulli method (Fig. 4c), all 15 median CS Flow group errors were positive. Bernoulli expert errors (i.e., red circles) were within the IQR of CS Flow group errors for 3 out of 15 sites. Two Bernoulli expert errors and 50 CS Flow group errors were over 100 %. Similar to float results, Bernoulli error medians and distributions were more variable in the Dhobi watershed than the Nakkhu watershed.

Overall, citizen scientists ranked the float method most favorably (43.2 % of rank 1 selections; average of blue bars) compared to Bernoulli and salt dilution methods, at 30.3 % and 26.5 %, respectively (Fig. 5). In contrast, citizen scientists ranked the salt dilution method least favorably (64.0 % of rank 3 selections; average of tan bars) compared to Bernoulli and float methods, at 18.6 % and 17.4 %, respectively. Most citizen scientists (72.7 %) thought the float method required the least amount of training (Q1), followed by the Bernoulli and salt dilution methods. Citizen scientists thought the Bernoulli method required the smallest investment in equipment (45.5 %; Q2), the fewest number of citizen scientists (54.5 %; Q3), and the least amount of data recording (42.4 %; Q4). Additionally, citizen scientists found the float method to be the least complex (48.5 %; Q5), most enjoyable (60.6 %; Q6), and safest (42.4 %; Q7). Finally, most citizen scientists (75.8 %) thought the salt dilution method was most accurate (Q8), followed by the float and Bernoulli methods. The complete results from the survey are included in the Supplement.

Observed flows from the CS Flow campaign (n=131 pre-monsoon; n=133 post-monsoon) were distributed among the 10 perennial headwater catchments of the Kathmandu Valley and ranged from 0.4 to 425 L s-1 and from 1.1 to 1804 L s-1 in the pre- and post-monsoon, respectively (Fig. 6a, b). The three locations in the Kathmandu Valley where the Nepal Department of Hydrology and Meteorology measures either water levels or flows (gauges) are included on Fig. 6a, b to illustrate the difference in spatial resolutions between the two datasets. Note that only one of the three DHM gauging stations is in a headwater catchment (i.e., Bagmati). Histograms of flow (Fig. 6c, d) and EC (Fig. 6e, f) show the increase in flows and the expected decrease in EC from pre- to post-monsoon.

Our first research question was the following: which simple streamflow measurement method provides the most accurate results when performed by “experts”? Based on phase 1 expert measurements, we found that salt dilution had the lowest absolute error (i.e., 15 %), compared to the float and Bernoulli methods (i.e., 23 % and 37 %, respectively; Table 4).

The largest salt dilution errors occurred for reference flows of 21 L s-1 or less, while float and Bernoulli errors appeared to be more evenly distributed through the range of observed flows. Because salt dilution measurements of low flows require less salt and water, it is possible that larger relative measurement errors caused while measuring these small quantities led to larger overall measurement errors. However, this is not substantiated in phase 2 results, so additional research is required in this area.

Our experience in the field was that float velocity measurements in slow-moving and shallow areas were difficult to perform. The combination of turbulence and boundary layer impacts from the streambed and the overlying air mass often made floating objects on the surface travel in nonlinear paths, adding uncertainty to distance and time measurements. In the literature, challenges with applying the float method in shallow depths are supported by USBR (2001) and Escurra (2004), who showed that uncertainty in surface velocity coefficients (i.e., the ratio of surface velocity to actual mean velocity of the underlying water column; C from Eq. 1) increased as depth decreased, especially below 0.3 m. The impacts of shallow depths on the surface velocity coefficient C should be the focus of additional research.

A primary challenge we experienced with Bernoulli measurements was keeping the flat plate at the same vertical location while rotating the plate from parallel to perpendicular to the flow direction (Sect. 2.2.3). This was usually due to the bottom of the flat plate being set on a streambed consisting of sands and gravels that could be easily disturbed during rotation. Slow water velocities, and correspondingly small changes in Bernoulli depths (Eq. 4), further compounded this issue. Adding a circular metal plate to the bottom of the flat plate used for Bernoulli depth measurements could help minimize these uncertainties.

Based on the 10 measured k values in phase 1, using an average k for all salt dilution measurements caused the largest percent difference in salt dilution flow (Eq. 2) for site 7 (8.6 % increase in flow) followed by site 19 (7.6 % decrease in flow). For phase 2, using average k values for all salt dilution measurements caused the largest percent difference in salt dilution flow (Eq. 2) for site D6 (13.7 % decrease in flow) followed by site D3 (12.6 % increase in flow). Because observed absolute error distributions from phase 1, and especially phase 2, are larger than errors introduced by using average k values (sometimes by more than an order of magnitude), we do not think our overall findings are negatively impacted by using average k values. However, because of the sensitivity of salt dilution measurements to k (Eq. 2), future work should focus on improving understanding of the variables affecting k. Specifically, spatial and temporal variability in k due to changes in stream water chemistry should be investigated prior to applying the salt dilution methodology described in this paper in other areas. For citizen science projects in other areas, we recommend that locally appropriate average k values be determined from measurements at multiple sites to understand spatial variability. Additional k measurements should also be repeated in different seasons to understand temporal variability.

Our second research question was the following: which simple streamflow measurement method provides the most accurate results when performed by citizen scientists? Based on phase 2 citizen scientist measurements, we found that salt dilution had the lowest absolute error (i.e., 28 %) compared to the float and Bernoulli methods (i.e., 63 % and 131 %; Fig. 4).

While absolute error distributions for citizen scientists followed the same trend to that of expert measurements, the relative increases in errors for float (41 % to 63 %; increase of 54 %) and Bernoulli (43 % to 131 %; increase of 205 %) methods were larger than that of salt dilution (21 % to 28 %; increase of 33 %). This could be due in part to the fact that salt dilution measurement errors may be less sensitive to a lack of field data collection experience. For example, as long as turbulent mixing conditions are present (which can be controlled by proper site selection during the experimental design phase), citizen scientists can primarily introduce errors into salt dilution measurements by (1) making mistakes in measurement or recording of amounts of salt and/or water used to prepare tracer solutions, (2) not thoroughly mixing tracer solution until all salt is dissolved, (3) not providing enough distance between salt injection and EC measurement points (recommended as 25 stream widths by Day, 1977; Butterworth et al., 2000; Moore, 2005), or (4) recording videos of EC changes that are difficult to read. Each of these sources of error can be minimized by implementing relatively easy to follow protocols such as “be sure to mix the salt and water until you cannot see the salt any longer.” In contrast, while performing float and Bernoulli measurements, citizen scientists need to accurately characterize (1) average stream depth, (2) stream width, and (3) average water velocity. Characterizing average depth and velocity requires several individual measurements, each coming with the chance of introducing measurement errors. Additionally, selecting the number of subsections required and the representative locations for each of these subsections can be difficult, even for people with extensive streamflow data collection experience. These factors may help explain the wider error distributions observed in float and Bernoulli methods compared to salt dilution (Fig. 4). Additional training might also help to close the observed differences between salt dilution error distributions and that of float and Bernoulli methods.

Our third research question was the following: what are citizen scientists' perceptions of the required training, cost, accuracy, etc. of the evaluated simple streamflow measurement methods? Based on a survey of 33 citizen scientists, we found that volunteers ranked the float method most favorably (43.2 % of rank 1 selections) compared to Bernoulli and salt dilution methods, at 30.3 % and 26.5 %, respectively (Fig. 5).

Regarding question number four from the perception survey (i.e., data-recording requirements), it is interesting to note that salt dilution received the least favorable ranking, meaning that citizen scientists perceived salt dilution to require the greatest amount of data. Our perception was that salt dilution, in terms of individual pieces of information, requires the least amount of data recording. This ranking may be explained by either (1) the amount of metadata collected about salt dilution measurements (i.e., GPS and photos of salt injection and EC measurement locations; see Sect. 2.4.2 and the Supplement for details) or by (2) citizen scientists' perception of using a digital EC meter and smartphone video for recording lots of individual pieces of data, when in some ways a video can be thought of as a single observation. Whereas results from float and Bernoulli method measurements are available immediately in the ODK form, the post-processing requirements of EC breakthrough curve data to solve for salt dilution flow may also lead to the perception that salt dilution measurements have higher data-recording requirements.

Citizen scientists ranked the float method as the safest, followed by salt dilution, and finally Bernoulli. We found this result to be somewhat counter intuitive, because salt dilution is the only method that can be performed without entering the stream, whereas for float and Bernoulli measurements the entire stream must be waded across to get depth and velocity data. Because the perception survey was performed after phase 2 evaluations where all three methods were performed consecutively, it may not have been obvious to citizen scientists that salt doses could be obtained without entering the stream from visual estimates of channel width, depth, and water velocity.

In terms of perceived measurement accuracy (question 8), 75.8 % of citizen scientists ranked salt dilution as the most accurate method. This ranking was performed before any quantitative results were reviewed. Our experience is that reading a value from a digital meter often gives an unfounded sense of measurement accuracy. The salt dilution method's perceived accuracy may be due to it being the only method that directly involves a digital measurement device (i.e., EC meter).

Expert absolute errors for float, salt dilution, and Bernoulli increased from 23 %, 15 %, and 37 % in phase 1 to 41 %, 21 %, and 43 % in phase 2. For the float method, this increase in error may be partially explained by the overall increase in flows from pre-monsoon (phase 1; average reference flow of 92 L s-1) to post-monsoon (phase 2; average reference flow of 243 L s-1). Our experience was that increased flow and velocity in high-gradient headwater streams made it more difficult to perform float measurements. This was mostly due to an increase in turbulence resulting in more nonlinear flow lines and increased relative measurement uncertainty for shorter float times (assuming distances were held constant). For the Bernoulli method, however, our hypothesis was that increased velocities would on average reduce measurement errors, because of decreased relative measurement uncertainty for larger Bernoulli depth changes. This hypothesis however was not supported by the data. The challenge of pulsing flows which require citizen scientists to visually average short-period (i.e., seconds or less) water level fluctuations may also counteract the otherwise larger Bernoulli depth changes. We do not have any explanations for the overall increase in salt dilution method absolute error from 15 % to 21 % from phase 1 to phase 2. Unlike the phase 1 results, we also do not see a concentration of larger errors at the lower reference flows in phase 2.

To proceed with phase 3, we had to select a preferred simple streamflow measurement method. Based on the results from phase 1 and 2, the salt dilution method had the lowest absolute errors, biases, and error standard deviations for both experts and citizen scientists. Therefore, from an accuracy perspective, salt dilution was the preferred approach. However, the results of our perception survey showed that citizen scientists thought the float method was most enjoyable (Q6) and required the least amount of training (Q1). Another important consideration was that salt dilution is the only method that does not require citizen scientists to enter and cross the stream and therefore can be safely performed over a broader range of flow conditions. While the enjoyment of measurements is an important motivational factor for citizen scientists, we concluded that accuracy and safety were ultimately more important. Considering all these factors, we selected the salt dilution method as the preferred approach.

Finally, our fourth research question was the following: can citizen scientists apply the selected streamflow measurement method at a larger scale? Based on measurements from pre- (n=131) and post-monsoon (n=133) in the Kathmandu Valley, citizen scientists can apply salt dilution streamflow measurements at a larger scale; however, challenges of recruiting, training, and motivating citizen scientists, along with data management issues, require further investigation.

The CS Flow campaigns provided us with a unique opportunity to evaluate the preferred salt dilution streamflow measurement method with more people at more sites. In addition to the valuable streamflow data that will help us characterize the water supply situation in the Kathmandu Valley with greater precision for pre- and post-monsoon periods, we also learned several practical lessons about how to scale citizen-science-based streamflow measurements. For example, our experience was that digitizing breakthrough curves from ODK-captured EC videos took roughly 15 to 30 min per site, depending on video length and quality. Additionally, managing EC change videos can be a significant challenge if videos are recorded at a smartphones' native resolution. In some cases, each minute of high-definition video can be nearly 100 MB. Uploading such large files, and subsequently storing and accessing them, can be challenging and costly. These difficulties can be solved by improved training and protocols regarding video collection settings and, when necessary, video compression.