Citizens and scientists collect comparable oceanographic data: measurements of ocean transparency from the Secchi Disk study and science programmes

Abstract

Marine phytoplankton accounts for approximately 50% of all photosynthesis on Earth, underpins the marine food chain and plays a central role in the Earth’s biogeochemical cycles and climate. In situ measurements of ocean transparency can be used to estimate phytoplankton biomass. The scale and challenging conditions of the ocean make it a difficult environment for in situ studies, however. Here, we show that citizen scientists (seafarers) using a simple white Secchi Disk can collect ocean transparency data to complement formal scientific efforts using similar equipment. Citizen scientist data can therefore help understand current climate-driven changes in phytoplankton biomass at a global scale.

Introduction

The ocean is a difficult environment to access for in situ study due to its scale, remoteness and challenging conditions. Although ocean science research is vital for our sustainable future¹, scientific research lags current, climate-driven ocean changes². The ocean’s phytoplankton account for approximately 50% of all photosynthesis on Earth^3,4 and temporal and spatial changes in the phytoplankton can influence marine productivity⁵, weather⁶ and climate^7,8. Monitoring the phytoplankton is therefore essential as an early indicator of regional and global ecosystem change^9,10,11,12.

Around 44% of the human population lives within 150 km of the coast¹³ and a number go to sea for work and recreation. The seafaring public often visits the same locations, whether as sailors on short day trips, commercial fishermen accessing fishing grounds, or offshore yachtsmen/women whose passages follow common routes dictated by the season, prevailing winds and currents¹⁴. Therefore, the seafaring public provides an opportunity to collect oceanographic data over varying spatial and temporal scales to contribute to scientific efforts and many now participate in marine citizen science^15,16.

The global Secchi Disk study (http://www.secchidisk.org)¹⁷ engages seafarers to use a Secchi Disk¹⁸ to collect in situ data on ocean transparency that can be used to estimate phytoplankton biomass¹⁹. The Secchi Disk study uses a simple, 30 cm diameter white Secchi Disk that is weighted and attached to a tape-measure, and lowered vertically into the water from a boat’s side. The depth (m) below the surface when the Secchi Disk disappears from sight is the Secchi depth (Z_SD), which measures ocean transparency. When the bathymetry is > 25 m depth and the distance > 1 km from shore, the primary influence upon ocean transparency is phytoplankton pigments and their breakdown products and therefore, Z_SD estimates phytoplankton biomass in the water column; re-suspended sediments and dissolved organic matter from rivers further reduce transparency and introduce optical errors in shallower water and closer inshore¹².

Marine scientists have used Secchi Disks to measure ocean transparency since 1865²⁰ and archives of Z_SD represent one of the longest-running, spatially extensive global marine datasets²¹. Recently, the Secchi Disk has fallen from widespread use among marine scientists²² due to spectrophotometric determination of chlorophyll a¹² and satellite measures of ocean colour²³. To re-establish the database of Z_SD measurements and facilitate contemporary Z_SD comparisons to historical data, the citizen science Secchi Disk study began in 2013¹⁷.

Measuring in situ data at sea from a small boat is challenging. For citizen science efforts of any type to be useful it is essential to ensure and assess data quality^24,25, and this was outlined recently in the ninth of the ten principles of citizen science outlined by the European Citizen Science Association (ECSA)²⁶. Marine and terrestrial, observational, citizen science studies have assessed data quality and shown that citizens can collect reliable observational data when given good instructions^27,28,29. We have already shown that Z_SD data collected by citizens agrees with satellite measures of chlorophyll a and satellite Z_SD estimates¹⁷. But, does Z_SD data collected by citizens of varying backgrounds using oceanographic equipment compare to Z_SD data collected by trained scientists using the same methods? Unfortunately, it is impractical to take citizens and marine scientists to sea together to measure Z_SD at the same time and place to see how they compare. To overcome this impracticality, we compare Z_SD measurements collected by citizens and scientists using Secchi Disks to independent satellite estimates of Z_SD, i.e. the comparison Secchi Disk measurements were not used to generate the satellite algorithm.

Results

We found positive relationships between satellite Z_SD and both the scientist and citizen Z_SD (Fig. 1a–l). Removing 20 samples did not strongly alter the regression lines (Fig. 1b,f,j). The red and blue polygons delineating the scientist and citizen data, respectively, as well as the regression lines, overlapped. Although a Kruskal–Wallis test revealed the regression lines were significantly different between scientist and citizen programmes (p < 0.01), these differences in slope were small and so there was only a small slope alteration when the data were merged. Correlation analysis revealed that scientist correlations were slightly higher when ocean + shelf and shelf only were considered (Fig. 1d,h,l). The difference in correlation was higher when open-ocean data only were analysed (Fig. 1d,h,l). In general, scientific data had less variance in correlation except for the shelf where a higher variance may be explained by outliers (red crosses below the regression line within the red, dashed polygons) (Fig. 1b,f).

Figure 1

Relationships between scientist (red) and citizen (blue) log₁₀(Z_SD) and satellite log₁₀(Z_SD). (a,e,i) Maps showing the positions of the scientist and citizen Z_SD data for shelf + open-ocean, shelf, and open-ocean, respectively. Maps were created using MATLAB version 2021a, https://www.mathworks.com/products/matlab.html. (b,f,j) Scatterplots, polygons and regression lines showing the relationship between scientist, citizen, and satellite log₁₀(Z_SD) for shelf + open-ocean, shelf, and open-ocean data, respectively. Red and blue polygons delineate the scientist and citizen data, respectively, and regression lines in black represent combined scientist and citizen data. (c,g,k) Histograms of regression coefficients between scientist, citizen and scientist/citizen (black) and satellite log₁₀(Z_SD) data for shelf + open-ocean, shelf, and open-ocean, respectively. (d,h,l) Histograms of linear correlations between scientist, citizen, combined scientist and citizen (black), and satellite log₁₀(Z_SD) for shelf + open-ocean, shelf, and open-ocean data, respectively.

Next, we analysed the data from WOOD, WOD, ICES and the citizens, separately (Fig. 2a–l). This analysis revealed that outliers could also be detected among scientific programmes (red crosses on the left below the regression line inside the red polygon, Fig. 2b,f). As before, all the regression lines were significantly different (Kruskal–Wallis test, p < 0.01). We noticed large variability in the regression lines, especially for WOOD when open-ocean data was considered; due to a low number of degrees of freedom. The histogram of the category of slopes showed a more significant separation and less overlap for all programmes in shelf seas than the open-ocean, which could reflect regional peculiarities (Fig. 2c,g,k). The histograms for each programme and for open-ocean + shelf, shelf and open-ocean (Fig. 3a–c) showed that part of the difference in regression coefficients may be due to location rather than the reliability of either citizen or scientist Z_SD. We also found a difference in the correlations between open-ocean and shelf. With the exception of the open-ocean where citizen correlations were slightly lower than WOD, correlations involving citizen data were higher than other programmes (Fig. 2d,h,l).

Figure 2

Relationships between WOD (magenta), WOOD (green), ICES (red) and citizen (blue) log₁₀(Z_SD) and satellite log₁₀(Z_SD). (a,e,i). Maps showing the positions of the WOD, WOOD, ICES, and citizen Z_SD data for shelf + open-ocean, shelf, and open-ocean, respectively. Maps were created using MATLAB version 2021a, https://www.mathworks.com/products/matlab.html. (b,f,j) Scatterplots, with red and blue polygons delineating the scientist and citizen data respectively, and regression lines showing the relationship between WOD, WOOD, ICES, citizen and satellite log₁₀(Z_SD) for shelf + open-ocean, shelf, and open-ocean, respectively. (c,g,k) Histograms of regression coefficients between WOD, WOOD, ICES, citizen and satellite log₁₀(Z_SD) for shelf + open-ocean, shelf, and open-ocean, respectively. (d,h,l) Histograms of the linear correlations between WOD, WOOD, ICES, citizen and satellite log₁₀(Z_SD) for shelf + open-ocean, shelf, and open-ocean, respectively.

Figure 3

Histograms of Z_SD for the three scientific programmes and for citizens. (a–c) Histograms for WOD (magenta), WOOD (green), ICES (red) and citizen (blue) Z_SD data for shelf + open-ocean, shelf, and open-ocean, respectively.

When we compared the citizen and scientist data for systematic under- or overestimation using satellite Z_SD as a reference, scientists always highly overestimate satellite Z_SD as shown by the percentage of overestimates that are consistently well above 50% (Table 1, Fig. 4). We found that citizens tend to slightly underestimate satellite Z_SD when all data are considered (44.5%) and to a greater extent when Z_SD is greater than or equal to 25 m (36.4%). Both citizens and scientists overestimate Z_SD compared to satellite Z_SD when the measurement is less than 25 m. In all cases however, the under- or over estimation is less for citizens than for scientists. In contrast to the percentage of overestimates, the mean deviation (Table 1) is a more quantitative measurement of the degree of under- or overestimation and can be influenced by individual deviation (the individual differences between in situ and satellite Z_SD measurements). The mean deviation shows nearly the same results as the percentage of overestimates although when we consider all Z_SD together the underestimation by citizens (− 1.6 m) is slightly higher than the overestimation made by scientists (1.5 m) (Table 1). (The discrepancy observed between the percentage of overestimates and the mean deviation is likely due to a few large, individual deviations that influence the latter estimator). When we separated measurements of Z_SD below and above 25 m the mean deviation is smaller in absolute value for citizens in agreement with what is shown by the percentage of overestimates.

Table 1 Indirect comparison of citizen and scientist Z_SD data with respect to under and overestimation and closeness.

Figure 4

Relationships between citizen and scientist in situ Z_SD, and satellite Z_SD. (a–c) Scatterplots showing the relationship between scientist (red crosses), citizen (blue circles), and satellite Z_SD for all the whole dataset, where Z_SD < 25 m and where Z_SD ≥ 25 m, respectively. Black dashes represent the line y = x. (d–f) Histograms of the mean square deviation (m²) for the whole dataset, where Z_SD < 25 m and where Z_SD ≥ 25 m, respectively. Orange, blue and magenta bars represent scientist, citizen and where scientist and citizen data overlap, respectively.

The mean square deviation (here, a measure of the level of the closeness between in situ and satellite Z_SD) revealed that Z_SD measured by scientists are consistently closer than Z_SD measured by citizens. Histograms prepared to examine the mean square deviation (Fig. 4d–f) suggest that a few observations made by citizens might cause the higher mean square deviation (Table 1). When we remove categories above 100 m² the mean square deviation for citizens reduced by more than 50%, from 39.2 to 18.9 m², and became very close to the scientist value (16.6 m²). Nevertheless, the degree of closeness between satellite and in situ Z_SD is higher for scientists than for citizens. When the first 3 categories of the histogram are considered together (Fig. 4d) they comprise 87% of the scientific data and 70% of the citizen data, which is a difference of 17%.

Discussion

The Secchi Disk is among the simplest pieces of oceanographic equipment with its method of use, optical properties, limitations and variability well understood^18,30; these features are taken into account in the instructions provided to citizens in the Smartphone application called Secchi that accompanies the citizen science study¹⁷. While our method of assessing citizen and scientist Z_SD data by comparison to satellite Z_SD is indirect and so imperfect (ideally citizens and scientists would go to sea together to make simultaneous measurements for such a comparison), our results show that scientists and citizens take reliable in situ measurements of Z_SD at sea.

Although we found that citizens tend to underestimate satellite Z_SD, scientists tend to overestimate to a greater extent (Table 1) when compared to satellite Z_SD. We do not know the cause of this under- and overestimation and cannot exclude this is a chance result, and because citizens and scientists under- and overestimate respectively, it is not possible at present to determine a correction factor; we nevertheless hope this can be resolved when more data is evaluated from both scientist and citizen programmes. We found higher variability in the mean square deviation for citizens and a 17% difference between citizen Z_SD and scientist Z_SD (when scientist data was combined from the three scientist datasets we used) (Table 1). When we performed the analysis on log₁₀ transformed data the differences between citizens and scientists were much less apparent and to see them we need to look at the third digit of the mantissa (Supplementary Table S1).

The difference in the mean square deviation between citizens and scientists could be explained by several factors. For example, the three science programmes we used likely involve measurements collected by just a few individual scientists (for example on a single cruise), which will reduce the amount of individual observer variability that is inherent in the in situ measurement of Z_SD³⁰ when compared to the citizen data where many individual participants each collect a few measurements. Interestingly, when we excluded categories of mean square deviation above 100 m² the closeness between citizens and scientists improved greatly. Consequently, as the citizen data set increases in size, we hope to be able to conduct analyses at a participant level to better understand individual variability and to remove outliers³¹. (The comparison between scientists and citizens may also change when data from many scientific programmes are considered and over longer time periods and larger spatial scales). Other influences on the data may include the fact that scientific measurements may be focused on a particular water body and so spatial heterogeneity will be different compared to the citizen data (Fig. 1a,e,i) and the fact that research vessels are likely to be more stable platforms compared to small boats and so measurements are likely to be easier to take. Finally, and importantly, scientists interact and train each other, which is different to a global citizen science study where similar training is impractical. Consequently, even though instructions are given to citizens in the Secchi smartphone application upon how to take Z_SD measurements, scientists are likely to show less variability because of their more formal training.

When we considered spatial heterogeneity, the Z_SD collected by citizens and scientists showed a better relationship to satellite Z_SD in the open-ocean (Fig. 2j,k). The poorer relationship found in shelf seas between in situ and satellite Z_SD (Figs. 1a,e,I, 2a,e,i) may therefore be due to the inherent properties of measurements made in specific oceanographic environments (Fig. 3) rather than the reliability of the Secchi Disk measurement in the hands of either citizens or scientists. Most likely, it is due to the greater spatial and temporal heterogeneity of shelf seas (where there may also be abiotic factors), which may all lead to an imperfect match between in situ and satellite Z_SD data due to violations in the assumptions underlying the satellite algorithm. The small overlap between volunteers (citizens) and professionals (scientists) (Fig. 1) when all the scientist data are combined together, compared to the clear overlap when we split the different science data sets (Fig. 2), may consequently, be explained by data originating from different ocean basins and environments, which is likely to alter in future when the data is more widespread.

Satellite estimates of Z_SD are influenced by the in situ Z_SD datasets used to generate the satellite algorithm and so will have inherent errors and biases. The Morel et al.³² algorithm we used is based upon a relationship between chlorophyll a and Z_SD for clear, open-ocean waters and a specific, scientist-collected, in situ Z_SD dataset where the majority of measurements were collected in the summertime in the northern hemisphere. Validation using MODIS Level-3 composite data and in situ Z_SD resulted in adjusted coefficients as the modelled values tended to exceed the measured values³², and it is these adjusted coefficients that have been used in the implementation, in other words, the algorithm was tuned to match the specific in situ Z_SD measurements.

In the open-ocean, satellite Z_SD can overestimate in situ Z_SD because the former assumes the layer extending down to Z_SD is homogenous. In contrast, chlorophyll a always varies with depth³²; chlorophyll levels are typically lower at the surface than at depth under stratified open-ocean conditions and so satellite chlorophyll underestimates the depth-averaged value. Although a recent algorithm to calculate Z_SD from current satellite sensors has improved the relationship between satellite and in situ Z_SD data going forwards³³, satellite remote sensing still only provides an integrated value that simplifies the biological and environmental vertical and horizontal heterogeneity^34,35; as our results in shelf seas may indicate in the latter case in particular. Consequently, even in the modern era of remote sensing of ocean colour, in situ measurements of Z_SD provide an essential insight into ocean transparency and phytoplankton biomass.

Despite the growth in marine science³⁶, due to the constraints of achieving widespread coverage by the limited scope of professional oceanographic cruises, the participation of the seafaring public provides an opportunity to extend oceanographic data collection. Citizen science has proven to contribute to gaps in our scientific knowledge in many fields^15,37. By collecting ocean transparency data to contribute to historical and current in situ scientific Z_SD efforts, citizen scientists can help better understand the effects of current global-scale, climate-driven changes in the phytoplankton.

In order to assess long-term, global, ocean phytoplankton change in situ from the beginning of the 20th Century, without combining different types of data, it is necessary to increase the coverage of present-day Secchi Disk ocean transparency measurements to enable a direct comparison with historical Secchi Disk data^12,18. Seafarers acting as citizen scientists provide a unique opportunity to collect in situ ocean transparency data. This study addressed the ninth principle of citizen science outlined by the ECSA, which recommends that citizen science should be evaluated for scientific data quality²⁶. We found that when citizen and scientist Z_SD are compared to satellite Z_SD calculated using the algorithm of Morel et al.³² citizen Z_SD slightly underestimates the satellite value and scientist Z_SD deeply overestimates the value (Table 1). The percentage of estimates and the mean deviation suggest that citizens perform as well as scientists. Although citizen data is more variable, it remains reliable and consequently, with this understanding, it is possible to combine citizen-collected ocean transparency data together with scientific studies (Figs. 1d,h,l, 2d,h,l). We would recommend the type of analysis we have conducted when Z_SD data from different studies are combined so that long-term changes in phytoplankton biomass are estimated correctly.

Methods

Scientist measurements of Z_SD

We obtained scientist Z_SD from 3 datasets. (1) The World Ocean Database (WOD) (https://www.nodc.noaa.gov/OC5/WOD/pr_wod.html)³⁸. (2) The World-wide Ocean Optics Database (WOOD) (https://accession.nodc.noaa.gov/0092528)³⁹. (3) The International Council for the Exploration of the Sea (ICES) (https://ocean.ices.dk/Project/SECCHI/)²². We used Z_SD data collected between 1997 to 2005, 1997 to 2001, and 1997 to 1998 from WOD, WOOD and ICES datasets, respectively.

Citizen measurements of Z_SD

We used measurements of Z_SD collected from 2013 to 2019 by seafarers participating in the on-going Secchi Disk study¹⁷.

Satellite estimates of Z_SD

Satellite estimates of Z_SD were the only possibility to compare citizen and scientist measurements of Z_SD since we do not have in situ data collected by citizens and scientists at the same locations contemporaneously. We obtained satellite measurements from the Hermes GlobColour website (http://hermes.acri.fr/index.php). The dataset, obtained by merging data from four sensors (SeaWiFS, MODIS, MERIS, VIIRS) with an arithmetic mean on a 4 km resolution spatial grid, provided overlapping time series of Z_SD for the time period 1997–2019. Estimates of Z_SD were calculated using the algorithm of Morel et al.³².

Bathymetry and distance from the nearest coast

We only used measurements of Z_SD collected in water > 25 m deep and > 1 km from the coast. Bathymetry data were obtained from a global ocean bathymetry chart (0.1 degrees latitude by 0.1 degrees longitude)⁴⁰. We calculated distance from the nearest coast from a global data set of distances from the nearest coastline estimated at a spatial resolution of 0.01 degrees⁴¹. We used the nearest method of interpolation^42,43 to give a bathymetry or distance from the nearest coast for each Z_SD.

Data pre-processing

We attributed a value of satellite-based Z_SD for each in situ measurement of Z_SD if a satellite estimate was available within 100 km of the in situ sample, plus three days before or after it’s measurement. To retain as many measurements as possible near to the coast we calculated the estimation using the nearest method of interpolation^43,44.

Data analyses

We conducted three analyses: (i) indirectly comparing all scientist Z_SD data together and citizen Z_SD data to satellite Z_SD (ii) comparing the three scientific programmes’ data separately and citizen data, to satellite Z_SD and (iii)  determining whether any biases occurred between citizen and scientist Z_SD data when compared to satellite data.

First, we examined the relationship between satellite Z_SD and both scientific (all programmes combined) and citizen Z_SD data using scatter plots for open-ocean and shelf systems; shelf where bathymetry ≤ 200 m and distance to coast > 1 km, and open-ocean where bathymetry > 200 m. We highlighted the datasets by drawing polygons around only-scientist, only-citizen and both scientist/citizen data. We performed three regression analyses³⁷ (i) all data (scientist/citizen), (ii) scientist and, (iii) citizen. Each regression analysis was performed 1000 times, each time removing 20 samples to examine potential outliers’ influence on the regression. Increasing the removal of samples did not alter our conclusions; the variance increased when the removal of samples was higher and conversely. We therefore created 1000 regression lines for each of the three regression analyses. We displayed the histograms of all regression coefficients for scientist, citizen and scientist/citizen data. A Kruskal–Wallis test was applied to test whether regression coefficients for each dataset were significant.

We also calculated the linear correlations between both scientist and citizen Z_SD and satellite Z_SD. We performed 1000 simulations for this analysis, recalculating the correlation after randomly choosing 100 pairs of points when all areas (shelf + open-ocean) were selected, 100 for just shelf data and 90 for open-ocean data. The number of pairs of points available drove the choice of these thresholds. We used histograms to display the correlation coefficients.

Second, we performed similar tests to investigate the three scientific programmes separately (WOD, WOOD and ICES) and citizen data. However, while we used 1000 simulations to calculate linear correlations, we considered 65 pairs of points for open-ocean + shelf data and shelf data, and 29 pairs of points for open-ocean. The ICES and Pacific Ocean data were not used when we tested the relationships between scientist Z_SD data and satellite Z_SD for the open-ocean and shelf, respectively. We also produced histograms of Z_SD for WOD, WOOD, ICES and citizen data for open-ocean + shelf, shelf, and open-ocean to determine if any changes in slopes (regression coefficients) might be explained by the sampling region.

Third, we determined if any biases occurred between citizen and science data when each was compared to satellite measurements independently. Specifically, we investigated whether citizen or scientist in situ measurements of Z_SD underestimated or overestimated satellite measurements of Z_SD that were derived from the algorithm of Morel et al.³²,Z_SD (i.e. without log₁₀ transformation). Although a logarithmic transformation improves the regression analysis by stabilising the variance of the variable (homoscedasticity) (Fig. 1b), it was more straightforward to compare real values; we present the results for a logarithmic transformation in Supplementary Table S1. We assessed three estimators that are generally used to examine the performance of diversity indices⁴⁵: (i) the percentage of overestimates, (ii) the mean deviation MD and (iii) the mean square deviation (MSD).

The percentage of overestimates is a measure of bias varying between 0 and 100%. An unbiased estimator should have a value of 50%, meaning that it underestimates or overestimates 50% of the time. In our context, we use this estimator to determine how frequently in situ citizen or scientist measurements were above or below satellite estimates. For example, a value of 20% for citizen Z_SD would suggest that citizens deeply underestimate Z_SD. The MD (expressed here in m) is the average of the differences between an estimator and its true or expected value. In the context of our study, it is the mean difference between in situ and satellite measurements and it is calculated as follows:

$$MD=\frac{\sum_{i=1}^{n}\left({O}_{i}-{S}_{i}\right)}{n},$$

where O is the in situ Z_SD measurement (either citizen or scientist), S is the satellite measurement of Z_SD and n is the number of couples of points (for example, citizen and satellite estimates of Z_SD). MD can therefore be negative if in situ values are lower than satellite estimates and inversely.

In contrast to the two previous estimators, MSD is used generally to estimate accur⁴⁵. It is calculated by estimating the average of the square of the differences between in situ and satellite measurements as follows:

$$MSD=\frac{\sum_{i=1}^{n}{\left({O}_{i}-{S}_{i}\right)}^{2}}{n}.$$

MSD (expressed here in m²) varies between 0 and $\infty$ and is a measure of closeness between in situ and satellite data. A value close to 0 indicates that the accuracy of the estimator is high and inversely. In our case, it suggests that the values originating from in situ measurement (either citizen or scientist) are constantly close to the values assessed from satellites using the algorithm of Morel et al.³².

We also calculated histograms of the deviations (a deviation is the difference between O and S) between in situ and satellite measurements to compare in greater detail the degree of accuracy between citizens and scientists.

The estimators and histograms, were calculated (i) on the whole set of untransformed data (see Fig. 4a) and data (ii) where Z_SD < 25 m and (iii) where Z_SD ≥ 25 m. They were also calculated between in situ citizen and scientist data. Therefore, we had 6 values of percentages of overestimates, MD and MSD as well as six histograms (displayed as two per figure) at the end of the procedure (see Table 1, Fig. 4).