Introduction

Freshwater algal blooms are usually caused by the excessive proliferation of cyanobacteria. The production of off-flavor compounds and cyanobacterial toxins that often accompany blooms pose a risk to human and animal health. Microcystis, Oscillatoria, and Anabaena are the major bloom forming phytoplankton in Korean freshwaters (Park et al., 1998; Ahn et al., 2003), and they often aggregate to form a floating scum or a greenish film on the water surface. The cell distribution of cyanobacteria in surface scums may vary over time and space. Their spatial distribution becomes more complicated by wind (Hutchinson & Webster, 1994; Verhagen, 1994). Thus, a considerable variance in the measurements of cyanobacterial abundance is unavoidable, even when samples are collected at the same site and time. Furthermore, if a consistent sampling method is not used, there can even be greater variations among measurements. In addition, different subjective decisions as regards selecting the point of sampling can further magnify measurement differences. Without considering such problems, field research on bloom monitoring cannot provide reliable and interchangeable data.

Nonetheless, in most research articles, the sampling method is usually described as “water samples were collected from the lake,” or “collected with a Van Dorn sampler,” or “collected 0.5 m below the water surface,” without any detailed explanation of the sampling procedure. While Codd et al. (1999) suggested fundamental requirements for sampling design and Blomqvist (2001) proposed a standard method for composite sampling, no such standard has been developed for cyanobacterial blooms, which are limited to the epilimnion. Different sampling methods result in different estimations for cyanobacterial abundance, especially in the presence of cyanobacterial scums. By filtering just a larger volume of water sample, Microcystis aeruginosa can be detected in winter, even in the surface water (Latour et al., 2004). Different sampling areas can also influence the interpretation of results of bacterial community structure (Toda & Morikawa, 2003).

In Australia and Korea, concentrations of cyanobacteria cells are used as a guideline of water quality or as an alert system for algal blooms (Falconer, 1994; Ahn et al., 2003). Therefore, a standardized procedure for sample collection is required to guarantee data reliability and minimize possible discrepancies resulting from different sampling practices. The current study compared five widely used methods for sampling cyanobacteria in the field, with the sampling confined to surface waters, where most cyanobacteria occur.

Materials and methods

Comparison of sampling methods

The tested sampling methods, which differed mainly in the degree of scum inclusion and sampling depth, area, and volume (Table 1 and Fig. 1), were selected based on their common use in sampling (Wetzel & Likens, 1991; APHA, 1995): (1) the pumping method sucked up water through a tube at a precise depth, in this case 10 cm, (2) the integrating (or tubular) method collected an intact water column sample from the surface to a depth of 20 cm, (3) the Van Dorn sampler (Vertical Alpha Water Sampler, Wildlife Supply Company, Saginaw, MI) collected subsurface water at a depth of 50 cm, (4) the inflow method streamed surface water into a half-submerged bottle, and (5) the mixing method collected water from the surface to a depth of 30 cm after some stirring around the sampling point. Since cyanobacteria migrate vertically with a diurnal rhythm (Reynolds & Walsby, 1975), sampling was conducted four times in a day at 9:00, 12:00, 15:00, and 18:00 h (GMT + 9:00 h) using the five methods.

Table 1 Comparison of sampling methods used in the present study
Full size table
Fig. 1
figure 1

Illustrations of the five sampling methods. A, pumping; B, integrating; C, Van Dorn; D, inflow; E, mixing

Full size image

Sampling

Two locations were chosen for the present study, Samjeong-dong (36°26′51″ N; 127°28′02″ E) and Chuso-ri (36°20′57″ N; 127°33′40″ E) in Daechung Reservoir (Fig. 2). In the summer of 2004, an Oscillatoria bloom occurred in Samjeong-dong, while Microcystis proliferated in Chuso-ri, with a 14-km linear distance between the two sites. The investigation was conducted on August 25 and September 1 at Samjeong-dong and Chuso-ri, respectively. The wind velocity was under 1 m s−1 on the sampling days. Two different individuals collected the water samples at the same site and time, to test individual differences, when using the same technique.

Fig. 2
figure 2

Sampling sites of Oscillatoria and Microcystis blooms from Daechung Reservoir, Korea

Full size image

Measurement of cyanobacterial biomass

Oscillatoria cells tend to form densely packed filaments and Microcystis cells often form densely packed colonies, making it difficult to count cyanobacterial cells under a microscope. In order to avoid counting errors due to inherent variations between individual counters and to ensure the objectivity of the relative comparisons, the cyanobacterial biomass was measured indirectly using absorbance and in vivo fluorescence. As Oscillatoria and Microcystis cells were observed to occupy over 95% of the total standing crops based on the cell counts for the respective sampling sites, this provided sufficient reason to attribute most of the measured absorbance and in vivo fluorescence to Oscillatoria and Microcystis. The absorbance was measured at 680 nm using a spectrophotometer (UV–160A, Shimadzu, Kyoto, Japan), and in vivo fluorescence was measured using a fluorometer (Turner 450, Barnstead/Thermolyne, Dubuque, IA).

Results

Correlation between absorbance and in vivo fluorescence for cyanobacteria

The distributions of the measured absorbance and in vivo fluorescence were highly skewed in the upper direction for both Oscillatoria and Microcystis. Therefore, all the ensuing ANOVA analyses were performed using log-transformed data to utilize the statistical tools based on a normal distribution (Zar, 1996). Since the absorbance and in vivo fluorescence exhibited a strong positive correlation (Fig. 3), only absorbance data were used for the ensuing statistical analyses. When fluorescence data were used, the statistical outcomes did not differ significantly. There was a stronger relationship between fluorescence and absorbance for Oscillatoria (slope = 0.911, r 2 = 0.729, P < 0.0001) than Microcystis (slope = 0.517, r 2 = 0.370, P < 0.0001). The data points for Microcystis were less scattered and had higher absorbance values than those for Oscillatoria at the same fluorescence.

Fig. 3
figure 3

Correlation between absorbance and in vivo fluorescence for Oscillatoria and Microcystis

Full size image

Comparison of methods and times for Oscillatoria bloom

There were significant differences in absorbance values between the five sampling methods (P < 0.0001) and four sampling times (P < 0.0001) with no interactive effects between sampling methods and times (P = 0.0745). When compared with the other methods, the integrating and inflow methods resulted in significantly higher measurements (Fig. 4A), as they collected more scum during the sampling. The box and whisker plots of the integrating and inflow methods were considerably extended in the upper direction. Conversely, the Van Dorn sampling at a depth of 50 cm produced significantly lower values with the least variation, most likely due to a lower coincidence with surface scums. Among the four sampling times, the absorbance and its variation were both lowest at 18:00 h, which was attributed to the sinking of Oscillatoria to a lower depth in the afternoon (Fig. 4B).

Fig. 4
figure 4

Box and whisker plots according to sampling method (A) and sampling time (B) for Oscillatoria bloom. Each sampling method in panel A includes all the data for the four sampling times, while each sampling time in panel B includes all the data for the five sampling methods. The different letters in the graphs indicate a significant difference at P < 0.05

Full size image

The biomass of Oscillatoria sampled at 9:00 h using the integrating, inflow, and mixing methods was the highest with the largest variations (Fig. 5). In contrast, the peak biomass occurred at 15:00 h when using the pumping and Van Dorn methods. The dispersed floating scum gradually disappeared from the water surface in the afternoon, and by 15:00 h the sinking Oscillatoria cells were sampled to the greatest extent by the two subsurface methods. The floating scum completely disappeared from the water surface at 18:00 h.

Fig. 5
figure 5

Time-series comparison of five sampling methods for Oscillatoria bloom

Full size image

Different results were obtained between each sampling person using the mixing method, even when sampled at the same site and time (P = 0.00745). The inflow method was nearly also affected by the individual collecting the samples (P = 0.0546). None of the other methods showed any significant differences among sampling persons. The integrating method in the early afternoon seemed to be better than other compared methods and times for Oscillatoria blooms, considering the following requirements that should be satisfied, at least, for optimal sampling methods: (1) it is unbiased to any extremes, particularly to outliers, (2) independent on individual samplers, (3) reflects both the surface scum and the underwater colonies, and (4) minimizes the effect of vertical migration of cyanobacteria.

Comparison of methods and times for Microcystis bloom

There were significant differences among the five sampling methods (P < 0.0001) and four times (P < 0.0001) for Microcystis populations at Chuso-ri, Daechung Reservoir, with a significant interactive effect between methods and times (P < 0.0001). The variations in the measured values were smaller than those for Oscillatoria (Figs. 4A, 6A). The results for the inflow and Van Dorn methods were significantly different from those for the other methods. Contrary to the Oscillatoria boom, there was no difference in the measured absorbance between the pumping and integrating methods for the Microcystis bloom, probably due to the less intense surface scum formation. The peak time was also different for Microcystis-dominant population (12:00 h) compared to the Oscillatoria-dominant population (9:00 h), considering all the sampling methods (Figs. 4B, 6B). The discrepancy between the mean and median values was much smaller with Microcystis than with Oscillatoria. In particular, the integrating and inflow methods showed peaks with larger variations at 12:00 h, whereas the subsurface methods, i.e., the pumping and Van Dorn methods, peaked at 15:00 h, as Microcystis migrated downward during the afternoon (Fig. 7). The sinking of Microcystis then led to the minimum sampled biomass at 18:00.

Fig. 6
figure 6

Box and whisker plots according to sampling method (A) and sampling time (B) for Microcystis bloom

Full size image
Fig. 7
figure 7

Time-series comparison of five sampling methods for Microcystis bloom

Full size image

No individual differences were found with any of the sampling methods at a significance level of P < 0.05, for the Microcystis-dominant population. Although the pumping, integrating, and mixing method in the afternoon (12:00–15:00 h) showed similar results for Microcystis blooms, the integrating method between 12:00 and 15:00 h can be recommended as the most proper sampling method for both Oscillatoria and Microcystis blooms, because only this method satisfied the previously mentioned minimal requirements.

Expressing central tendency of bloom measurements

The box and whisker plot showed the distribution of the measured values effectively (Fig. 8), yet, this type of plot is not suitable for data of a small sample size. One could not gain a good understanding of the real data distribution with the mean and standard deviation (SD), particularly for a skewed distribution (Fig. 8A). The median with the first and third quartiles (Q1 and Q3), an abridged version of the box and whisker plot, provided a better grasp of the real data distribution than the mean ± SD, for a skewed distribution. Nonetheless, it still has the same disadvantages as a box and whisker plot. The median absolute deviation (MAD) can be an appropriate alternative. The MAD is the median of the absolute-value distances of the points about the median: MAD = median{|x i - median(x)|}, where x i is each data point.

Fig. 8
figure 8

Comparison of data-presenting methods. Oscillatoria (A) and Microcystis (B) bloom data was obtained using integrating method at 15:00 h

Full size image

One might be concerned that the sample size needs to be large to use the median ± MAD. Contrary to expectation, the median value remained stable when the sample size was decreased from 60 to 3 (Fig. 9). Samples were randomly chosen in the decreasing sample size and the most commonly observed pattern was illustrated in Fig. 9. The coefficient of variation in the median values for the respective sample sizes was 7.0%, which was much smaller than that for the mean values, 14.6%. In addition to the stability, the data distribution could be more effectively estimated by the median ± MAD than by the mean ± SD.

Fig. 9
figure 9

Influence of sample size on data presentation when using median ± MAD and mean ± SD. Oscillatoria bloom data was obtained using all sampling methods at 15:00 h. Data was randomly chosen with decreasing sample size. The most frequently encountered plot is shown above

Full size image

Discussion

Cyanobacterial scums can appear and then disperse in a short period of time, making it difficult to establish a simple-sampling strategy. Accordingly, monitoring approaches need to be flexible and sampling design may need to be adapted to each water body, taking account of hydrological, meteorological, and geographical characteristics of the lake and dominant species in it (Codd et al., 1999). Regular sampling at a fixed site can be supplemented with temporary sampling at mobile sites, where drifting cyanobacteria scums accumulate (Codd et al., 1999; Ishikawa et al., 2003). In this case, one must be careful in deciding where to sample, because each species behaves differently with wind, but it may be sufficient just to have a good spread of routinely sampled sites. For example, Hedger et al. (2004) found that aggregates of Ceratium were pushed upwind, whereas the near-surface scum of Microcystis was accumulated at downwind areas of Lake Esthwaite, UK. Hutchinson and Webster (1994) also observed accumulation of Microcystis at the downwind shore in wind-tunnel tank experiments. Ishikawa et al. (2003), however, reported that the distribution of Microcystis was not affected by wind, although Anabaena formed distinct bloom patch areas driven by wind. These results show that the collected samples could produce significantly different outcomes depending on the prevailing wind conditions. Even after decisions have been made on the most appropriate sampling sites, how the water samples are collected poses another problem. The focus of the current study was limited only to the field collection of cyanobacteria samples.

The five methods compared in this study were aimed mainly at sampling surface waters, as cyanobacterial scums are typically confined to the surface part of the water column in which the public is most likely to come into contact with cyanotoxins. As expected, the sampling methods that incorporated more surface scum resulted in higher measurements with larger variances. Consequently, one important aspect to the sampling method is to determine how far from the surface the sample should be drawn. A deeper sampling depth with the integrating method would dilute cyanobacteria that may accumulate at the water surface. However, the sampling depth was not found to be a critical factor down to a sampling depth of 50 cm in the afternoon (15:00 h), as no significant differences were found among the five sampling methods at that time (P > 0.05 for Oscillatoria and Microcystis populations). The cyanobacteria in Daechung Reservoir have been reported to remain mostly above a depth of 1 m during the daytime in August (Oh et al., 1995). Therefore, sampling depth less than 1 m is recommended for Daechung Reservoir.

Theoretically, as the area sampled and number of replicates increase, the variance decreases with fewer outliers. When the Oscillatoria data was treated as if it had been collected from an area two or four times larger, the extended upper limits in the data distribution became a little shorter. The mean values did not change with the larger sampling areas, yet some median values increased slightly. Even though sampling from a wider area is more advantageous for avoiding sensitivity to outliers and to achieve a greater coverage, it is often difficult due to limited time and costs. In that case, two or more samples of the same volume can be combined to achieve the same effect of broader spatial sampling. When several genera of algae dominate simultaneously and their horizontal distributions are separated by wind effects (Ishikawa et al., 2003; Hedger et al., 2004), this technique could be useful in showing the overall trend.

The arithmetic mean values generally tend to overestimate the actual central tendency and are easily influenced by high outlier values, which are commonly observed in water quality variables (Boyer et al., 2000). The higher similarity of the median values between the compared methods and times makes the median a more attractive substitute for the mean value. Lower variance in the median values, compared with the mean values, has been reported previously (Knowlton et al., 1984). As such, the median was proposed for data presentation in some medical sciences (Govaerts et al., 1998; Krummenauer, 2002). The long-term trend in water quality was also analyzed by using the median and non-parametric statistical methods (Ravichandran, 2003). Although the mean value has been widely used, the present study found the median value to be superior, particularly when the bloom was severe.

The remaining problem with the median value is how to display the error bars simply and effectively, for the graphical presentation. As stated earlier, the real data distribution can be better envisaged with the median ± MAD, than with the mean ± SD (Fig. 8). Although Q1 and Q3 can better express a long upper tail, they have the same disadvantage to that of a box plot, i.e., the requirement of a large sample size. Furthermore, the median ± MAD can be drawn using a smaller sample size without quality deterioration (Fig. 9), and it takes account of the actual differences between all the measurements around the median, while Q1 and Q3 only consider the rank of each measurement. The unique defect of the MAD is that it is unable to sufficiently reflect an extended upper limit. However, this is not peculiar to the MAD, but common to all error bar presentation forms.

Conclusions

This study compared, over four daily time periods, the performance of five widely used methods for sampling surface water when cyanobacterial blooms occurred. The integrating method in the early afternoon satisfied most fully the necessary requisites for standard sampling for both Oscillatoria and Microcystis blooms. In addition, the median ± MAD was proposed as a more practical and ideal method to express a central tendency for cyanobacterial biomass. Even though the mean ± SD has been conventionally used in field studies, it is too sensitive to outliers and quite susceptible to different sampling methods. The use of median ± MAD could overcome such defects.