In situ data in this study were collected until 2009 by the Finnish Institute of Marine Research, and by the Finnish Environment Institute (SYKE) from 2009 onwards, within the Alg@line network of Baltic Sea ferryboxes. Here we consider systems installed on two cargo vessels, M/S Finnpartner (2000–2006) and M/S Finnmaid (2007–2014), which served between Travemünde (Germany) and Helsinki (Finland) as depicted in Fig. . Three routes were sailed during the study period. Depending on weather conditions, the passage between Gotland and the mainland of Sweden (39 % of all transects) was favoured over the direct route east of Gotland (52 %), while the route with a lay-over in Gdansk (Poland) was only occasionally served during 2009 to 2012 (7 %). Several transects (2 %) were sailed for refuelling or maintenance in other ports and not used for this study.

Details on the instrumentation of the Alg@line ferrybox systems can be found in , , , and . In summary, the systems record in vivo fluorescence of chlorophyll a (Chl a), salinity, and temperature throughout the studied period (2000–2014). Turbidity and (in summer) phycocyanin pigment fluorescence were recorded from 2005 onwards and are not used here. At cruising speed (20–23 knots) the sampling interval of 20 s resulted in a nominal spatial resolution of 200 m.

Quality control flags were derived from (1) sensor reading thresholds on speed, flow rate, hull, and sampled water temperature, and (2) data variability, expressed as lower and upper bounds for standard deviation between neighbouring measurements, as described below. Measurements at low (< 5 knots) or zero ship speed are typically collected in the harbour and were omitted. Erroneous records, e.g. caused by instrument communication errors, were removed using a moving window mean filter. A window length of 25 observations (approximately 8.3 min) was used for records of ship speed, and a window length of 100 observations (33.3 min) was used for flow rate and temperature records. Low flow rates can indicate blocked passages, pump failure, or leaks. Flow meter readings were available for approximately one-third of all records. A proxy for flow disruption is the difference in ship-hull temperature and in-line temperature. Flow rates < 0.3 L min-1 or a temperature difference > 2 ∘C were used to flag records as suspect. Instrument failure, communication or digitizing errors may lead to “stuck” values, which were detected by calculating standard deviation in a moving window of 100 samples. Observations corresponding to low standard deviation (σ<1e-4) of Chl a fluorescence measurements or GPS-derived latitude were omitted. GPS-derived latitude was additionally filtered for exceptionally high short-term variability (σ>0.5, window size 50 samples), caused by poor satellite reception or serial communication errors. Table  provides an overview of the applied quality control flags.

Chl a fluorescence data were corrected for sensor drift and discontinuities by transect-wise normalization (division by transect mean). This was necessary to account for changes in instrumentation, signal contamination due to bio-fouling, trapped bubbles and particles, and changes in sensor sensitivity due to deterioration or manual adjustments. Laboratory analysis results of bottle samples are typically available from every sixth transect, with up to 24 samples collected by automated, refrigerated water samplers (Teledyne Isco). Laboratory analyses included inorganic nutrient concentrations (nitrate+nitrite, phosphate and silicate), Chl a concentration, and occasionally inverted light microscopy counts of phytoplankton species. Laboratory Chl a concentration results were used to convert transect-normalized Chl a fluorescence to units of Chl a concentration (in mg m-3). First, a linear (generalized least squares) regression fit of normalized Chl a fluorescence against corresponding Chl a lab measurements was carried out for each sampled transect. If the regression failed (R2<0.3 or p>1), a moving window regression was carried out (window length 10 samples), and the subset with the highest R2 was used to determine the correction factor. The threshold for R2 was determined manually based on the distribution of R2, while p>1 indicated numerical instabilities during the fitting procedure. Each transect without corresponding bottle samples was corrected by individually applying the regression parameters of the two neighbouring sampled transects. These two solutions were then interpolated linearly, weighted by their temporal distance to the respective transect. Negative concentration values occasionally occurred for weak fluorescence signals, and were set to zero.

The diurnal variability of the fluorescence signal was estimated from quality-controlled observations in all seasons. First, these observations were divided by their respective transect mean to remove biomass-driven first-order variability in the fluorescence signal. Then, diurnal cycles were derived by dividing these observations into hourly bins and sun elevation angle ranges (0.1 rad bins).

Photosynthetically active radiation (par), sea surface temperature (sst), and wind speed (wind) were derived from the ECMWF (European Centre for Medium-Range Weather Forecasts) ERA-Interim reanalysis data set . The spatial resolution of the model is constrained by the underlying atmospheric model, which is stored on a spatial T255 grid corresponding to approximately 79 km cell size when projected to a reduced Gaussian grid. Four values per day were retrieved for each parameter and the entire Baltic Sea. Parameter values for each Alg@line observation were extracted using spatio-temporal spline interpolation of third order. The first-order seasonal signal (e.g. rising par and sst in spring) was removed from the observations by subtracting multi-year (2000–2014) daily sea area averages, approximated by second-order polynomials. The seasonally detrended parameters were then averaged over the bloom period and are further referred to as par, sst, and wind.

A single term for nutrient availability was adopted from , calculated as nut = (NO3+NO2)×PO4×SiO43, where NO3+NO2, PO4, and SiO4 are the concentrations of nitrite+nitrate, phosphate, and silicate, respectively. These concentrations were derived from laboratory analysis of bottle samples that were regularly collected along the transect (further detail in Sect. ). nut was spatially binned for each investigated sea area and resampled to daily averages and consecutively smoothed with a 21-day centred-running-mean filter. This treatment resembles the processing of Alg@aline observations (see Sect. ) to enable consistent interpretation of the joint data set. Nutrient concentrations and depletion timing are described using the following metrics. The nutrient concentration prior to bloom start (nut-peakvalue) was defined as the yearly maximum nutrient concentration (day of year between 31 and 160). The day of year when the nutrient concentrations equalled 100, 50, and 25 % of their peak values are referred to as nut-peakday, nut-deplday-50, and nut-deplday-25. The day and value of the lowest nutrient concentration index are referred to as nut-minday and nut-minvalue. The rate of nutrient depletion between 75 and 25 % of the peak value (nut-slope) was determined through linear regression.

Extraction of bloom timing and intensity was carried out for five Baltic Sea areas, where each area follows definitions of the HELCOM Combine program . Figure  illustrates the location of the areas: the western Gulf of Finland (gof: > 59.5 ∘N latitude, along-transect), the northern Baltic Proper (nbp: 58.4–59.5 ∘N latitude, along-transect), the combined western and eastern Gotland basins (got: 56.2-58.4 ∘N latitude, along-transect), the southern Baltic Proper (sbp: 54.5–56.2 ∘N latitude, along-transect), and the Bay of Mecklenburg (bom: < 54.5 ∘N latitude, along-transect). For the got and sbp areas, only routes that passed by Gotland were selected, whereas routes via Gdansk were excluded. This is because the route through Gdansk was sailed only from 2009 to 2012. If not otherwise stated, all further steps are carried out individually for each of these areas and for day of year between 31 (31 January) and 160 (9 June). The ship-of-opportunity (Alg@line) measurements typically commenced in the second half of January, which is why 31 January was chosen as the start of our analysis. The end date was chosen such that it covers all spring bloom events in all basins but excludes summer bloom.

Alg@line Chl a concentrations (see Sect. ) were resampled to daily sea area averages, using linear interpolation, and subsequently smoothed with a 21-day centred-running-mean filter e.g. to fill in gaps and reduce short-term variability. We derive several metrics, all of which have in common that the bloom peak concentration (peakheight, see Table  for explanations of acronyms) and timing (peakday) are defined as the maximum Chl a value at the corresponding day-of-year, respectively. Two threshold-based metrics and one distribution-fit-based metric were calculated.

Chl a concentration exceeding a fixed threshold of 5 mg m-3 was defined as bloom by , further referred to as const5. A 21-day centred-running-mean filter was used to keep results comparable to the other metrics considered, whereas used a 7-day centred-running-median filter.

Principal component analysis (PCA) was carried out to attribute seasonally detrended meteorological conditions (sst, par, wind) and nutrient concentrations (nut-peakvalue, nut-minvalue) to the inter-annual variability in bloom intensity (bloomidx, concavg, peakheight) and timing (startday and peakday, duration). Outliers were defined for each parameter as departure by more than 3 standard deviations from the parameter mean, and replaced with the region median. Z score normalization (subtraction of mean, division by standard deviation) was carried out on a per-region basis. Region-equalized, zero-mean, and unit-variance data were then subjected to the PCA function in the Python framework scikit-learn .

The Alg@line ferrybox systems collected over 9.5×106 observations between 2000 and 2014, of which 3.8×106 observations were sampled during spring (day-of-year 31 to 160). Availability and rejection rates for each quality control parameter are listed in Table . In total, quality control procedures removed 4.55 % of all observations.

Determination of the fluorescence yield was supported by an “adaptive regression” method. Where necessary (R2<0.3 or p>1), it selected the subset of bottle-sampled and laboratory-analysed Chl a concentrations that yielded the best linear fit to Chl a fluorescence observations for a given transect. This procedure allowed 318 (98 %) out of 324 transects for which bottle samples were collected to be successfully fit. Only 266 (82 %) transects could have been used (R2>=0.3 and p≪1) without applying this technique.

Figure a shows normalized fluorescence observations as a function of sampling time-of-day. Results are presented separately for summer (May to August), winter (November to February), and the transition periods (autumn, spring). Diurnal variability was most pronounced in summer, when the fluorescence signal varied on average 50 % over the course of a day. In winter and during the transition periods (spring, autumn) a diurnal variability of 35 and 38 %, respectively, was contained in the fluorescence signals. This seasonal effect is likely caused by variations in average irradiance intensity, which are modulated primarily by sun elevation, but also by atmospheric conditions (e.g. cloud cover, aerosol optical thickness) and optical properties of the water body (e.g. ice cover, attenuation). Figure b depicts normalized fluorescence as a function of solar elevation. In this representation, seasonal differences in diurnal variability are essentially absent and the correspondence between solar elevation and average fluorescence response was approximately linear for daytime observations.

Blooms generally developed first in the south and progressed towards the north (see Fig.  and Table ). Bloom peak timing (not influenced by choice of metric) followed this pattern, as did metric-dependent bloom start and end dates. The fixed-threshold bloom metric const5 suggested longer blooms in high-biomass sea areas like the gof, compared to low-biomass areas such as the sbs. The spatially variable-threshold metric median5 applies area-specific bloom thresholds (nbp: 3.52, gof: 4.95, got: 2.51, sbs: 2.62, bom: 4.02 mg m-3) and resulted in approximately stable bloom duration in all sea areas. The weibull metric, which is not sensitive to absolute bloom intensity, also resulted in comparable bloom durations for all sea areas. The year-to-year variability of start, peak, and end days generally increased towards the south for all metrics.

Spring bloom intensity was described by three parameters: the metric-independent bloom peak concentration (peakheight), the Chl a concentration average during bloom conditions (concavg), and the sum of daily Chl a concentrations over the bloom period (bloomidx). Similar patterns were observed for all these parameters and bloom metrics, as illustrated in Fig. . The highest bloom intensity was found in the gof and nbp, followed by the bom. Low-intensity blooms were observed in the sbp and the got. Variability was generally proportional to bloom intensity, highest in the high-biomass and coastal gof and bom. Variability in bloomidx was comparable to that in peakheight, while concavg was considerably more stable. All calculated bloom phenology parameters can be found in the Supplement.

Figure  shows normalized (subtraction of area-average concentration) concavg and peakheight for all sea areas combined, as a function of bloom year. peakheight is independent of bloom metric and shows a highly significant (R2=0.12, p≪0.01) negative trend of -0.30±0.10 mg m-3 yr-1. concavg is dependent on bloom start and end days and was therefore calculated for all applied metrics. Statistically significant, negative trends resulted from all metrics: -0.12±0.04 for const5 (R2=0.11, p≪0.01), -0.11±0.05 for median5 (R2=0.12, p<0.05), and -0.22±0.07 mg m-3 yr-1 for weibull (R2=0.11, p≪0.01).

No significant trends were found for bloomidx, startday, and peakday with any of the applied metrics, while endday showed weakly correlated but statistically significant (R2=0.06, 0.08, p<0.05) positive trends for const5 and weibull with slopes 0.6 to 0.7±0.3 day yr-1, respectively.

Bloom duration resulting from the weibull metric stands out in the result set with a positive trend of 1.04±0.20 day yr-1 (R2=0.28, p≪0.01, Fig. ). No significant trend in bloom duration was found for any fixed- or variable-threshold metric.

Peak nutrient concentrations showed no significant trend, in contrast to post-bloom nutrient concentrations with a highly significant, negative trend -0.020±0.004µmol L-1 yr-1 (R2=0.23, p≪0.01). Peak nutrient concentration timing shifted to earlier dates (-0.7±0.3 day yr-1, R2=0.06, p<0.05), while the 25 % of peak value was reached progressively later (0.67±0.31 day yr-1, R2=0.06, p<0.05). No significant trends were found for the nutrient depletion slope, 50 % of peak value timing, or the day of minimal nutrient concentrations.

Pre-bloom nutrient concentrations were positively correlated to bloom peak height (no normalization, R2=0.39, p≪0.01) and concentration average (no normalization, R2= 0.37–0.57, p≪0.01, depending on metric). After applying area-wise mean and variance (z score) normalization, however, a negative correlation was found for peakheight (R2=0.11, p≪0.01, metric-independent) and concavg (R2=0.12, 0.11, p≪0.01 for const5 and weibull, respectively).

The timing of nutrient depletion, specifically nut-deplday-50, was positively correlated to the bloom peak day (R2=0.47, p≪0.01), and to bloom-averaged, detrended par-levels (R2= 0.14–0.29, p≪0.01). Average wind speed and par were negatively correlated during bloom conditions (R2= 0.10–0.23, p≪0.01). The bloom timing parameters (startday, peakday, endday) were weakly but statistically significantly intercorrelated (results not shown).

PCA scores and loadings of the first three principal components (PCs) are shown as biplots in Fig. . The first PC is dominated by negative correlations to bloom intensity parameters (peakheight, concavg, bloomidx). This component is positively correlated to pre-bloom nutrient concentration (nut-peakvalue) and bloom duration, illustrating that bloom intensity is driven by pre-bloom nutrient availability. The second PC is linked to bloom timing, with strong positive correlations to startday and peakday. Correlations to par (positive), sst (positive), and wind (negative) suggest that weather conditions affect bloom timing. Bloom duration is positively correlated to the third PC, as well as to bloomidx. Additional negative correlations to nut-minvalue and wind, as well as a positive correlation to par, suggest a link between favourable meteorological conditions (low wind-mixing, high light level) and efficient nutrient depletion.

Thresholds for speed, flow rate, and data variability were iteratively adjusted to the data set and may not be applicable to other ferrybox implementations. Particularly flow rate, derived from differences in line and hull temperature, will likely require tuning to each ferrybox installation. However, here we analysed data from two ferrybox installations, which could be treated with the same set of thresholds. Transect-wise normalization of the quality-controlled fluorescence data was adequate to consistently interpret observations collected by different generations of instrumentation. However, this approach crucially depends on continuous temporal coverage of reference measurements for calibration to Chl a concentrations. Adaptive regression analysis improved the handling of statistical outliers which would otherwise hamper determination of fluorescence yield, while transects for which no bottle samples are available were corrected with an interpolated fluorescence yield derived from the closest bottle-sampled transects. The present procedure allows for automated and reproducible processing, which is an improvement over manual quality control. Applying the proposed interpolated fluorescence yield helps in reprocessing and long-term data analysis of ferrybox fluorescence observations to better represent natural variability.

Diurnal fluorescence patterns showed low seasonal dependence after accounting for solar elevation. Unsurprisingly, light intensity is the predominant factor in Baltic Sea phytoplankton fluorescence yield variability. Other seasonal differences in fluorescence response can be attributed to typically higher cloud cover in winter compared to summer and spring/autumn, which was not accounted for in our analysis. The seasonal cycle of species composition, from dinoflagelate- and diatom-dominated spring communities to cyanobacterial summer bloom , influenced fluorescence yield considerably less than diel cycles.

The diurnal variability in fluorescence response of 50 % during an average summer day is within the range of earlier findings, e.g. 66 % (±33 %) for near-surface observations in upwelled waters of the equatorial Pacific reported by or 30 % for near-surface seaglider observations in northeast Pacific waters off the Washington coast, United States ; although differences in normalization impede direct comparison. The sampling depth of 5 m for Alg@line systems and the high attenuation of the Baltic Sea in comparison to clear Pacific Ocean waters are likely to dampen the observed diurnal variability.

In this study, fluorescence observations during spring, when diurnal variability reached on average 38 %, were binned for five large Baltic Sea areas. At a typical cruising speed of approximately 23 knots, each sea area is sampled for at least several hours. This limits the influence of diurnal variability in fluorescence yield along a transect on derived Chl a concentration, which is therefore of lesser relevance for the present study. However, if fluorescence measurements were to be quantitatively evaluated at a higher spatial resolution, locally varying fluorescence yield should be accounted for. Analysis of signal coherence offers an alternative to quantitative interpretation of fluorescence observations and can be used to qualitatively detect cyanobacterial surface bloom. If light history is known, e.g. from a dedicated irradiance sensor, a correction of diurnal fluorescence yield variability might be possible, and further research in this direction is recommended.

The presented bloom phenology expands the time series presented by and is in good agreement for the overlapping period (2000–2004) when comparing the const5 metric results. Remaining differences are likely due to quality-control and preprocessing procedures on the fluorescence records. The authors reported for gof, nbp, and the Arkona Sea that bloom typically started in the south and ended in the north, while bloom intensity increased towards the north. These observations are confirmed here. Sea areas not covered in , e.g the high-biomass bom and low-biomass sbp and got, followed the reported south–north trend in bloom development. Present results also support and expand the findings of , who showed, with simulations and monitoring data from 1994–1996 for the western Baltic Sea, that surface heating in early spring needs to overcome the temperature of maximum density to repress convective mixing and allow spring bloom to emerge. The temperature of maximum density increases with decreasing salinity; therefore convective mixing is sustained longer in less saline northern Baltic Sea waters when spring temperature is on the rise. At the same time, incident solar radiation increases slower in the north due to lower solar elevation.

Inter-annual variability in coastal systems exceeds long-term trends by orders of magnitude . Consequently, trends were observed at relatively low coefficients of correlation. The importance of appropriate data preprocessing and gap handling e.g and choice of metric has been demonstrated in literature and is further emphasized by the present analysis. Robustness of the reported decadal trends is documented by high statistical significance levels (p≪0.01, Figs.  and ), which were supported by spatially binning phenology parameters from all examined Baltic Sea areas. Similar trends were observed earlier for individual Baltic Sea areas, however, usually outside 95 % confidence intervals e.g.

reported stable or increasing Chl a concentrations for the period 2007–2011 in several Baltic Sea areas despite signs of declining nutrient concentrations. More recently, eutrophication trend reversal and oligotrophication processes were reported by , based on analysis of 112 years of consolidated Baltic Sea observations. Both reports considered surface-layer Chl a concentration in summer as one of the direct indicators of eutrophication, but did not include spring bloom in their assessment. The time series for 2000–2014 that we present here fills this gap: a negative trend in bloom intensity was also found for spring bloom, providing further evidence for their hypothesis of gradual nutrient load reduction.

Thresholds of const5 and median5 are fixed for the whole time series. The observed negative trend in peak concentration was expected to introduce an artificial negative trend in bloom duration because an increasingly higher percentile of the distribution is seen below the bloom threshold (Fig. ). Contrary to this expected behaviour, however, const5 and median5 revealed no significant trends in bloom duration. This indicates that the anticipated negative trend in bloom duration was countered by a positive trend, e.g. in bloom intensity. The Weibull metric is based on concentration distribution ratios that are calculated individually for each bloom. Therefore, Weibull-metric results for bloom duration are not sensitive to long-term trends in peak concentration. Weibull-distribution metrics confirmed a highly significant, positive trend in bloom duration. These two sets of results corroborate the conclusion that spring blooms in the Baltic Sea have become longer, while Chl a peak and average concentration levels have declined.

This “flattening” of the concentration distribution is supported by the absence of a trend in time-integrated biomass bloomidx and by shifts in nutrient concentration timing (earlier nutrient peak concentration, later 25 % of peak value day). These results indicate that annually generated spring bloom biomass has not changed significantly over the study period, in contrast to bloom timing. found a similar development for cyanobacterial summer surface bloom, and reported decadal oscillations, yet no long-term trend, of surface area covered by cyanobacteria in the period 1979–2013. In the same period, summer bloom initiation moved to earlier dates by -0.6 day yr-1. These results suggest that the gap has decreased between dinoflagelate- and diatom-dominated spring bloom and cyanobacterial summer bloom. Due to the shorter period covered here as compared to the time series presented by , it cannot be ruled out that the spring bloom trends are caused by decadal oscillation. Moreover, Alg@line nutrient records often did not commence sufficiently early in the season to record bloom onset. Trends in bloom start and nutrient peak timing can therefore not be derived at the same accuracy and precision as the other phenological parameters. In future, additional data and longer time series may revise this analysis. To this end, nutrient metrics derived in this work are provided in the Appendix.

Our findings emphasize that bloom timing is an essential indicator to monitor marine ecosystem dynamics, and thus eutrophication status. Observations at high temporal resolution and choice of bloom metrics are crucial to derive bloom timing trends. Eutrophication status assessment frameworks such as HEAT3.0 may be adapted to embrace available high-frequency data sources to include bloom timing in their analysis. The present results may also prove useful in the calibration and validation of ecosystem models of the Baltic Sea.

Gradually decreasing nutrient concentrations , as well as rising average air and sea-surface temperatures have been reported for recent years, corresponding to a combination of nutrient reduction efforts and global climate change. Several scenarios for future change are plausible but extrapolation of the present results to climate scenarios is beyond the scope of this study. We nevertheless make an attempt to attribute the observed bloom phenology shifts to reported changes in environmental drivers.

Wintertime nutrient concentration and bloom intensity were positively correlated if no spatial normalization was applied. This supports the paradigm that the first-order driver of bloom intensity is nutrient availability. Lacking alternative explanations, we attribute the reported negative trend in bloom peak concentration to declining nutrient concentrations. First-order spatial trends in bloom intensity and timing can be removed by an area-wise z score normalization, which effectively constrains the analysis to inter-annual variability. After this normalization, both regression and PCA resulted in negative correlation between wintertime nutrient concentration and bloom intensity. This negative feedback can be understood as a subtle interaction between meteorological forcing and nutrient supply: strong wind-forced mixing can cause upwelling of deep, nutrient-rich waters to surface layers. Wind speed, however, was found to be negatively correlated to the prevalent light level, as well as to bloom duration and bloom index. Therefore, in years when additional nutrients are available due to strong wind forced mixing, low-light regimes that can slow down bloom development are also likely to prevail.

Bloom duration primarily co-varied with weather conditions; e.g. high irradiance levels and low wind speeds were frequently observed for long-lasting blooms (and vice versa). Although the same pattern was observed for bloom timing, no trend was found for bloom start and peak day. Increasingly favourable meteorological conditions in late bloom phases are thus a likely driver for the observed increase in bloom duration. Similar weather-driven modulations of bloom timing were reported earlier for spring, and especially cyanobacterial summer bloom .