For millennia, the marine environment has been subjected to various sources of pollution. Major inputs of nitrate, phosphate, and other pollutants have been causing detrimental effects on the marine environment, including harmful algal blooms (HABs) and eutrophication (Le Moal et al., 2019). Since phytoplankton are at the base of the food web, their blooms can affect higher trophic levels and cause serious changes in marine biodiversity and water quality (e.g., oxygen deficiency) (Kazmi et al., 2022; Young et al., 2020). HABs can produce toxins that degrade water quality and cause health problems in humans and marine animals (Ross Brown et al., 2022; Young et al., 2020). They can also be associated with mass foam accumulations on beaches, as with Phaeocystis globosa blooms (Blauw et al., 2010; Spilmont et al., 2009). Furthermore, they can detrimentally cause economic losses in sectors such as fish farms, shellfish aquaculture, tourism, and recreational activities, as well as public health (Derot et al., 2020; Hallegraeff et al., 2021).

Understanding the processes underlying HABs and eutrophication necessitates continuous monitoring of the marine environment in order to prevent the associated deterioration effects and achieve optimized environmental assessment and management strategies. Traditionally, monitoring aquatic and marine ecosystems was done using low-frequency in situ measurements (weekly to monthly sampling frequency). It was performed by collecting water samples through Niskin bottles and then performing several laboratory analyses to determine various physical, chemical, and biological parameters, including salinity, temperature, conductivity, and organic and inorganic matter, as well as phytoplankton biomass, abundance, and diversity. Despite the fact that these data helped scientists to have an overview of the processes taking place in the marine environment, they are of insufficient temporal resolution to advance the understanding of phytoplankton dynamics and eutrophication.

In order to implement proper management strategies that prevent further deterioration of the marine ecosystem, it is crucial to enhance our understanding of algal blooms, eutrophication; recurrent, rare, and extreme events; and phytoplankton dynamics. Thus, it is necessary to collect continuous measurements not only on a monthly or weekly basis but also on an hourly or even sub-hourly timescale. Satellites and models can provide data of high spatio-temporal resolution (Chai et al., 2020), but such data must be validated with in situ data (Lefebvre and Schmitt, 2016). This motivated scientists to study the marine environment using in situ high-frequency (high temporal resolution) monitoring systems such as buoys, FerryBox systems, etc. (Dickey and Bidigare, 2005).

Over the past decades, the advancement of sensor technology and data science has shed light on the importance of time series in marine research. This urged the construction of autonomous systems capable of supporting long-term time series for key physical, chemical, and biological parameters. The implementation of such automated systems enabled the measurement of essential ocean variables (EOVs) and essential biodiversity variables (EBVs) at high frequency, which aided in reorienting marine research from low-frequency measurements to high-frequency measurements (Blain et al., 2004).

In the eastern English Channel (EEC), HABs are mainly caused by the Prymnesiophyceae Phaeocystis globosa, which is often associated with Pseudo-nitzschia (Karasiewicz and Lefebvre, 2022). When the temperature of the water rises in the spring and summer and when nutrient concentration is optimal, Phaeocystis globosa forms a large biomass. Phaeocystis globosa was identified as a potentially harmful species for several reasons. First, it releases dimethyl sulfide (DMS) gas, which can irritate people's eyes, skin, and respiratory system (Riegman and Van Boekel, 1996). Second, mucopolysaccharides are abundant in its colonies (Zhu et al., 2021). These polysaccharides are broken up by external factors like turbulence as well as internal factors like lysis and aging, which cause the accumulation of a thick, odorous foam on the coast. Besides, needle-shaped Pseudo-nitzschia can stick into Phaeocystis globosa colonies and form structures that irritate filter feeders during Phaeocystis globosa blooms (Sazhin et al., 2007). These structures can also injure fish, making them more susceptible to bacterial and viral infections (Lefebvre and Devreker, 2023). Moreover, the neurotoxin domoic acid (DA) produced by Pseudo-nitzschia is responsible for the neurological disorder known as amnesic shellfish poisoning (ASP) in humans (Bates et al., 2018; Petroff et al., 2021). Additionally, marine mammals and seabirds may get poisoned if they consume DA-contaminated planktivorous prey (Delegrange et al., 2018).

The French monitoring of phytoplankton population and associated environmental factors in the eastern English Channel (ECC) started in 1979 with RNO (Réseau National d'Observation) or RNC (Réseau Nationale de Contrôle). Then, in 1984, a national network called REPHY (le REseau de surveillance du PHYtoplankton et des phycotoxines) was established by Ifremer to estimate the abundance and taxonomic composition of phytoplankton, describe their spatio-temporal dynamics, detect toxin-producing species, and monitor and alert for harmful blooms (10.17882/47248, REPHY – French Observation and Monitoring program for Phytoplankton and Hydrology in coastal waters, 2022). After that, in 1992, the Artois-Picardie Water Agency and Ifremer decided to establish SRN (Suivi Régional des Nutriments) to accurately monitor nutrient concentrations (Lefebvre and Devreker, 2023). Although these monitoring networks enhanced our knowledge of phytoplankton dynamics, they remain inadequate to thoroughly understand recurrent, rare, and extreme events occurring in the marine environment.

In 2004, the MAREL (Mesures Automatisées en Réseau pour l'Environnement Littoral) Carnot monitoring station was installed in the French part of the ECC. The MAREL Carnot station, developed and implemented by Ifremer (French Research Institute for Sea Exploitation), is a moored buoy protected by a tube and equipped with physical, chemical, and biological measuring devices and sensors that operate continuously and autonomously. This multi-sensor station is located in the Boulogne-sur-Mer harbor (eastern English Channel), which is influenced by both marine water and freshwater. It is equipped with high-performance systems for seawater analysis and data transmission in near real-time. It measures the following parameters with high-frequency resolution (20 min): estimated sea level, wind direction relative to true north, horizontal wind speed, photosynthetically active radiation (PAR), seawater temperature, practical salinity, pH, dissolved oxygen, oxygen saturation, fluorescence, and turbidity. For nutrients, including nitrate, phosphate, and silicate, the sampling frequency is set to 12 h.

The purpose of this article is to describe the MAREL Carnot dataset and provide an overview of the variability of its physical, chemical, and biological parameters. For future users of the related dataset, we will offer a thorough description of the MAREL Carnot station, including its deployment and measurements. Based on previous research papers, we aim to demonstrate that the MAREL Carnot dataset can be used to evaluate the environmental or ecological status and conduct research in marine phytoplankton ecology and oceanography.

Table 4 represents the descriptive statistics for the main parameters measured by MAREL Carnot from 2004 until 2022. The results show a high percentage of missing data, denoted as NA or “not available”. A major problem in time series is missing data. It is primarily due to sensor failure, communication problems, or sensor maintenance disability.

Figures 4 and 5 show the time series of parameters collected from MAREL Carnot station from 2004 until 2022. We noticed that some signals have seasonal cycles, such as water and air temperatures as well as photosynthetically active radiation (PAR). In addition, the signals contain episodic or continuous missing values over several time periods. For instance, a large number of missing values can be found around the year 2014 in most time series. This is due to station and sensor alterations that occurred during that time, particularly the replacement of several sensors with a multi-parameter probe (Lefebvre and Schmitt, 2016). Likewise, the signals of air temperature, PAR, wind speed, and sea level were lost for several years while waiting for new funding resources to ensure the renewal of sensors and associated electronic systems. Conductivity data prior to 2015 were deleted by the Coriolis data center, probably under the presumption that salinity is more relevant to the scientific community. This highlights the added value of our research, which is to ensure that all observations collected by MAREL Carnot remain permanently available and accessible to everyone.

Nutrient signals such as phosphate, nitrate, and silicate are only available until 2010. This is caused by a previous sensor failure and the inability to replace it.

Figure 6 shows a wind rose showing the frequency (%) and wind speed (m s-1) for different wind directions measured by MAREL Carnot from 2004 until 2022.

Scientists from several disciplinary backgrounds have utilized MAREL Carnot data to accomplish a wide range of research objectives. In the following paragraphs, we will go over some of the most significant findings from several research efforts. The scientific community that is interested in the MAREL Carnot dataset may find this evaluation useful in determining which topics may or may not require further study based on the results of this evaluation. This dataset allows researchers to investigate the dynamics of phytoplankton as well as to detect blooms caused by human activities and/or climate change.

For instance, Rousseeuw et al. (2015) developed an unsupervised hidden Markov model (uHMM) for monitoring the marine environment, specifically for detecting algal blooms and understanding phytoplankton dynamics. In their uHMM, parameters were estimated using spectral clustering rather than the commonly used iterative expectation maximization. The results obtained using the MAREL Carnot dataset showed that the proposed system is efficient at detecting the main productive and non-productive periods, as used for the purposes of the EU Water Framework Directive to assess good environmental status, and refining knowledge about phytoplankton bloom dynamics in a temperate ecosystem, temporarily dominated by a harmful algae, e.g., Phaeocystis globosa. Thus, the suggested uHMM system successfully characterizes phytoplankton dynamics from new incoming data (in near real-time), and it will enable researchers to gain a better understanding of the main controlling or forcing parameters (e.g., nutrient pressure, light availability, turbidity), the environmental status (e.g., phytoplankton biomass), and the direct and/or indirect effects of algal blooms (e.g., oxygen concentration) (Rousseeuw et al., 2015).

Following the unsupervised approach by Rousseeuw et al. (2015), Grassi et al. (2019) suggested a multilevel spectral clustering (M-SC) to split multivariate time series from general patterns to extreme events without a priori knowledge. The results obtained from the MAREL Carnot dataset have shown that we can extract knowledge on dynamics of events or environmental states. In addition, it was shown that M-SC allows for unsupervised labeling of time series, which is a basic part of machine learning needed to build an event prediction system and improve sampling strategies to operate in near real-time (Grassi et al., 2019). As a result, scientists should be able to create a HAB early-warning expert system to warn shellfish farmers and prevent both public health risks and commercial losses in the shellfish farming business.

The application of M-SC and uHMM on the MAREL Carnot dataset can reveal rare, recurrent, and extreme events, which may aid in improving coastal assessment and defining what constitutes a desirable environmental state. This can indirectly help improve management strategies established by the Water Framework Directive (WFD), Marine Strategy Framework Directive (MSFD) and Oslo and Paris (OSPAR) Convention.

The MAREL Carnot dataset can also be beneficial to data scientists and machine learning specialists. This dataset contains some missing data due to sensor failure and harsh weather conditions that prevent immediate sensor maintenance. It was used to evaluate the performance of a proposed “dynamic time warping” method to fill in successive missing values of univariate time series (Phan et al., 2020) and low, uncorrelated multivariate time series (Phan et al., 2017). It was also utilized in the application of a fuzz-logic-based similarity measure to impute large gaps of uncorrelated multivariate time series (Phan et al., 2018). These data imputation approaches are published on the Comprehensive R Archive Network (CRAN) and accessible through DTWBI (Imputation of Time Series Based on Dynamic Time Warping) and DTWUMI (Imputation of Multivariate Time Series Based on Dynamic Time Warping) packages, respectively.

This dataset can also be utilized to assess the performance of time series analysis methods on marine datasets. For instance, Kbaier Ben Ismail et al. (2016) used four parameters measured by MAREL Carnot to compare the classical techniques of time series analysis to recent ones. Also, Huang and Schmitt (2014) performed empirical mode decomposition (EMD) to study time-dependent intrinsic correlation of temperature and dissolved oxygen time series measured by MAREL Carnot.

Derot et al. (2020) investigated the impact of different sampling frequencies on forecasting harmful algal blooms. They applied a random forest (RF) and sliding-window strategy on 12 parameters derived from the MAREL Carnot dataset. The research demonstrated that the sampling frequency has a direct impact on the forecast performance of a RF model as high-frequency datasets might provide useful information to the RF model. This type of model sets the groundwork for the creation of a numerical decision-making tool that could help mitigate the impact of algal blooms and can recreate interactions that closely resemble the real biological processes (Derot et al., 2020).

Moreover, the MAREL Carnot dataset might be useful for studying turbulence. Derot et al. (2015) studied the phytoplankton biomass during bloom events by applying empirical mode decomposition (EMD) on a fluorescence dataset from MAREL Carnot. Results revealed that bloom events include considerable internal variations. Blooms are not smooth and “mountain-like” but exhibit high-frequency oscillations possibly due to turbulent advection and complex population dynamics (Derot et al., 2015). Besides, Zongo and Schmitt (2011) demonstrated that pH fluctuations in marine waters are strongly influenced by turbulent hydrodynamical transport and may be considered a turbulent active scalar.

Moreover, the sensors placed on the lighthouse provide valuable data for meteorological research and may improve local weather forecasts by measuring variables including wind speed, wind direction, and air temperature. Also, the MAREL Carnot high-frequency dataset can be used to validate satellite-derived products such as fluorescence. It also provides measurements for parameters that cannot be measured from space such as nutrient concentration (Lefebvre and Schmitt, 2016). Our dataset may assist fisheries research. For instance, Toomey et al. (2023) incorporated MAREL Carnot water temperature time series in the supplementary material of their study on the impact of temperature on Downs herring.

Overall, the MAREL Carnot station provides automatic, continuous, and long-term observation of various physical, chemical, and biological parameters that enhance our knowledge about the environmental state of the coastal environment and bloom events. Hence, the MAREL Carnot dataset aligns with objectives of SRN (Suivi Régional des Nutriments in French, Regional Nutrients Monitoring Program), especially by assessing the influence of continental inputs on the marine environment and their implication on possible eutrophication, which can assist in estimating the effectiveness of development and management policies in the marine coastal zone (Lefebvre and Devreker, 2023). To clarify, MAREL Carnot is the first coastal sampling station for the SRN transect. Thus, it assists in understanding phytoplankton dynamics by determining recurrent, extreme, and rare events in this highly impacted and vulnerable coastal area.

Furthermore, the MAREL Carnot dataset can be complementary to both REPHY (Observation and Surveillance Network for Phytoplankton and Hydrology in coastal waters) (10.17882/47248, REPHY – French Observation and Monitoring program for Phytoplankton and Hydrology in coastal waters, 2022) and REPHYTOX (Monitoring Network for Phycotoxins in marine organisms) (10.17882/47251, REPHYTOX – French Monitoring program for Phycotoxins in marine organisms, 2022). The goal of REPHY is to measure the biomass, abundance, and composition of marine phytoplankton as well as hydrological parameters in coastal and lagoon waters. REPHYTOX is designed to find and track three types of toxins that can build up in bivalve mollusks and cause DSP (diarrheic shellfish poisoning), PSP (paralytic shellfish poisoning), and ASP (amnesic shellfish poisoning) (Belin et al., 2021). Monitoring carried out by MAREL Carnot in parallel with REPHY and REPHYTOX permits continuous adaptation to the objectives, developing analysis strategies with extensive and complex data, thereby ensuring sustainability, which were challenges faced by REPHY and REPHYTOX before.

While MAREL Carnot has made substantial progress toward automating marine ecosystem monitoring, there are still some significant challenges to overcome. Indeed, it can be interrupted by rough sea conditions, such as strong tidal currents and storms. In addition, biofouling presents a major problem for sensors in the coastal environment, which explains why only a few moored autonomous systems have been deployed in the coastal environment (Blain et al. 2004). Due to sensor failure, phosphate, nitrate, and silicate measurements are not available after 2010. To better explain the large data gap, we should emphasize that we were in an interim phase, facing difficulties in maintaining a system developed and built in the early 2000s, with electronic parts that were no longer available and waiting for the improvement of the smart multi-sensor marine observation platform, COSTOF2, which was driving all of the sensors and dataflow.

As our knowledge and understanding of coastal ecosystems is growing with time, the EOVs (essential ocean variables) and EBVs (essential biodiversity variables) may be updated in the future. This may necessitate the installation of new sensors on the MAREL Carnot station to measure these new variables or parameters.

In future work, we plan to use a multi-scale, multi-source, multi-criteria, and multi-parameter approach to characterize and predict harmful algal blooms in the eastern English Channel caused by Phaeocystis globosa and Pseudo-nitzschia spp. We will do this by combining high-frequency datasets from MAREL Carnot, satellite, and modeling data with low-frequency datasets from other sources. This integrated observing system will be used to identify environmental states present in the region and develop an early-warning system that can anticipate harmful algal blooms in particular, as well as changes in water quality and environmental state in general.

The raw data are present on the official Coriolis website. These data were collected and made freely available by the Coriolis project and programs that contribute to it (http://www.coriolis.eu.org). The dataset after quality control procedures is present on the SEANOE (SEA scieNtific Open data Edition) website (10.17882/39754) (MAREL Carnot, 2023) in file “2004–2022 Coriolis processed data”. Our data are made available according to the FAIR approach (Findable, Accessible, Interoperable, and Reusable).

In conclusion, high-frequency data from the MAREL Carnot instrumented station are useful in many scientific fields, such as phytoplankton ecology, data science, and oceanography. They can be used to describe the environmental state and forecast algal blooms in the eastern English Channel, which is important to warn shellfish farmers and prevent economic losses and health problems. It can also be used with satellite, modeling, and low-frequency in situ data to enhance our understanding of the marine ecosystem.