For the in vitro experimental setup, a 24 h experiment was conducted as a joint experiment comprising the five Helmholtz Institutes Alfred-Wegener-Institute, Helmholtz-Centre for Marine and Coastal Sciences (AWI), Hereon (HZG), Helmholtz-Zentrum für Ozeanforschung Kiel (GEOMAR), Helmholtz-Zentrum für Umweltforschung (UFZ), and Deutsches GeoForschungsZentrum (GFZ) within the framework of Modular Observation Solutions for Earth Systems (MOSES) (Weber et al., 2021). The main goal of the experiment was to compare different sensors as they came from different cooperating research institutes, which are often involved in comparative measurements during joint field campaigns. Thereby, the main focus of the here presented experimental approach was, if the different sensors provide comparable results and not the numerical deviation of the single sensors from an assumed “true” value. This latter topic can only be addressed in a certified sensor calibration lab under strictly controlled conditions.

In the in vitro experiments, 14 different temperature sensors continuously measured the seawater temperature in an experimental tank of 100 cm × 60 cm × 100 cm (600 L). The sea water temperature in the tank was gradually lowered from 18.8 to 15.8°C within a period of 13 h (Figure 2) by continuously adding freshwater of a constant temperature of 15.8°C. Complete mixing in each basin was ensured owing to the circulation circuit. Parallel rails were installed above the basins and the sensors were deployed from them in the basin at the same depth of 60 cm. The sampling frequency of each sensor was set according to the sensor manuals to the highest frequency possible for the sensor [Table 1, column “Max. Sampling Frequency (sec)”], as this is, to our experience, the most applied sampling frequency set-up in operational science.

For the data analysis, all data were averaged over 1 min to reduce the bias toward sensors with a higher sampling frequency. As the experiments targeted the interoperability of the sensors, we used the median temperature of all 14 sensors as reference data for each time step.

The sensors included low-cost water temperature data loggers, water level loggers, instrument clusters, such as CTD probes, and flow-through systems, such as FerryBoxes (4H Jena Engineering GmbH). These sensors represent a price range between 200 and 15,000 EUR. The measurement principle of the different sensors varied with different resolutions and accuracies (Table 1). The sensors are usually applied to a wide range of environmental research questions covering ground water, fresh water, coastal, and marine compartments.

Pre-experiment sensor handling followed standardized routines defined by the sensor manufacturers and individual routines. All sensor operators prepared their sensors exactly, as they normally do for standard scientific missions. Specific across-institute concerted sensor preparation procedures were explicitly not provided, as we wanted to focus on possible variations of measurement between different standard sensors under SOPs as applied by different scientific operators and institutes. Therefore, we did not provide any guidelines with respect to sensor calibration and routine maintenance prior to the experiment, except that the routines must be in full agreement with the respective institute guidelines for good sensor handling practice prior to scientific measurement campaigns.

All analyses were performed with R-Studio (R-Studio Team, 2020) and Python, and the packages are listed in Supplementary Appendix 1. For the in vitro experiments, a Bland–Altman analysis was performed to evaluate the agreement between the two series of measured values. In particular, it provides information about the influence of the height of the measured values on the magnitude of the deviations/differences. The Bland–Altman analysis is based on the quantification of the agreement between two quantitative measurements by determining the bias or mean difference as a measure of accuracy and construct limits of agreement (LOA) as a measure of precision (Altman and Bland, 1983; Bland and Altman, 1986). This is used to evaluate the agreement between two different instruments or two measurement techniques (Montenij et al., 2016). For the in situ and in silico analysis of temperature changes in water temperature in Svalbard over time, we applied ANOVA with subsequent linear regression analysis (temperature versus time). For all virtual samplings where the possibility for a random choice of the sampling hour, day, or week was given (all sampling schemes except for hour and d-2), 100 computer-generated repetitive samplings were performed. For sampling scheme w-2, for example, the computer conducted 100 (virtual) repetitive samplings with a random choice of sampling day (Tuesday, Wednesday, or Thursday) and sampling hour (between 10:00 and 15:00). The results of these 100 samplings were used as input variables for subsequent statistical tests on the effects of the parameters “sampling time,” “sensor-id,” and (in silico) “replicate sampling” on the possibility of detecting an increase in water temperature over time. Furthermore, the 100 slopes of the “samplings” were used to calculate the predictive capacity of a significant increase in temperature over time.

In addition to evaluating the comparability of the tested sensors with respect to sensor accuracy and precision, we also evaluated whether specific sensors may be more appropriate for scientific tasks based on measurement characteristics. We specifically checked if there were sensors that were more appropriate for (1) measuring the “true” temperature as accurately as possible or (2) determining minimal small-scale temperature changes over time. While task (1) has less strict requirements regarding the accuracy of the measurements, task (2) requires a high precision of the temperature measurements.

The median of all sensor data for each time step was determined to analyze the accuracy of the different sensors within the intercomparison experiment. Usually, sensor accuracy is calculated using a “true” value. However, measurements of the true value are extremely challenging, and reference techniques can only provide an approximation. In this set-up, the real “true” temperature was unknown. Therefore, the difference between each measured data and the calculated median at each time step was determined. Figure 2 displays the temporal behavior of all the sensors and the deviation of each sensor to the calculated median. Figure 2 indicates that data from all multiparameter probes were below the median value, except for DEV_3. In addition, FerryBox revealed warmer temperatures. The loggers’ water temperatures were either very close to the median value or above the median value. DEV_8–DEV_10 were of the same type, but DEV_8 measured colder than the median water temperatures, while DEV_3 and DEV_10 measured higher water temperatures, indicating a high variability of this type of sensor.

As described above, one aim was to find reasonable and feasible metrics to select the most appropriate sensors for a specific scientific task. It is essential to be certain that the sensor of interest is as accurate as a reference or as an assumed “true” value. Therefore, it is crucial to measure the agreement between the two sensors.

One approach to evaluate the agreement is using Bland–Altman analysis between two sensors, rather than validating the sensors to a “true” reference. Bland–Altman analysis is based on quantifying the agreement between two quantitative measurements by studying the mean difference and constructing the LOA (Giavarina, 2015). The LOA is a confidence interval, and it is commonly computed by ±1.96 × standard deviation of the difference for each comparison. LOA describes how far apart the measurements of two methods are likely to be for most individuals. Here, Bland–Altman analysis was performed using the median of all sensors as a reference to determine the mean accuracy (bias) (Figure 5). The Bland–Altman diagram in Figure 5 shows the differences between each measuring device (DEV), the median of all sensors (REF) on the Y-axis (DEV − REF), and the mean of these two parameters [(DEV + REF)/2]. Figure 5 shows that the means are different for a bias = 0. A bias close to zero indicates an accurate sensor. The sum of the distance of the upper and lower LOAs to 0 indicates the precision (Figure 5).

In Figure 5, it can be seen that the sensors are precise if the limits are close to the bias. In this example, DEV_9 was the most accurate and DEV_12 was the most precise in relation to the median.

These percentage differences between the reference and the sensor show that mean accuracy is the best benchmark to find the sensor that measures the temperature as accurately or “true” as possible. In Figure 6, the results indicate that within the intercomparison experiment, DEV_7 (FerryBox System, in Figure 6A1) and DEV_12 (logger, in Figure 6A2) are the most appropriate. In addition, parameter precision is suitable for selecting repeatable sensors and can record small temperature changes. For the latter research questions (resolving small temperature changes), DEV_6 (Multiparameter in situ probe, in Figure 6P1), DEV_11, and DEV_12 (Logger, in Figure 6P2) should be selected.

Figure 7 shows the base datasets of water temperature of the Svalbard AWIPEV underwater observatory from 2012 to 2017 of the four single sensors and the quality-controlled data set. The statistical analysis revealed a highly significant increase (p < 0.001, lower right panel, column significance slope) in water temperature over time for all five datasets, with a numerical range of the temperature increase between 0.14 and 0.33°C year^–1 (lower right panel, column slope).

This basic dataset (subsequently referred to as an “hourly” dataset) was used for all subsequent virtual sampling campaigns. Table 4 shows the results of hourly, daily, weekly, and monthly sampling schemes, as well as the effects of a very strict sampling plan with no free choice of the hour or day of the sampling versus a liberal sampling plan where the station personnel (here the computer) can schedule the hour or the day of the sampling within the predefined range according to their wishes.

The results in Table 4 show that in all sampling schemes, “sampling time” and “sensor id” had a significant effect on the temperature measurements. In contrast, “repetition” had no significant effects. Furthermore, the interactions between “sampling time × sensors_id” were highly significant (p < 0.01) for hourly, daily, and weekly sampling schemes, indicating that for these sampling frequencies, the measured temperature change (slope) over time was significantly different among the different sensors.

However, when sampling monthly, this difference between the sensors over time could only be resolved with the restricted sampling scheme when the monthly sampling was performed exactly on the 15th of each month (m1) over the entire time period and only with p = 0.05. In contrast, when sampling monthly with the “half” or “full” sampling scheme (m2, m3), no significant interaction between sampling time and sensor was observed.

Finally, the interaction term including the factor replicate (time × repetition as well as time × sensor × repetition) was not significant in any sampling scheme, indicating that the random selection of the sampling time or day in the 100 virtual sampling events within each time slot did not confound the analysis with respect to hidden temporal patterns introduced by the sampling scheme.

Figure 8 shows the predictive capacity of the different sampling schemes in detail. For this analysis, 100 replicate samples within one sampling scheme (d1, d2, w1, w2, w3, m1, m2, m3) were analyzed. The percentage of 100 repetitive samplings in each sampling scheme was calculated to detect the temperature increase over time with the same significance level as that found in the hourly sampling.

Our analyses show that daily sampling (either full random or restricted, d1, d2) revealed identical results as hourly sampling for all sensors and sampling times, and for the interactions between sampling time and sensor ID.

Weekly sampling was similar except for the interaction term “sampling time × sensor_id.” For this interaction term, the predictive capacity dropped to 0% in the weekly restricted sampling scheme. This shows that only daily sampling allows statistical disentangling of the effects of sampling frequency on the effects of sensor type when analyzing temperature increase over time.

When switching to the monthly sampling scheme, the predictive capacity dropped sharply. This means that with this sampling scheme, it is no longer possible to reliably determine the temperature increase over time from undirected signal noise. In the 100 repetitive samplings, a statistically significant relationship between temperature increase and time was found in only 40%, and the effect of the different sensors on the temperature measurements was found to be less than 20%.

In the next step, we analyzed the temperature measurements of each sensor in detail to focus on the effects of using different sensor types to determine long-term temperature changes (referring to the factor “sensor-id” and “sensor-id × sampling time” in the above analysis). Figure 9 shows the calculated mean temperature increase per year (slope) measured with the different sensors from January 2013 to December 2017 for the individual sampling schemes. Additionally, the standard deviations of the slope measurements of the 100 replicate samplings within each sampling scheme are shown. The analysis revealed that the temperature increases per year (slope) measured by the individual sensors were quite different. The highest temperature increase per year was detected by sensor 13 with an average increase of 0.34°C per year, a minimum value of 0.31°C (±0.08°C), and a maximum value of 0.39°C per year (±0.01°C). With all daily and weekly sampling schemes, the temperature increase measurements of this sensor were significantly higher (p < 0.001) than those of all other sensors. When the sampling frequency dropped to monthly, this difference was no longer significant, and the p-value between the sensors dropped to 0.05 (Table 4), even though the numerical differences in the temperature increase measurements of this sensor to all others were still obvious.

In the last step, which is identical to the calculation of the predictive capacity of the different sampling schemes, we calculated the predictive capacity of each individual sensor to discriminate a real increase in water temperature over time from random temperature fluctuations (Figure 10). This analysis revealed that when using a single sensor, reliable determination of real temperature increase over time (100% detection) from random temperature fluctuations was only possible with hourly and daily sampling schemes. When switching to weekly sampling, the predictive capacity dropped in almost all sensors to a value of only 50% in the validated dataset, most sensors were even below 40%. A higher value above 90% predictive capacity was only shown by the ADCP temperature sensor, which, however, showed significantly higher temperature increase values than all other sensors. When finally switching to a monthly sampling scheme, the predictive capacity of all sensors approached values was close to or at 0.

Our results from the in vitro experiments show that the simple formula of “the more expensive a sensor is, the better is the data quality” does not hold true. The intercomparison experiment showed that the sensors used in the experiment span from sensors with a high accuracy but a lower precision (example DEV 3) to sensors with a lower accuracy but a higher precision (DEV 6). Most interestingly, these sensor characteristics were not linked to the sensor’s cost. A sensor with a high accuracy means that the sensor measures are highly accurate with a low deviation from the expected temperature. On the other hand, a sensor with high precision can discriminate small-scale temperature changes even though the absolute measured temperature may deviate from the expected temperature more than in the highly accurate sensor (JCGM, 2008). This discussion, however, becomes highly complex especially when approaching the resolution limit of a sensor. In this operational limit, it is almost impossible to decide if small-scale temperature variations detected in a highly precise but less accurate sensor are the result of real changes in the field, or simply accuracy deviations generated in the sensor itself. The rule of thumb could be that any validation of small-scale variations in water temperature should be larger than the known accuracy deviation of the particular sensor in question. Following this argumentation, the precision of a sensor becomes a relative value that depends also on its accuracy. The best choice to obtain climate-quality SST data for instance would be to have both, a highly accurate as well as highly precise sensor. These theoretical considerations clearly show, that additional to comparative and more operational experiments, as performed in this study, in depth benchmark studies for each type of sensor are very useful and required for researchers to make a choice according to their available budget and objective of the measurements (e.g., short-term local variations versus long-term global trends across a range of sites).

The sensor systems used in our intercomparison experiment varied in measurement technique and price, from high-end FerryBox systems and multiparameter in situ probes to low-cost single-parameter temperature loggers. Depending on the scientific question and the local circumstances, scientists need to decide whether a single expensive sensor with long-term high frequency logging and multi-parameter measurement options is more appropriate for their program than multiple cheaper sensors or exchanging sensors every month when cheap sensors are available with similar or sufficient sensor accuracy and precision. For instance, the low-cost loggers of HOBO (DEV 11 and 12) measured the temperature during our intercomparison experiment with good accuracy and precision.

Another remarkable finding of the intercomparison experiment was that the behavior of each sensor is sensor-specific, and even sensors of identical type and manufacturer sometimes do not show the same behavior or provide different data under defined experimental conditions (e.g., DEV 8, 9, and 10). It is often assumed that periodic calibration ensures accurate and precise data, as a known reference standard with high accuracy is used for the calibration. In our experiments, all participants confirmed that their sensors were in a calibrated state; thus, we assumed that all sensors were well calibrated and ready for accurate use. The results of the experiments indicate that even proper calibration has the potential to retain errors. Calibrated sensors are assumed to be initially true and have a bias less than their precision error. Our results clearly show that calibrated sensors need to be checked against each other frequently, especially for comparative measurements with multiple sensors, for example, on joint cruises with different ships or synoptic measurements at different places.

This also holds true for the accuracy and precision of new sensors from manufacturer datasheets. These values are often only valid for a brand-new sensor and sometimes do not reflect the specific sensor, but only the sensor type. In this case, the question arises if such reported manufacturer values are trustworthy for field experiments. On the other hand, it has been reported that in some sensors, the accuracy and precision actually improve over time as the sensor stabilizes while in other cases it was observed that the accuracy and precision deteriorates over time, e.g., when the battery power decreases below a certain threshold. This shows that intercomparison experiments as well as proper sensor preparation prior to field campaigns should be a standard routine to assess and document the sensor performance prior to each campaign and that those operational “metadata” of the sensor should be accessible for later data analysis.

Our experiments show that during the planning and implementation phase of measuring or monitoring programs with multiple (different) sensors and in programs where different institutions with different sensor handling procedures are involved, it is highly recommended to perform intercomparison experiments. Such experiments are easy to perform, foster information and knowledge exchange and transfer among sensor operators, and help to select suitable sensors with regard to resolution, accuracy, and price. Our analysis showed that even low-cost sensors can be suitable and the low price allows the implementation of a measurement array at the same cost as a single, more expensive sensor. These kinds of considerations including the respective accuracy and precision information have to be properly documented in the data’s metadata especially when submitted to global datasets. In this context, it must be considered, however, that data of lower accuracy and precision, even though sufficiently accurate and precise for local scientific questions, may compromise global datasets obtained by higher accurate and precise sensors. Therefore, data portal administrators should be aware especially of such considerations when accepting data from various institutions using different sensor types and deployment methods.

In addition, the observed variability between sensors of the same type from the same manufacturer in our experiments supports the need for intercomparison experiments to assess reliability across sensors of the same type. In particular, larger research communities with different departments and cooperation partners need to establish standardized facilities to compare sensors and to carry out standardized calibrations with defined reference values. This information on data intercompatibility is necessary for data blending or common analysis and interpretation, and therefore contributes to the FAIR principles (Wilkinson et al., 2016) of scientific data. Furthermore, such standardized sensor intercomparison and calibration facilities also serve the goal of combining resources to preserve financial and human resources by avoiding the repeated re-examination of problems/issues across institutions.

In addition to proper sensor selection, the re-analysis of the Svalbard dataset from 2012 to 2017 revealed that the overall measurement strategy, in particular the sampling frequency, is crucial for a possible statistical-reliable discrimination of long-term interannual temperature changes. In particular, for long-term field measurements over several years, setting up the sampling scheme must include not only accuracy and precision consideration of the sensors themselves but also the long-term availability of the workforce on site for sensor maintenance, possible weather constraints preventing sampling for some time, and possible temporal or spatial restrictions with respect to access to the area. While scientists often want to achieve a strict sampling plan with fixed sampling days or even hours at as high as possible temporal frequency, logistics station personnel who have to conduct the sampling in the field prefer the sampling plan to be as flexible as possible to fit their daily, weekly, or monthly routines, as well as their preferred field times. Unfortunately, such discussions are often not based on an in-depth knowledge or evaluation of the consequences of the proposed sampling scheme for data reliability and data quality for a certain question, but rather follow the “experience” factor either of the scientist or the “feasibility” factor of the station personnel.

A proper long-term reliable sampling plan and the respective preparation including all the above-mentioned technical, human, and legal points will facilitate the long-term success of a monitoring program and will better focus on the scientific question, instead of technical or logistic issues.

Our experiments revealed that the sampling frequency is most critical for the chance to determine long-term changes in a parameter (here temperature) with a relevant statistical significance. We detected average increases in temperature over time in the shallow area of the Kongsfjorden ecosystem close to the settlement NyÅlesund between 0.1 and 0.4°C per year depending on the sensor used. Using our best estimates based on our quality control dataset, an average increase of 0.22°C per year was calculated. These values fit quite well with the overall estimate of the effect of global warming in the Arctic realm. Recent studies have shown a significantly faster increase in Arctic temperatures due to global warming than the global average, with Svalbard lying in the global hot-spot area in recent decades (AMAP, 2012, 2021). A recent study by Hop et al. (2019) revealed an average increase of 0.14°C year^–1 in the deeper water layers of the Kongsfjorden ecosystem, and AMAP (2021) showed an average increase in air temperature of 3.1°C year^–1. It is assumed that the shallow waters of the Kongsfjorden get an additional temperature pulse from the atmosphere, explaining the superior warming of the shallow water ecosystem compared to the deeper water masses. Considering arctic amplification and the global relevance of Arctic water mass temperature elevation for climate change, it is most important to be able to assess water temperature and the temperature increase over time as accurately and precisely as possible. Our results show that, depending on the sampling frequency per time, a measured increase in water temperature was significant over time or not and this was additionally dependent on the sensor in use. In our experiment, we found significant data only when a daily sampling scheme was applied. All sensors revealed that the temperature increase over time was significant at a p level of at least 0.05, independent of the time of sampling during the day. Hourly or daily sampling, therefore, proved to be a robust sampling scheme when attempting to prove the observed increase in water temperature statistically over time. However, this result is, so far, only validated for our Arctic dataset from Svalbard and it would be interesting to test these findings for other non-Arctic long-term data.

When looking at the ability to significantly prove the observed increase in water temperature by weekly sampling, the probability that the observed temperature increase over time reaches a statistical significance (Figure 9, predictive capacity of slope detection) dropped from 100% (as in the daily sampling scheme) to lower values and became sensor-dependent and partly erratic. While one sensor (Figure 9, sensor 13) detected a significant predicted increase in temperature over time in almost all of the 100 simulated samplings independent of the weekly sampling strategy (w-1, w-2, or w-3), another sensor (sensor 16), had a detection rate less than 20% under identical conditions. Using the quality-controlled dataset (qc), the detection rate of the weekly sampling scheme also dropped to less than 60% in the w-3 sampling scheme, meaning that the chance to prove a long-term increase in water temperature statistically only reached less than 60%, and therefore is almost random.

When shifting to a monthly sampling scheme, the chance of detecting a significant increase over time was almost 0, independent of the sensor used and if the sampling was done on the same day of the month or on a random day of the month.

Summarizing these results, in our monitoring program, a sampling frequency of less than “daily” is inappropriate for trying to discriminate random fluctuations in arctic water temperature from a directional change in temperature over time.

Wiltshire and Manly (2004) discussed the option of not using mean values but rather minimum or maximum values, and whether this may elevate the probability of detecting significant changes in environmental parameters over time. In an additional calculation (Supplementary Appendix 2), we tried this approach for the Svalbard data and used the minimum and maximum temperature values for the calculations. When using maximum temperatures for data integration over time, the calculated temperature increases are partly distinctively deviated from the mean value approach. A temperature increase of up to 0.41°C year^–1 was found when using maximum temperature values integrated over months for sensor 13. In contrast, when using the maximum temperature values per month for sensor 15, a negative trend in temperature was identified at −0.06°C year^–1. Thus, in our scenarios, the use of maximum or minimum values is not recommended for calculating long-term trends in temperature in shallow water areas.

Another interesting issue emerged when examining the results of sensor 13. Independent of the sampling scheme, this sensor revealed the highest temperature increase, with an average of 0.33°C year^–1. Evaluating this value in the context of all other sensors and the quality-controlled dataset strongly suggests that this value is a sensor-specific overestimation of the real temperature increase over time. This may be due to the larger number of measurement gaps due to technical failures. It is well known that data gaps can confound underlying “real” trend signals in long-term datasets especially when the covered overall time period is relatively short and data have a pronounced seasonality (Slater and Villarini, 2016). The observation that in this study, the highest warming trends in the time-series is recorded by the sensor with the largest gaps in the measurements seem to confirm this hypothesis and underlines the overall importance of as complete as possible datasets for monitoring programs in environmental trends studies especially in highly dynamic coastal ecosystems. On the other hand, larger gaps were also observed for sensor 15, which showed an average increase of only 0.18°C year^–1. In addition to the higher temperature increase identified by sensor 13, the probability that this temperature increase was statistically significant was above 90% for all three weekly sampling schemes (Figure 9). So, even though sensor 13 had large gaps and showed a suspiciously high temperature increase over time, the general trend of this sensor was the same as for the other sensors, but the numerical value was most certainly a distinct overestimation.

This result may be explained by the method of linear regression, as the calculation of statistical significance of a slope is done by analyzing the increase in the measured value after time using t-statistics. Therefore, when the rate of the measured temperature increase over time is comparatively constant over time, even though it is too high for a specific area, the statistical test will show a significant result, even though the absolute numerical value is too high. This consideration, however, also indicates that when the overall temperature increase rates over time in a certain area are higher compared to our study site, a weekly sampling strategy may also be valid and provide reliable results.

In contrast, in integration scenarios m-1, m-2, and m-3, almost all calculated slopes were insignificant, except for the 0.33°C year^–1 increase from sensor 13, which showed a p-value of 0.05. This additionally can be taken into account when rating this sensor as suitable for such long-term analysis. Adding the manufacturers metadata to this assessment, the user can extract the information that even though the accuracy of this sensor is 0.01°C, its precision is only 0.4°C. This indicates that this type of sensor may not be appropriate for measuring temperature for long-term trend analysis as the expected changes over time are in the same order of magnitude as the sensor precision. This kind of information and background considerations should be a more prominent part of any datasets metadata as there is a certain risk that such data are used in science especially when available as temperature data for a certain area in international databases.

Summarizing the observed patterns for the hourly, daily, weekly, and monthly sampling schemes, a consistent picture emerges. While hourly and daily sampling provided stable results independent of the sensor and independent of the aggregation procedure (minimum, mean, or maximum values), weekly sampling may show significant results in long-term temperature changes over time; however, these results are highly sensor-dependent and are potentially associated with a high probability of error. In our analysis, monthly sampling schemes did not provide significant results for long-term temperature changes over time, independent of the sensor used and the sampling scenario.

Sampling aquatic environments, especially in remote areas, is time-consuming and expensive. Our results showed that for the Svalbard dataset, only hourly and daily sampling is a reliable sampling strategy for monitoring long-term changes in water temperature for climate change monitoring programs. Even daily sampling programs based on discrete water samplings, for example, with a small ship or even from a pier or any other access point to the water, are not practical, even when considering a year-round operated research base in the Arctic, such as the AWIPEV research base in NyÅlesund used for our study. Winter conditions with extreme outside temperatures and Arctic polar nights make such a human-based sampling program not feasible. Furthermore, measuring further environmental parameters such as pH or chlorophyll a by discrete water sampling is also not feasible on a daily basis, even in more friendly environments, as the workload is too high and hiring extra personnel for such programs is often not possible. Cable observatories are often assumed to be expensive and technically demanding. However, when considering the financial expense and workload effort required for daily sampling based on human operation, cabled observatories are often more cost-effective, not only in remote areas. Cable connected fully automated sampling facilities have become operational standards over the last decade, and data handling procedures for quality control and storage have been developed and established in most scientific institutions. Considering these technological developments and the findings from this study that at least daily sampling is most appropriate and reliable in terms of statistical power to discriminate random fluctuations in water temperature from directional changes over time, observatory technology with sensors measuring at least on a daily resolution is a cost efficient and reliable method for environmental monitoring. Our results are transferable to other aquatic research questions and non-polar regions. Increases in surface water temperatures constitute a global challenge and are monitored in many coastal and terrestrial regions. Hence, it is important to evaluate sensor behavior and provide elaborate and feasible sampling schemes.

However, these results do not address the problem that sensor-based measurements have a higher potential for bias than discrete water samples. Therefore, we propose a synergistic approach of sensor-based measurements of at least a daily frequency, with a regular discrete sampling scheme several times per year to validate the sensor data and ensure a high accuracy of the continuous sensor data. In our experience, such a validation by discrete water samples must be done in pre-defined intervals, depending on the variable and the environment, as well as on the requirements of the data quality.