The instrumentation that was used for monitoring the FSC of sailing ships is shown in Fig. 1. The UAV was a MATRICE 600 PRO (SZ DJI Technology Co., Ltd., Shenzhen, China). This type of UAV cannot be used on rainy days or when the wind speed is higher than 8 m s-1. The white box installed underneath the UAV in Fig. 1 is the aforementioned second-generation pod for measuring exhaust gas. When the UAV approaches a ship's plume, the gas pump in the pod draws air using the gas probe. The water vapor, particles, and the soot in the gas are subsequently removed by a hose filter valve. The sensors detect the gas, and measurement information is sent out by communication modules. The pod has dimensions of 20 cm × 12 cm × 9 cm and weighs 900 g.

The sensors used were able to measure both SO2 and CO2 and were purchased from Shenzhen Singoan Electronic Technology Co., Ltd., China. The SO2 sensor is based on the electrochemical method and has a measuring range of 0–10 ppm, an accuracy of ±3 % (0.3 ppm), and a response time (T90) of ≤30 s. The CO2 sensor is based on the nondispersive infrared analyzer method and has a measuring range of 0–10 000 ppm, an accuracy of ±3 % (300 ppm), and a T90 of ≤30 s. The T90 represents the time taken to reach 90 % of the stable response following a full range change in the sample concentration. These sensor characteristics were provided by the instrument manufacturer and were ensured to be within the tolerances by calibration. The zero and full scales are usually calibrated by a standard mixed gas when the equipment is used on a daily basis. The major parameters of the UAV system are listed in Table 1.

As illustrated in Fig. 2, the monitoring region was the channel of the Yangtze River estuary, near the Waigaoqiao port area to the north of Shanghai. The Yangtze River is the longest river in China and the third-longest river in the world. Shanghai is one of the most prosperous cities in the world and had a permanent resident population of approximately 24 million people at the end of 2017 (Shanghai Municipal Bureau of Statistics, 2017). The Waigaoqiao port area is only 20 km from the city center, and the air pollution caused by ship emissions directly affects the urban air environment and the health of residents (Wang et al., 2019; Feng et al., 2019). The experimental area of the MISEE project is mainly within the Waigaoqiao port and the Yangtze River estuary.

During the experiment, the operator took a patrol boat to the channel and selected a target ship at random. After identifying the target ship for monitoring, the patrol boat would accelerate to a few hundred meters to the left or right ahead of the vessel. The patrol boat would then stop and the UAV would take off from its deck and fly towards the plume of the target ship to measure the concentrations of SO2 and CO2 in the plume (Fig. 3).

During the measurements, the operator adjusted the position of the UAV to ensure that it was in the ship's plume. Real-time measurements of SO2 and CO2 were made such that the pod could effectively detect the plume. Generally, it was necessary for the UAV to follow the ship's funnel mouth for approximately 5 min, as illustrated in Fig. 4. The target ship continued to move during the measurements; hence, it was followed by the patrol boat in order to avoid the UAV moving too far away from the operator. When the operator was sure that valid data had been collected, the patrol boat stopped and the UAV returned and landed back on the deck.

The FSC in this study was obtained directly by sampling the gas concentrations in the ship plumes using the UAV. The enhancements of SO2 and CO2 in measurements that were affected by exhaust gases were calculated, and the ratio of these SO2 and CO2 peaks was used to calculate the FSC (Eqs. 1, 2). This method has been widely used to calculate the FSC in related studies (Alföldy et al., 2013; Pirjola et al., 2014; Balzani Lööv et al., 2014; Beecken et al., 2014, 2015; Kattner et al., 2015; Zhou et al., 2019). In the calculation, the molecular weights of carbon and sulfur are 12 and 32 g mol-1, respectively, and the carbon mass percent in the fuel is 87±1.5 % (Cooper et al., 2003). By assuming that 100 % of the carbon content of the fuel is emitted as CO2 and sulfur is emitted as SO2 and other forms, the FSC mass percent can be determined using Eq. (1): 1FSC%=Skgfuelkg=SO2ppm⋅A(S)CO2ppm⋅A(C)⋅87%+R=0.232∫SO2,peak-SO2,bkgdtppb∫CO2,peak-CO2,bkgdtppm%+R=120EFgSO2kgfuel-1+R, where R represents the sulfur content that is emitted in forms other than SO2, as preliminary studies have shown that 1 %–19 % of the sulfur in fuel is emitted in other forms, possibly SO3 or SO4 (Schlager et al., 2006; Alföldy et al., 2013; Balzani Lööv et al., 2014). EF is the emission factor, and bkg represents background. In Eq. (1), if the sensors measuring SO2 and CO2 have approximately the same response time and can be set to be synchronized, the peak concentrations of SO2 and CO2 can be used to calculate the FSC; otherwise, integrals need to be used. In our research, the sampling rates of the SO2 and CO2 sensors were 1 s, and integrals were used because the two sensors could not be completely synchronized.

The continuous measurement data for two typical plumes (2019-4-15B and 2019-3-29A) are exhibited in Fig. 5. The data for plume 2019-4-15B (Fig. 5a) were considered to be “good” quality, whereas those for plume 2019-3-29A (Fig. 5c) were considered to be “poor” quality. Data were determined to be good quality when obvious, easily distinguished peak values were observed, whereas less obvious peaks that still corresponded to a result were considered to be poor-quality data. Meanwhile, the correlation between the SO2 and CO2 time series is a key factor in judging quality. Assuming that the gas is completely mixed, the variation trend of the SO2 and CO2 measurements should be the same (although there may be some deviation because the corresponding time of the sensors was not consistent) and can be identified in the peak area.

The selection of peak values leads to uncertainty because when the area ratio is selected for the calculation, the starting and ending time points of the area are still associated with substantial uncertainty. Figure 5b and d depict the average concentrations of the SO2 and CO2 measurements (from Fig. 5a and c, respectively) for 10 s periods. The peak value of each average concentration was selected for the calculation. This process is equivalent to selecting the area ratio of SO2 to CO2 within 10 s for the calculation, as shown in Eq. (2).

2FSC%=0.232∫SO2,peak-SO2,bkgdtppb10∫CO2,peak-CO2,bkgdtppm10%+R≈0.232AVGSO2,peak-AVGSO2,bkgAVGCO2,peak-AVGCO2,bkg%, where AVG (⋅) is the calculated function for the average measurement value within 10 s; hence, the data in this study are the average values of measurements in 10 s. When the UAV took off from the patrol boat and flew high into the air, the SO2 and CO2 concentrations were relatively low. The background values were obtained at this stage as the minimum SO2 and CO2 concentrations. As the UAV flew into the plume, the measured concentrations of SO2 and CO2 increased. The obvious, stable maximum values in the observations of the average measurement values should be selected as the peak values. It can be seen that using the average values of measurements within 10 s makes it easier to select the peak values, especially with respect to poor-quality data. However, as there can still be several options for peak values, the data treatment methods reported by Zhou et al. (2019) were incorporated in this study to select the most appropriate peak values. In Fig. 5b, the time point of selected peak values is at 10:19:11 LT (local time). The measurement values from 10:19:57 to 10:20:15 LT were not used because the CO2 concentration covered the full range. In Fig. 5d, the time point of the selected peak values is at 10:38:27 LT. The measurement values from 10:39:57 to 10:41:41 LT were not used because we excluded data exhibiting either dramatic changes or errors in continuous observations. The details for selecting the peak values are given in Table 2.

In previous research (Zhou et al., 2019), the main uncertainties of UAV measurements were summarized as sensor uncertainty, measurement uncertainty, calculation uncertainty, and exhaust uncertainty. The instrument calibration method, UAV flight procedures, and the data treatment methods were designed to reduce these uncertainties. However, some uncertainties remain, as discussed below.

To make the UAV lightweight and convenient, the second-generation pod was only equipped with SO2 and CO2 sensors and a simple filter. We did not account for the interference that some factors might have caused, including interference due to (1) the cross-sensitivity of the SO2 sensor to NO2, (2) the impact of a large temperature change in the exhaust plume, and (3) water vapor and/or particle contamination of the instruments.

The average gas concentration within 10 s was chosen for the FSC calculations; however, this does not mean that 9 or 11 s could not have been selected. To demonstrate this, a comparison calculation was carried out using both 9 and 11 s, which showed that these led to very little difference in the results. Nevertheless, it is necessary to ensure that the gradient of the gas measurements is stable within the sampling time (the interval length of the integral). Moreover, the interval length cannot be too short (e.g., 2 s) or too long (e.g., 20 s). If the interval length is too short, it is difficult to determine whether the measurements are stable and undisturbed over time. Similarly, if the interval length is too long, it is also difficult to ensure that all of the measurements in the integral interval are stable and undisturbed. In addition, during the flight of the UAV in this study, the time available for measuring the plume was ∼5 min. As both the ship and the UAV were moving at this time, it was virtually impossible to ensure that the UAV was flying consistently within the plume and obtaining stable measurements. Accordingly, 10 s is also a relatively appropriate value for the measurement process.

Nevertheless, there is also some uncertainty associated with choosing the peak values. After ruling out the peak values across the full range as well as those corresponding to dramatic changes, the global maximum values were selected as the peak values to calculate the FSC. The maximum values probably correspond to the measurements taken in the center of the ship's plume. At that location, the measurement values were relatively stable, and the probability of interference from other factors was lower. Furthermore, the higher the peak value is, the greater the proportion of exhaust gas is; hence, the impact from the incomplete mixing of the exhaust gas with clean air is relatively small.

In summary, the obvious and stable maximum values are selected as peak values to calculate the FSC. There are, of course, situations where multiple similar peaks can occur simultaneously. In this case, their calculated FSCs may be very similar, and the results obtained by the calculation of the highest peak should have high credibility, such as the measurements of plume 2019-4-15B. Meanwhile, the occurrence times of the peak SO2 and CO2 values sometimes have a small deviation that usually corresponds to a few seconds. This is due to two different sensor response times and leads to three different options for selecting the peak values: (1) the time point of the peak SO2 value with the CO2 value at the same time; (2) the time point of the peak CO2 value with the SO2 value at the same time; (3) the peak SO2 and CO2 values at different time points. Option 3 was selected in this research.

Additional uncertainties were encountered during our monitoring of sailing ships because the UAV was usually hundreds of meters away from the operator. The location of a plume depended primarily on the following three aspects. (1) The position of most plumes with black smoke could be identified using the operator's visual judgment. (2) The real-time image shot by the camera could be used to assist in finding the ship's funnel mouth. (3) In the measurement process, the real-time measured concentration sent to the receiving equipment gradually increased, indicating that the UAV was approaching the center of the plume. However, the operator occasionally faced difficulties in accurately determining the ship's plume, which led to failed measurements. We attempted to measure more than 40 ship plumes in open water; however, only 27 of them resulted in good- or poor-quality data, i.e., usable data.

The deviation of the estimated FSC value obtained by the first-generation pod was <0.03 % (m/m) for an FSC level ranging from 0.035 % (m/m) to 0.24 % (m/m) (Zhou et al., 2019). The second-generation pod was also verified on berthing ships by using this method at a similar FSC level, and the accuracy was approximately the same (see Sect. 3.1). These verifications of the deviation were based on the FSC measurement of berthing ships, which did not exceed the Chinese DECA FSC limit of 0.5 % (m/m). However, some of the sailing ships did exceed this limit. It should be noted that the deviations for different FSC levels were not the same. Based on previous studies, the deviation of the FSC obtained from high-sulfur plume should be greater; for example, Van Roy and Scheldeman (2016a, b) estimated relative uncertainties of 20 % at a level of 1 % (m/m) FSC and 50 %–100 % at 0.1 % (m/m) FSC. Therefore, the deviation of sailing ships may be >0.03 % (m/m) when the FSC exceeds 0.5 % (m/m). Nonetheless, our UAV system was still able to accurately detect an FSC that obviously exceeded 0.5 % (m/m).

Before monitoring the sailing ships, we first monitored 11 berthing ships between March and April 2019 in the Waigaoqiao port to verify the accuracy of the second-generation pod. While one person operated the UAV to monitor one of the plumes, two maritime law enforcement officers boarded the corresponding ship to collect a fuel sample. Both processes took approximately 10–20 min. The fuel samples, which are considered to represent the true FSC values, were then sent for chemical analysis in a laboratory. The estimated (UAV) and true FSC values are listed in Table 3 along with the identification number of each plume and the time and serial number. Table 3 shows that the deviation generally did not exceed 0.03 % (m/m) for an FSC level of between 0.03 % (m/m) and 0.22 % (m/m), except for plumes 2019-3-22A and 2019-4-3B. Additionally, when the FSC of a target ship was low, for example, when light diesel fuel was used, the measured SO2 concentrations were mostly zero. When this occurred, the FSC was generally <0.02 % (m/m), as seen for plumes 2019-4-3A and 2019-4-12A.

Between March and December 2019, effective monitoring of 27 sailing ships was undertaken using the UAV that took off from the patrol boat (Table 4). The FSC of 23 berthing ships measured by the first-generation monitoring equipment and the FSC of 11 berthing ships (Table 3) measured by the second-generation monitoring equipment in this study were taken as the FSC monitoring results for berthing ships. We compared the distribution of the FSCs of these 34 berthing ships with those of the 27 sailing ships. Figure 6 shows that the FSCs of the sailing ships were considerably higher than those of the berthing ships; the FSC of all 27 sailing ships exceeded 0.1 % (m/m), and the FSC of 12 of these ships exceeded the Chinese DECA FCS limit of 0.5 % (m/m), which included five exceedances of 1.5 % (m/m). The uncertainty in the assessment is not small; however, the results thus far do not lead to optimism with respect to the FSC used by ships sailing in the area. The reason for this is that although berthing ships are sometimes boarded by maritime law enforcement officers for examination, an effective approach for monitoring the FSC of sailing ships in open water that led to prosecution by China's maritime authorities did not existed prior to the present study.

According to the monitoring results, law enforcement officers of the Pudong Maritime Safety Administration intercepted four sailing ships for which the UAV FSC results were of a good quality and exceeded 1.5 % (m/m). The officers boarded these ships for inspection on 15 July, 14 August, 20 August, and 27 September 2019, respectively, and took fuel samples, which were sent for chemical analysis in a laboratory. The FSC of all four respective fuels was also found to exceed 0.5 % (m/m): 0.534 % (m/m), 0.744 % (m/m), 0.813 % (m/m), and 1.991 % (m/m) (in chronological order). The reason that three of these laboratory results did not exceed 1.5 % (m/m) was related to the fact that ships cannot stop immediately in the channel for inspection and have to sail to the anchorage point; when the officers boarded the ships to take samples they found the crew taking various measures to drain the high-sulfur fuel in the main engine fuel oil pipeline. This means that the chemical analysis results of the sampled fuels were obviously lower than those of the UAV monitoring. Nevertheless, the four inspections successfully confirmed that the FSC of the fuels exceeded the standard for sailing ships. The inspection on 15 July 2019 was the first time that a sailing ship's FSC failed to meet Chinese regulations, and this aroused wide concern in the shipping community.