Identification and detection of high NO x emitting inland ships using multi-source shore-based monitoring data

In urban areas situated along busy waterways like the Yangtze River, the diesel engines of inland navigation ships emerge as significant contributors to air pollution. Among these vessels, certain high-emission ships exhibit considerably higher levels of nitrogen oxides (NOx) emissions compared to others. To effectively identify such ships, this study employed a cost-effective ship emission monitoring sensor platform, comprising high-precision gas sensors, automatic identification system receiver, and sensitive meteorological sensors, along the Yangtze River in Wuhan City. By combining multi-source shore-based monitoring data, we identified ship emission signals and proposed a high-emission ship detection method using inverse modeling. Using this method, we successfully detected inland high-emission ships based on two months of monitoring data. Furthermore, the relationship between different ship types, sizes, speeds, and ship NO x emission rates were investigated. The results of this study are beneficial for strengthening the regulation of high-emission vessels in inland waterways, thereby reducing the adverse impact of ship emissions on the environment and climate. It also encourages the inland shipping industry to adopt more environmentally friendly technologies and fuels, as advocated by the International Maritime Organization.


Introduction
Shipping constitutes a significant source of air pollution in coastal ports and inland waterways (Eyring et al 2010, Merico et al 2017, Chen et al 2018, Feng et al 2019, Huang et al 2022, Zhou et al 2023).The Yangtze River, as a vital inland waterway in China, has seen its annual freight volume surpassing 3 billion tons, consistently holding the top position in global inland shipping for over a decade (Zhao et al 2020, Zhou et al 2023).The bustling ship traffic in the Yangtze River contributes to emissions of pollutants such as sulfur dioxide (SO 2 ), nitrogen oxides (NO x ), particulate matter (PM), and aerosols, which not only pose threats to the human respiratory system but also contribute to environmental issues such as water eutrophication, acid rain, and the greenhouse effect (Corbett et al 2007, Murena et al 2018, Liu et al 2019, Ramacher et al 2020, Xie et al 2021).
Currently, within the context of Emission Control Area (ECA) policy, many scholars focus on detecting high SO 2 emission ships, especially for ships using fuel with high sulfur content (Peng et al 2021, Wu et al 2023).However, there is a lack of a unified definition for vessels with high NO x emissions and effective identification methods.Kurchaba et al (2023) proposed a method for detecting anomalous NO 2 emitting ship using TROPOMI satellite data and machine learning, which does not prioritize high emitters but rather focuses on deviant emitters.In recent years, there has been a growing emphasis on the NO x emissions from ships, particularly the high emitters operating on inland waterways (Eger et al 2023, Topic et al 2023).Since the establishment of the inland ECA in the Yangtze River in 2020, which imposed restrictions on the sulfur content in ship fuels, emissions of SO 2 and PM 2.5 have seen a significant reduction (Zhang et al 2022a, Anastasopolos et al 2023).However, the emissions of NO x from ships have scarcely decreased.Contini and Merico (2021) point out that emissions of gaseous pollutants, especially NO x of shipping are generally larger than those of other pollutants like PM and that could be a concern for meeting air quality standards in coastal areas.Therefore, how to identify and detect these highemission ships is the key to reducing pollution from inland vessels.
For vehicles in road traffic, high emitters are typically identified based on emission factors (EFs), with significantly higher nitrogen oxides EFs observed in high-emission vehicles compared to other vehicles (Shen et al 2022).Similarly, this method can also be applied to the identification of high-emission ships.There are various methods for testing ship EFs, including bench testing (Geng et al 2016), on-board testing (Yang et al 2022, Fan et al 2023), and plume testing (Zhao et al 2013, Ekmekçioglu et al 2020, Krause et al 2023).Bench testing is conducted in controlled environments using high-precision instruments, yielding accurate results but struggling to reflect the emission characteristics of marine engines under actual navigation conditions (Fan et al 2023).On-board testing refers to the use of portable emission measurement systems (PEMSs) directly at the ship's exhaust pipe for testing, enabling the acquisition of EFs under various sailing conditions.Bench testing and on-board testing based on PEMS are considered the standard methods for measuring ship EFs.However, they require expensive instruments and intricate manual operations, and can only measure individual ships, limiting the potential for screening conventional high-emission ships under real-world conditions.Shore-based plume testing is an indirect method that involves capturing ship plumes diffusing in the atmosphere using sensors (Contini et al 2011, Krause et al 2023).Compared to bench testing and on-board testing, shore-based monitoring methods offer the advantage of automated snapshot measurements, enabling data collection over a broader range and improving screening efficiency.
With these considerations, this study proposed a method for identifying and detecting high NO x −emitting ships based on multi-source monitoring data provided by a shore-based monitoring platform, as depicted in figure 1, outlining a specific workflow.The approach comprises two main steps: (1) from the monitoring data series, we extract peaks and, by integrating automatic identification system (AIS) data, identify those peaks caused by ship emissions.These are then correlated to form a large sample dataset that includes both the peak values and information about the ships, and (2) based on each peak's information, calculating the NO x emission rate for the associated ship, converting it into an EF, and comparing it with existing ship emission limits to identify high-emission ships.The organizational structure of this study is as follows: section 2 provides a detailed introduction to the shore-based monitoring platform and the proposed models; section 3 analyzes and discusses the results from the proposed methods; and section 4 presents the conclusion.

Shore-based monitoring platform
In this study, we developed and deployed a monitoring sensor platform designed for continuously measuring passing ships and their pollutant emissions along the Yangtze River (see figure 2).The monitoring platform consists of gas analyzers, meteorological sensors, and an AIS receiver, which use Internet of Things technology to transmit real-time monitoring data and store it at a data terminal.The gas analyzers integrate low-cost sniffing sensors to collect the diffusion plume concentrations of CO 2 and NO x from passing vessels at a 1 s time resolution.The meteorological sensors measure wind speed and wind direction using a rotating anemometer and a wind vane, with a time resolution of 5 s.Additionally, the measurement station is equipped with an AIS receiver to identify ship information and specific timing as ships pass through the measurement location.Table 1 provides the detailed specifications of the instruments used.
The measurement site is located on the northern bank of the Yangtze River in the Wuhan section (114.27• E, 30.54 • N; figure 2(a)), where ship traffic is busy, significantly increasing the possibility of monitoring ship exhaust emissions.Considering the rising effect of the plume, the monitoring equipment is installed on a pole at the top of a barge structure on the riverbank and positioned downwind of the prevailing wind (figure 2(b)).The top of the barge is approximately 10 m above the water surface, and the sampling port of the monitoring equipment is about at 11.5 m, utilizing a pump-suction design to maximize the capture of ship emission plumes (figure 2(c)).The testing was conducted in July and August 2022, and the predominant southeast winds in summer aided in transporting most of the exhaust plumes from passing ships to the measurement location.

Peak extraction and ship association algorithm
Statistical correlation is only established when the monitoring sensors detect the passing ship plume signals and show fluctuations.In this section, we proposed an algorithm that combines monitoring time series data with AIS data to identify peak fluctuations caused by ship emissions and match them to the corresponding vessels.
The first step is to extract peaks in the raw monitoring data series.Initially, the sliding time window algorithm is applied to calculate the background concentration baseline, which represents variations in background concentrations due to meteorological factors and sources of emissions other than ships.To capture fluctuations in normal background concentration, we choose to set the size of the sliding time window to 10 min.Within each time window, the Isolation Forest algorithm is employed to eliminate noise points and outliers caused by external conditions and instrument factors.Since outliers are significantly different from the majority of the data points, multiple decision trees are constructed to 'isolate' the observations.An isolation score is calculated for each data point based on the length of the path.The lower the score, the more likely the point is an outlier.Detailed information about the Isolation Forest algorithm and its procedure can be found in the literature (Ding and Fei 2013).Based on the isolation score within each window, normal background concentration is determined, and the median of the normal concentrations in that window is selected as the background concentration baseline.Thus, subtracting the baseline concentration from the observed values allows the separation of peak variations.After that, the occurrence time, duration, and peak value of each identified peak are recorded.
The second step involves identifying peaks caused by ship emissions and conducting target association, requiring synchronization with meteorological data (such as wind speed and direction) and AIS trajectory data.The main idea of the ship association algorithm is as follows: (1) match each peak value with nearby vessels.If multiple ships are matched, the corresponding peaks will be eliminated, as it is not possible to attribute them definitively to a single ship; (2) combine wind direction and ship heading to assess the probability of each ship's plume diffusion to the monitoring site; (3) automatically identify potential ship peaks by meeting several criteria defined by the algorithm (see table S1) and associate each peak with the corresponding ship.The criteria and detailed steps of the algorithm are outlined in the supplementary document, and an illustrative example of the criteria for associating peaks with ship emissions is depicted in figure 3. Once a ship is identified as the source of NO x peaks, the relevant information of the ship (such as Maritime Mobile Service Identity, ship name, ship type, position, heading, and speed), along with the monitoring data within the time period when the peak occurred, will be saved as a dataset.

Emission rate inversion and high-emitter detection
The data from monitoring site reflects the NO x concentration at site locations rather than at the ship's stack outlet.Thus, it is necessary to employ an inverse modeling to infer the ship emission rate contributing to the concentration increase at the measurement site.In this study, we propose a model that combines peak concentration data with the enhanced Gaussian plume model to back-calculate ship emission rates.
The Gaussian plume model is a continuous steady-state gas dispersion model that has been widely employed in recent years for inverse modeling simulations of emission source strength (Ye et al 2020, Liu et al 2022).Considering the mobility of ships, this study extends the traditional Gaussian model to establish a pollutant emission dispersion model for inland ships at any given moment during navigation.The improved Gaussian model is expressed by the following equation ( 1) where, C (x, y, z, t) represents the concentration of NO x at the monitoring station location (x, y, z) at time t (g m −3 ); Q ship denotes the emission rate from the ship's stack (g s −1 ); u represents the composite wind speed in the x-direction, (m s −1 ); H e is the height of the plume (m), where the plume center rapidly bends downward due to the wind and the vessel's movement; σ x i , σ y i , σ z i respectively signify the diffusion coefficients in the x i , y i and z i (m).
For each ship, the simulation starts 2 min prior to the corresponding peak time.AIS signals within the given time range are collected and interpolated to a 1 s time resolution.Assuming the model operates under steady-state conditions, wherein meteorological parameters such as wind speed, wind direction, atmospheric stability, and others remain constant during the pollutant dispersion period.To adhere to the model's fundamental assumption, we utilize the 5 min average values of wind speed and wind direction obtained from measurement station data as inputs.Meteorological inputs for the model include relative wind speed and wind direction.The diffusion coefficients are selected based on atmospheric stability, with stability classes determined from conventional meteorological data such as ground wind speed, cloud cover, and sunlight conditions (Kahl and Chapman 2018).The chimney height for ships is set at 8 m above the average water surface, considering that cargo ships dominate in inland waterways, and chimneys at the rear of these ships tend to have similar heights.
As the Q ship is unknown, the model is assumed to operate at an arbitrary but constant emission rate.To derive the ship's emission rate, a comparison is made between the actual measured concentration at the monitoring site and the model-simulated dispersed concentration, an example is illustrated in figure 4. Assuming that the model has reconstructed the entire process of ship emissions of pollutants, including their dispersion in the atmosphere and the corresponding changes in concentration at monitoring site, the only difference between the simulated and the measured concentrations arises from the difference in emission rates.Therefore, the actual emission rate of the ship (Q ship ) can be derived using the simulated peak concentration area and the measured peak concentration area by equation ( 2) where, C measured represents the area under the peak concentration measured at the monitoring site, and C simulated represents the area under the peak concentration simulated by the dispersion model.Utilizing the peak area proves to be a more dependable metric than the peak maximum, as it considers the overall quantity of pollutants reaching the measurement site (Krause et al 2023).Furthermore, the ship emission rates, measured in g s −1 , are converted to ship EFs (EF) in g kWh −1 where, Q ship is the instantaneous emission rate of the ship in g s −1 ; and P represents the engine power of the ship from AIS data in kW.Therefore, the estimated ship EFs can be used for detection of high-emission ships by comparing them with existing emission regulations (see table S2).

Traffic and meteorological characteristics
Figure 5(a) illustrates the distribution of ship traffic flow in the region based on vessel types.During the two-month monitoring period, utilizing data from AIS receivers and excluding vessels with zero speed (indicative of anchorage or mooring situations), a total of 16 225 instances of vessel transits were observed within the monitoring area.Among them, cargo ships dominated with a total of 11 600 instances, accounting for about 71.5%, followed by tankers (about 11.2%), passenger ships (about 5.8%), container ships (about 3.8%), engineering ships (about 2.9%) and enforcement ships (about 1.8%), respectively.On average, the daily ship traffic flow amounted to approximately 270 instances.It is worth noting that the mentioned instances include cases where vessels were recorded multiple times due to round-trip voyages.As a result of the statistics conducted by MMSI code, the actual counts of monitored cargo ships, tankers, container ships, passenger ships, engineering ships and enforcement ships were 3,652, 816, 152, 32, 23 and 6 respectively, with an additional 286 ships classified under other types.Figure 5(b) shows a rose diagram of regional wind speed and direction during the monitoring period, which clearly presents the frequency of wind directions across 16 compass points and the distribution of wind speed intervals for each direction.It can be seen that, the prevailing winds originated from the east and southeast during the monitoring period.Wind speeds were mainly concentrated in the 2-4 m s −1 range, with an average wind speed of approximately 2.65 m s −1 .It indicates that meteorological conditions during the monitoring period were favorable for the dispersion and transport of pollutants emitted from ships to the monitoring station.

Results of the identification of peak from ship emissions
A total of 2841 peaks were identified as being associated with ship emissions based on the peak extraction and ship association algorithm.To understand the factors affecting the monitoring of emissions from inland ships, the identified peaks were analyzed from four dimensions: ship type, sailing direction relative to water currents, ship speed, and the distance of ships to the monitoring stations.
Figure 6 shows the results of the statistical analysis of the identified peaks with respect to different factors.In terms of ship type, cargo ships emerge as the predominant contributors to the identified peak, accounting for about 85.3% of the total.They are followed by tankers, passenger ships and container ships, aligning with the general ship type distribution observed in ship traffic flow (figure 5(a)).Notably, engineering ships and enforcement ships were not found among the peaks, likely attributable to their comparatively lower emission levels and greater distance from the monitoring stations.In terms of sailing direction relative to water currents, a higher number of upstream vessels were detected, accounting for about 80% of the total, while downstream vessels constituted only 20%.This can be attributed to two main reasons.Firstly, upstream vessels are positioned on the side of the monitoring station, making them closer and thus more easily detectable.In contrast, downstream vessels are situated farther away from the monitoring station, and pollutants within the ship's funnel may become extremely diluted after dispersing over long distances, posing difficulties for monitoring.Secondly, due to the influence of water currents, upstream vessels need to overcome the resistance of the current, requiring more fuel consumption and consequently resulting in higher emissions (Huang et al 2020).
Figures 6(c) and (d) respectively illustrate the relative frequency of detected ship emissions at different speeds and distances from the monitoring stations.It can be observed that, within the identified peaks, ship speeds range from a minimum of 7.1 knots to a maximum of 16 knots.The majority of ships operate at speeds between 11-13 knots.In general, downstream vessels exhibit higher speeds than upstream vessels.Thus, it is evident that when the speed exceeds 13 knots (indicating downstream vessels), the frequency of detected ship emissions significantly decreases.Examining the distances between ships and monitoring stations, the highest number of detected ships falls within the 200-300 m' range, making up about 38% of the total.Following closely are the ranges of 100-200 m and 300-400 m.Additionally, as the distance increases, the probability of detecting ship plumes gradually decreases.The furthest distance at which ships were detected by the monitoring stations in this section is 790 m.

Analysis of NO x emission rates from ships
Based on the identified peak concentrations associated with ship emissions, the method proposed in section 2.3 was employed to estimate the instantaneous emission rates at the stack outlet for each vessel.In general, the dependence of ship emission rates on ship parameters is complex and varies from ship to ship.Therefore, this section focusses on analyzing the impact of ship static parameters such as ship type and ship length, as well as dynamic parameters like speed and draft, on ship emission rates.Figure 7 shows box plot analyses of the influence of different ship parameters on emission rates, while table 2 displays the statistical results of emission rates under various classifications.
From figure 7(a) and table 2, it can be seen that there are variations in the emission rates among different types of inland ships.Passenger ships exhibit the highest average NO x EF at 2.44 g s −1 , followed by cargo ships, tankers, and container ships at 2.30 g s −1 , 2.19 g s −1 , and 2.01 g s −1 , respectively.One possible explanation is that passenger ships, due to their  frequent maneuvering or operating in high-load conditions, result in higher NO x emissions.In terms of maximum emission rates, cargo ships have the highest rate at 14.34 g s −1 , followed by tankers, container ships, and passenger ships at 11.15 g s −1 , 5.90 g s −1 , and 5.78 g s −1 , respectively.In terms of ship size, as shown in figure 7(b), larger vessels exhibit higher NO x emission rates compared to smaller vessels.This is because larger vessels typically feature engines with higher power to provide the necessary propulsion and maneuvering capabilities, as confirmed by Eger et al (2023).
In addition to ship static parameters, the ship speed also influences the NO x emission rate, as shown in figure 7(c).Except for vessels with speeds in the range of 0-4 knots, the average NO x emission rate of vessels increases with higher speeds.Vessels operating at low speeds exhibit higher NO x emission rates, with an average emission rate of 2.96 g s −1 , consistent with the findings of Fan et al (2023).The primary reason is that during low-speed operation, vessels experience significant acceleration, leading to uneven mixing of oil and air within the engine cylinder.This results in the formation of locally high-temperature and oxygen-rich zones, consequently causing greater fluctuations in NO x emissions (Zhang et al 2022b).In addition, the draught of a ship reflects its cargo carrying capacity to some extent, and a fully loaded ship has a deeper draught compared to an empty ship.However, figure 7(d) shows that changes in draught have little effect on the emission rate of a ship, and there is no obvious correlation between ship draught and emission rate.

Analysis of high-emitters in inland ships
Evaluating the NO x EFs of inland ships in the Yangtze River based on the current NO x emission limits established by the IMO and some developed countries (see table S2), we then identified high-emitters within the monitoring dataset.Figure 8 shows the results of high-emitters under various emission standards, including China II, USA IV, Euro V, and IMO Tier III.Each scatter in the figures represents the NO x emission level of a vessel, with blue dots indicating low-emission vessels, red markers indicating high-emission vessels, and dashed lines representing the maximum limits of each emission standard.Meanwhile, the number of identified high-emitters and low-emitters is annotated in figure 8.For the detected 2841 ships, under China's current ship engine emission standards (China II), 845 ships were identified as high-emitters, accounting for 29.7%.These ships exhibit elevated pollution emission levels, making a significant contribution to the overall ship emissions.When comparing this result with other standards such as those of the United States, the European Union, and IMO, the proportion of highemitters shows a substantial increase, with percentages of 89.4%, 58.7%, and 77.3%, respectively.It indicates a severe non-compliance of NO x emissions from inland ships in China relative to those standards.
High NO x emitters are typically associated with ship engines and fuel.The engine technology of some older vessels may lack the efficient combustion control systems or exhaust control systems, making them more prone to generating higher levels of NO x .Figure S1 further provides a statistical analysis of the age distribution characteristics of the 845 high-emission vessels.Among these vessels, the age ranges from a minimum of 7 years to a maximum of 30 years, with an average age of about 12 years.This indicates that a significant number of older ships emitting NO x above the standards are still operational in the Yangtze River.The research results above highlight the formidable challenges facing the control and management of NO x emissions from inland ships.Urgent measures are needed to expedite the upgrade of ship engines in inland waterways and promote the application of more environmentally friendly ship emission reduction technologies.

Conclusion
In this study, the primary focus was on the application of multi-source monitoring data in the inversion and identification of high-emission ships in inland waterways.An algorithm for identifying and correlating peak ship emissions based on multi-source monitoring data was proposed.Building upon this, a method for inverse calculation of ship emission rates and identification of high-emission ships was established using a Gaussian inversion model.Finally, a monitoring sensor platform was developed and implemented for a continuous two-month measurement of passing ships and their pollutant emissions in the Wuhan section of the Yangtze River.The research results indicate that, based on shore-based pollutant monitoring data and ship AIS data, a total of 2841 peaks related to ship emissions were identified.In inland waterways, a significant number of older ships with NO x emissions exceeding standards are still in operation.These findings demonstrate the feasibility of sensor-based shore-based measurements for identifying high-emission ships.It also emphasizes the urgency of improving NO x emissions from inland ships, highlighting the need for further regulation and technological improvements to reduce NO x emissions.
However, this study faces two main shortcomings: the derived ship emission rates are subject to uncertainty from the dispersion model and monitoring data, requiring further validation, and the criteria for identifying high-emission ships need additional examination and refinement.Future efforts could further refine the method presented in this study to enhance its applicability to emerging atmospheric pollutants, such as volatile organic compounds (VOCs), nonmethane VOCs and black carbon.Furthermore, exploring the integration of advanced monitoring technologies and machine learning algorithms could significantly improve the method's precision and real-time monitoring capabilities, enabling more effective identification of high emission sources.

Figure 1 .
Figure 1.The workflow of the method for detecting high-emission ships.

Figure 2 .
Figure 2. Schematic of the location and photographs of the monitoring sensor platform installation.

Figure 3 .
Figure 3. Example and criteria for the association between peaks and ship emissions.

Figure 4 .
Figure 4.An example for the estimation of emission rates based on the comparison between simulated and measured concentrations.

Figure 5 .
Figure 5. Regional ship traffic and meteorological characteristics.(a) Shows the distribution of ship traffic flow by ship type, (b) shows a rose diagram of wind speed and direction.

Figure 6 .
Figure 6.Statistical analysis of identified peak versus (a) ship type, (b) ship direction, (c) ship speed and (d) distance between the ship and the monitoring site.

Figure 7 .
Figure 7.Comparison of the effect of different ship parameters on emission rates.

Figure 8 .
Figure 8. Results of high NOx−emitting ships under different emission limit standards.

Table 1 .
Specifications of monitoring instruments.

Table 2 .
Statistics on emission rates from ships under different classifications.