The Effect of Crowdsourced Police Enforcement Data on Traffic Speed: A Case Study of The Netherlands

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The topic presented is interesting and constitutes a contribution to the literature. This study examines the impact of crowdsourced data about police enforcement on traffic speed in the Netherlands. The researchers use navigation apps to extract police enforcement information, collect historical traffic speed data, and develop a deep learning prediction model to predict traffic speed. The paper investigates a very interesting subject, it is well-written and well-structured, and supported by valid experimental data. In my opinion, the article can be published, after considering the following observations:

-In section 2, it is recommended to review the following articles that could provide more background:

-Ryeng, E. O. (2012). The effect of sanctions and police enforcement on drivers’ choice of speed. Accident Analysis & Prevention, 45, 446-454.

-Simpson, R., McCutcheon, M., & Lal, D. (2020). Reducing speeding via inanimate police presence: An evaluation of a police‐directed field study regarding motorist behavior. Criminology & Public Policy, 19(3), 997-1018.

-Simpson, R., Frewing, Q., & Bayer, J. (2023). The Effects of Saturation Enforcement on Speed (ing) Along a Highway Corridor: Results from a Police-Directed Field Study. Justice Evaluation Journal, 6(1), 20-31.

-In section 3.1 it is mentioned: “The adjacency matrix A, which contains elements of 0 and 1, is used to describe the connection between nodes. The element is 1 if the sensors are connected, and 0 otherwise. Traffic speed is treated as a feature of nodes.”. I think it would be better to place after equation 2.

 

-The Conclusions section is written very general and should be extended by adding some details presented in the paper research. Emphasis could also be placed on the practical implications of the accuracy of speed prediction models, such as the GCN-GRU model, which can be affected by the presence of reported police activity. Additionally, it would be nice to know the potential policy implications based on the results.

Author Response

1. Summary

Dear reviewer,

We sincerely appreciate your favorable feedback and express our gratitude! Thank you for allowing us to resubmit a minor revised version of the manuscript. We have incorporated the suggested changes from yours.

Authors 

2. Point-by-point response to Comments and Suggestions for Authors

We appreciate for your valuable consideration. We eagerly anticipate your insights during this review round.

Comment 1: The literatures have been added in the Section 2 as follows:"There are many researchers have delved into the matter that how the police enforcement influences drivers' driving speeds. \cite{RYENG2012446} investigated factors influencing speed choices on rural roads in Norway with an 80 km/h speed limit. It looked at how drivers' perceptions of police enforcement, penalties for speeding, and other drivers' speeds affect their choices. The study found that making most other drivers slow down or increasing law enforcement had the most significant impact on reducing individual speed choices, while stricter penalties had only a minor effect. \cite{simpson2020reducing} examined the impact of police presence on speeding in urban areas using a realistic-looking police cut-out named "Constable Scarecrow" in British Columbia, Canada. The findings showed that deploying the cut-out along major roads helped reduce speeding among motorists. \cite{simpson2023effects} looked at how police saturation enforcement impacts speeding on a highway corridor in Western Canada. They used radar devices at different locations to measure speeds during enforcement and non-enforcement periods. The results showed that police saturation enforcement effectively reduced average vehicle speeds and the proportion of speeding vehicles in the enforcement area, contributing to discussions on policing and road safety. However, there is still a scarcity of relevant research on crowdsourced data for police enforcement."

Comment 2: We have moved the sentences below the Eq.2.

Comment 3: We rewrite the conclusions and add contents as you suggested: "In addition, we find that the report of police activity lessens the performance of the speed prediction model GCN-GRU. Existing models have considered the influence of external factors such as weather and accidents on traffic speed. However, the presence of police enforcement on a road segment also affects the performance of deep-learning prediction models. Therefore, during the model training phase, it is necessary to incorporate the impact of external factors related to the reported police activity. Furthermore, the existence of crowdsourced police activity data allows drivers to be aware of the locations and times of police presence in advance. While this might somewhat diminish the effectiveness of law enforcement, such as reducing the number of traffic citations, it can encourage drivers to slow down in advance, which is also a safety-enhancing measure."

Reviewer 2 Report

Comments and Suggestions for Authors

1.      You may want to provide more information regarding how the police enforcement data was collected and processed e.g. filtering, mapping to road segments etc.

2.      The F-values are listed in all your tables, but not discussed. The authors may want to discuss the purpose of listing the F-values.

3.      On line 249, the authors stated that “However, the actual average speed is 1 [km/h] lower than the predicted average speed during the police enforcement periods, while it is slightly higher when no police enforcement is reported.”  This effect happens in multiple segments. The author may want to add some discussion regarding this effect.

4.      On line 294, the author stated that one of the limitations is that “A limitation of this study is selecting road segments that have both police activity reports as well as speed loop detectors. Most of the reported police enforcement takes place on rural or secondary roads, where few or no speed sensors and consequently speed data exist to conduct such investigation” The authors may want to discuss how they plan to address this limitation in the future.

5.      In the conclusion section, the authors may want to elaborate on the practical implication of the results for both traffic management and enforcement.

Author Response

1. Summary

Dear reviewer,

We sincerely appreciate your favorable feedback and express our gratitude! Thank you for allowing us to resubmit a minor revised version of the manuscript. We have incorporated the suggested changes from yours.

Authors 

2. Point-by-point response to Comments and Suggestions for Authors

We appreciate for your valuable consideration. We eagerly anticipate your insights during this review round.

Comment 1: The description has been added in the Section 4 as follows: Erroneous reports were removed, and reported police activities were accurately mapped to specific road segments using their location information.

Comment 2: We have added the analysis in Section 4: "Furthermore, all the results consistently demonstrate that in the absence of police, the F-score exhibits a higher value, while in the presence of police, the F-score is generally much lower. This indicates that when police are present, the model's performance tends to decline."

Comment 3: We discuss this results in Conclusion 5 and add the content as follows:"In addition, we find that the report of police activity lessens the performance of the speed prediction model GCN-GRU. Existing models have considered the influence of external factors such as weather and accidents on traffic speed. However, the presence of police enforcement on a road segment also affects the performance of deep-learning prediction models. Therefore, during the model training phase, it is necessary to incorporate the impact of external factors related to the reported police activity".

Comment 4: This limitation can be addressed by collecting more crowdsourced police enforcement data. So we can obtain more road segments that include enough reported police activity data. The contents has been added in Conclusion "For a future study,  we plan to collect more crowdsourced police enforcement data".

Comment 5:  We elaborate on the practical implication of the results for both traffic management and enforcement in Section 5:" Furthermore, the existence of crowdsourced police activity data allows drivers to be aware of the locations and times of police presence in advance. While this might somewhat diminish the effectiveness of law enforcement, such as reducing the number of traffic citations, it can encourage drivers to slow down in advance, which is also a safety-enhancing measure."

Reviewer 3 Report

Comments and Suggestions for Authors

Aims:

The goal of this academic paper is to use a Graph Convolutional Network (GCN)-Gated Recurrent Unit (GRU) model to predict traffic speed in various scenarios, specifically when there is or is not police enforcement. This prediction model's performance is measured using three key indicators: mean absolute error (MAE), root mean square error (RMSE), and accuracy (ACC). The paper intends to determine the effectiveness of the prediction model by comparing the predicted speed to the actual speed. The model's utility is demonstrated through case studies involving two different highways.

 

The main research question:

How does the driver's response to police enforcement, as determined by navigation apps, affect driving behavior and overall traffic speed on the road network, and how can we accurately predict this driving speed? The researchers intend to answer this question by employing a Graph Convolutional-Gated Recurrent Network (GCN-GRU) model for speed prediction and analyzing crowdsourced data from Dutch police enforcement.

 

Methods:

For traffic speed prediction, the paper employs a prediction model based on Graph Convolutional Network (GCN) and Gated Recurrent Unit (GRU). This model is tested in a variety of scenarios, including periods with and without police enforcement.

Three evaluation indicators are used to assess the model's performance:

1. Mean Absolute Error (MAE) - calculates the average magnitude of errors in prediction results, regardless of direction.

2. Root Mean Square Error (RMSE) - a quadratic scoring rule that calculates the average magnitude of an error. It is the square root of the average of squared differences between predicted and actual values, with larger errors penalized more severely.

3. Accuracy (ACC) - measures the similarity of predicted and actual values. The greater the ACC, the more accurate the prediction model.

A one-factor Analysis of Variance (ANOVA) test is used to determine whether there are any statistically significant differences between the means of different groups. The groups in this case refer to periods with and without police enforcement.

 

Conclusions:

The paper concludes that the Graph Convolutional Network (GCN) and Gated Recurrent Unit (GRU) models can effectively predict traffic speed in different conditions, specifically when police enforcement is present or not. 

This model's performance was assessed using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Accuracy (ACC). These performance indices demonstrated the model's ability to calculate the difference between predicted and actual speeds. 

The statistical analysis using the one-factor ANOVA test confirmed the model's effectiveness. It discovered that the differences between the actual and predicted speed means were not statistically significant, indicating that the model can predict traffic speed accurately.

This model's ability to predict traffic speed has potential applications in traffic management and planning. It is important to note, however, that the prediction efficacy is dependent on the conditions and situations at hand, such as the presence or absence of police enforcement.

 

Added value:

The paper claims to provide a model that can better capture the topology features of transportation networks and efficiently forecast traffic speeds by treating the road network as a graph and integrating the GRU for temporal data characteristics. While other deep learning models have been used to predict traffic, the combination of GCN and GRU proposed in this paper provides an advanced alternative that takes both the topological and temporal characteristics of traffic data into account.

This approach adds a fresh perspective to academic discussions about traffic prediction, with the potential to accelerate the development of more effective, precise prediction models based on real-world data structure. It paves the way for future research into other complex data prediction challenges using similar architectures.

 

Main deficiencies:

1. Limited dataset: If the prediction model is only tested on a small or specific set of data, the results may not generalize to larger or more diverse scenarios.

2. Lack of comparison: If the GCN-GRU model's effectiveness was not compared to other established traffic prediction models, it may be difficult to claim superiority or improvements.

3. Real-world application: Concerns may be raised about the feasibility of implementing this model in real-time traffic prediction due to factors such as computational cost or the requirement of large-scale high-quality traffic data, which may not be readily available.

4. Contextual factors: While the model described in this paper addresses intrinsic data factors such as the road network, it may overlook extrinsic dynamic factors such as weather, social events, or variable human behavior, which can all have a significant impact on traffic patterns. 

 

Questions:

1. What influenced your decision to use the Graph Convolutional-Gated Recurrent Network (GCN-GRU) as the model for your research? Why not use one of the existing models?

2. Has your model been tested on a variety of dataset scenarios to ensure that it is not overfitted and that it can generalize well?

3. How practical is your model's implementation in real-time traffic management in terms of computational efficiency and data requirements?

4. How would your model handle dynamic changes in traffic patterns caused by extrinsic factors such as weather, local events, or other uncontrollable factors?

5. Does your research take into account the effect of various types and intensities of police enforcement strategies on driving behavior and overall traffic speed?

6. Can your model predict traffic speed effectively in road networks of varying complexity, or is it optimized only for the selected road networks?

7. Are there any limitations to your study that could be addressed in future research to improve traffic speed prediction?

8. How would the model perform in different geographical locations, given that driver behaviors vary greatly depending on culture or prevalent driving rules?

9. Have you considered including a comparison group or control group in your experimental design to strengthen your findings and control for potential outside influences? 

10. Can you explain how you chose your method for evaluating prediction performance? Why were these methods chosen over other potentially viable alternatives?

 

Other comments:

1. The paper does not meet the MDPI journals' template. Please reformat it according to the requirements!

2. The format of citation in the main text does not fit to the required format in MDPI journals. Please reformat all of them according to the prescriptions!

3. Please do not use nouns and sentences with "we", "I", etc. Scientific journal papers mustn't contain this kind of writing in an academic style! Please eliminate all of them from the manuscript and use passive sentences, or apply the noun "the authors" instead of "we", "I", etc.

4. The references have to be supplemented by more, relevant, up-to-date (mainly) journal papers from the time interval 2020-2023 (2024).

5. Please prepare a "Nomenclature" at the end of the paper that contains all the notations/parameters with a short text explanation and the applied units in/according to the SI system.

Comments on the Quality of English Language

Minor editing of English language required

Author Response

1.Summary

Dear reviewer,

We sincerely appreciate your feedback and express our gratitude! Thank you for allowing us to resubmit a revised version of the manuscript. We have incorporated the suggested changes from the reviewers and have further refined the manuscript. We eagerly anticipate your insights during this review round.

Authors

2. Point-by-point response to Comments and Suggestions for Authors

(1) We revise the manuscript as the reviewer's suggestions:

• The nouns and sentences with "we", "I", etc. have been corrected.
• We add literatures in Section 2:"There are many researchers have delved into the matter that how the police enforcement influences drivers' driving speeds. \cite{RYENG2012446} investigated factors influencing speed choices on rural roads in Norway with an 80 km/h speed limit. It looked at how drivers' perceptions of police enforcement, penalties for speeding, and other drivers' speeds affect their choices. The study found that making most other drivers slow down or increasing law enforcement had the most significant impact on reducing individual speed choices, while stricter penalties had only a minor effect. \cite{simpson2020reducing} examined the impact of police presence on speeding in urban areas using a realistic-looking police cut-out named "Constable Scarecrow" in British Columbia, Canada. The findings showed that deploying the cut-out along major roads helped reduce speeding among motorists. \cite{simpson2023effects} looked at how police saturation enforcement impacts speeding on a highway corridor in Western Canada. They used radar devices at different locations to measure speeds during enforcement and non-enforcement periods. The results showed that police saturation enforcement effectively reduced average vehicle speeds and the proportion of speeding vehicles in the enforcement area, contributing to discussions on policing and road safety. However, there is still a scarcity of relevant research on crowdsourced data for police enforcement."
• We add contents in the conclusion as you suggested:"Existing models have considered the influence of external factors such as weather and accidents on traffic speed. However, the presence of police enforcement on a road segment also affects the performance of deep-learning prediction models. Therefore, during the model training phase, it is necessary to incorporate the impact of external factors related to the reported police activity. 
  Furthermore, the existence of crowdsourced police activity data allows drivers to be aware of the locations and times of police presence in advance. While this might somewhat diminish the effectiveness of law enforcement, such as reducing the number of traffic citations, it can encourage drivers to slow down in advance, which is also a safety-enhancing measure."

(2) Answers to the review's questions: 

1. 1. GCN-GRU serves as the foundational framework for numerous deep learning prediction models, with many models being variants built upon it. This is why we have chosen to use it. This model is, in fact, an existing model proposed by Zhao in 2019 and is known as T-GCN.

   2. Yes, this model has already been tested. Furthermore, you can also review its literature: L. Zhao et al., "T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction, IEEE Transactions on Intelligent Transportation Systems, Vol. 21, No. 9, pp. 3848-3858, September 2020, DOI: 10.1109/TITS.2019.2935152.

   3. We aim to investigate whether police enforcement crowdsourced data affects model performance. The conclusion we have drawn is that it does indeed have an impact, which in itself holds significant practical significance.
   4. Thank you for your question. We did not consider this factor because our primary objective was to compare speed changes during a similar time period with and without enforcement, making it the sole variable.We’ll investgate the influence of other external factors on our conclusions in the furture research.

   5. Thank you for your curiosity. This is an interesting question, but due to the limited scope of our current data, we were unable to address this issue in this study. However, in future research, we plan to collect more data and investigate this further.

   6. This question awaits further investigation into the model, primarily from an algorithmic perspective. We will embark on subsequent studies to delve deeper into this aspect.

   7. Existing models have considered the influence of external factors such as weather and accidents on traffic speed. However, the presence of police enforcement on a road segment also affects the performance of deep-learning prediction models. Therefore, during the model training phase, it is necessary to incorporate the impact of external factors related to the reported police activity.
   8. This is a profoundly intriguing question. In this study, our data collection was limited to enforcement data from the Netherlands. We have indeed contemplated the idea of contrasting this with data from other countries or cultures, like in China, where on-site enforcement primarily targets drunk driving rather than speeding. This question demands further investigation, and we are committed to obtaining additional data for a more comprehensive analysis.
   9. Thank you for your question. We did not include a baseline because our primary objective was to investigate whether enforcement affects the performance of GCN-GRU. Therefore, 'with police' and 'without police' are the sole variables, and the model's inherent predictive accuracy does not influence our conclusions.
   10. MAE and RMSE are classic indicators used in traffic flow prediction models. In addition, we also conducted a statistical comparison of the results using F-score and P-value. We believe this approach provides a comprehensive assessment of our study.