Hourly Particulate Matter (PM₁₀) Concentration Forecast in Germany Using Extreme Gradient Boosting

Round 1

Reviewer 1 Report

Title: Hourly Particulate Matter (PM10) Concentration Forecast in Germany using Extreme Gradient Boosting

The authors combined the XGBoost algorithm and land use regression model to predict the levels of PM10 in cities, in Germany. They collected hourly PM10 concentration from 400 monitoring stations from 2009 to 2017 to train the models. They choose 9 available variables from COSMO-REA6 to build the model to predict the levels of PM10. There are some major issues to address. I recommend the authors present the details of the method sections.

2.1 study area. As the authors declare, the study collected hourly PM10 concentrations from 400 monitoring stations. The authors should include a description of monitoring station details including locations, sampling periods, etc. Or, the authors could present a map showing the locations of the monitoring stations. The current paragraph is not relevant to this study. In my opinion, the readers would like to know the details of monitoring stations not the general information of Germany.

2.3 Dataset. 

The authors should include a table to show the description of the data summary of measured PM10 in Germany. 

   Line 177. Please show which nine variables are included in the model configuration. I guess the details of the nine variables are shown in Figure 1. Please define some norms including DOY, PBL, PS, LAT, LON, QV, DOW, CLCT, and TOT_PRECIP. Most of the readers are not familiar with the COSMO-REA6.

Line 173. Please give a brief description of COSMO-REA6. What is this model used for? 

In addition, why some variables derived from COSMO-REA6 are not used in the prediction of PM10 using extreme Gradient Boosting? What are the criteria for discarding variables?

Figure 10. How many sampling points are included in the prediction of PM10 obtained from the test dataset in 2018? Why is this test dataset of 2018 selected not the dataset from other years (e.g., 2019, 2020)?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comment

Manuscript “Hourly Particulate Matter (PM10) Concentration Forecast in Germany using Extreme Gradient Boosting” submitted to Atmosphere Stefan Wallek, Marcel Langner, Sebastian Schubert, Raphael Franke, Tobias Sauter is devoted to pressing and serious problems of our time, that is, the problem of air pollution, especially in urban areas. The authors in this study examine point-based forecasting of hourly PM10 concentrations using the XGBoost algorithm to assimilate them into a geostatistical land use regression model for high-resolution spatial and temporal forecast maps. To achieve the goal, the configuration and training of the model included station meteorological data and time variables based on statistical values and expert knowledge. Hourly measurements from approximately 400 stations between 2009 and 2017 were used for training. The paper provides a statistical assessment of the selected model depending on the location of the selected points. The authors note that despite the identified limitations, the model can effectively predict hourly PM10 concentrations. The study area covers the Federal Republic of Germany, offering a wide range of geographical and climatic conditions influencing air pollution dynamics.

The authors have done a tremendous amount of work. The full output of the authors' COSMO-REA6 model includes a set of 150 variables. The authors focused on 35 available 2D parameters and selected 9 of them to use as features to build the model.

The study area covers the Federal Republic of Germany, which offering a wide range of geographical and climatic conditions influencing air pollution dynamics. From the mountainous terrain of the Bavarian Alps to the coastal plains along the North and Baltic Seas, Germany's geography is characterized by a mix of landscapes including forests, agricultural areas, urban centers and industrial areas.

The authors note that the model is not able to accurately reflect situations occurring entirely outside the study area, such as dust emissions from the Sahara or long-range transboundary air pollution from sources such as thermal and coal-fired power plants from neighboring countries, and author note that these impacts can be partially compensate by taking into account the wind direction, which varies depending on the season of the year.

The reviewer believes and hopes that the model developed by the authors will be modernized in the future for into account transboundary long-distance transport of pollution

A huge amount of work has been done and I wanted to clarify some issues.

Figure 9-Figure 11 shows a comparison of the measurement data and the model for PM10 concentration in connection with this

1.  high PM10 concentrations in the presented research results are due to local emissions or long-range transboundary transfer of pollution please explain.

2.What is the reason for the differences in the correlation coefficient in the study areas?

3. When comparing measurement data and model for PM10 concentration in Fig. 12 the model is limited till 50 µg/m3 please explain.

The article is written in an understandable language, not overloaded with unnecessary terminology. The conclusions of the authors are well founded. The article is of scientific interest and is recommended for publication after minor revision.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

accept