This study develops a novel microwave-based method for retrieving aerosol optical depth (AOD) using ground-based radiometer measurements at K- and V-bands. A random forest model trained with daytime sun-photometer data enables continuous day-night AOD monitoring, revealing higher nighttime values. Validation against lunar measurements, radiosonde data, and model simulations confirms the method's reliability. The approach overcomes traditional limitations of nighttime aerosol monitoring, providing new insights into diurnal AOD variations. This practical technique offers valuable applications for air quality assessment and climate studies, particularly for investigating nocturnal aerosol-cloud interactions and radiative effects. The operational simplicity and all-weather capability make it suitable for comprehensive environmental monitoring networks. While the study presents valuable findings, several aspects require further consideration prior to publication.

1. In Figure 2, the channel importance analysis in Figure 2 would be significantly improved by clearly labeling all channel names/frequencies. Additionally, please elaborate on the methodology used to calculate channel importance scores, as this is crucial for interpreting the variable selection process in your random forest model.

2. The criteria for identifying clean sky cases in Figure 5 require more detailed explanation, particularly since this selection directly impacts your nighttime algorithm performance. Given the microwave sensor's limited sensitivity to cloud layers, please clarify: (a) your cloud screening methodology, and (b) how potential cloud contamination was addressed in the analysis. I think this algorithm developed in this study will be the operational algorithm during nighttime, the first step is to determine the clean sky cases.

3. The physical interpretation in section 3.3 would benefit from incorporating established microwave scattering theory. Specifically, please discuss how your findings relate to known scattering and penetration characteristics of microwave channels for different aerosol particles, citing relevant literature (e.g., [reference 1], [reference 2]). This would strengthen the theoretical foundation of your approach.

4. Figure 10 current layout makes data interpretation challenging. I recommend reorganizing it using a 2×2 panel format to: (a) better separate day/night comparisons, (b) improve visualization of temporal trends, and (c) allow direct visual comparison between different measurement types.

This study presents the results of a new and innovative method based on microwave irradiances for retrieving Aerosol Optical Depth (AOD) at VIS/NIR wavelengths (440 nm, 675 nm, 870 nm, 1020 nm, and 550 nm). The method used the BT retrievals of a ground-based mircrowave radiometer (MWR) at K- and V-bands. The study has been led at the Beijing Nanjiao Meteorological Observatory in China, where the MWR instrument was operated.

The method (a machine learning based algorithm) is unfortunately not described. It is only informed, that a random forest model trained with daytime sun-photometer data has been used.

A validation of the results is presented comparing the AOD obtained with the new MWR method to the AOD measured with the photometer of the closest AERONET station (Beijing-CAMS) during day (solar photometry) and night (lunar photometry). The validation experiment has been runned over at least 10 months from December 2019 to October 2020. The validation results shows a good agreement between the new method and the photometry results. 

The new method enables continuous day-night AOD monitoring in cloudless conditions. This is a gain for night measurements compared to the widespread lunar photometry technique that is restricted to the presence of the moon and the moon cycle.

As application, a study over diurnal (24 hours) AOD variations at the MWR site (Beijing Nanjiao Meteorological Observatory) is presented.

The text of the manuscript and the figures and caption is clear, the english is good. It is written in a simple and understainable way. The new method is welcome and innovative. The comparison/validation study shows convincing results. It is a manuscript of good quality showing a very good work

Nevertheless, this paper has a relevance to be published because of the new and innovative method. The paper has a relevance to be published if this method is 1) described in details and precision and 2) if the validation of the method is presented. Unfortunately, even if the validation study is well presented (step 2), the first and most important step (presentation/description of the method) is missing. The method is not descripted but only presented as a black box: "machine learning based retrieval method focusing on the RFR method (Svetnik et al., 2003)". Before acceptance for publication, the section "Retrieval algorithm" has to be considerably extended with precise explanation with algorithm schemes and equations showing how the AOD in VIS/NIR wavelengths is extracted out of the Brightness Temperatures (microwave irradiances) retrieved by the MWR in the K-bands and V-bands. The paper must clearly explain on which atmospheric sciences physical processes between aerosol amount (AOD in VIS/NIR) and radiation in the micowave, the retrieval algorithm is based.

Without this description, in my opinion, this article is not relevant for publication in AMT.

I suggest the authors of the manuscript to work again on it, develop an extended section about "Retrieval algorithm" given all the details about physicall processes mentionned above and to submit again.

In addition of this mandatory change (add the detailled description of the AOD retrieval algorithm), I suggest following major changes before acceptance:

1) In the introduction (Part 1): Increase your knowledges about the cureent reference technique regarding AOD measurements: The photometry using sun photometers during the day and lunar (widespread) or stellar (rare) photometers during the night. Cite relevant papers of these techniques. Be clear how you describe the restrictions of lunar photometry: Only the half of the nights because at least half moon is needed (moon cycles) / mostly only the half of each measurement night because the moon set/rise cycle is not anti-correlated to the sun set/rise cycle.

2) In the method presentation (Part 2): make a description of the sites where you make the measurements. For the three sites (MWR site, AERONET site and Radiosounde site), but of course most of all for the MWR site (Beijing Nanjiao Meteorological Observatory). Make a subparagraph explaining about the sites for each site (coordinates, position regarding the urban area of Beijing, expected aerosol/pollution, rural/urabn/suburban site, generalities about the climate: cold/warm, wet/dry, cloudy/sunny in the different seasons), make a table summarizing the most important data (adress, geocoordinates, distance to MWR site...) and show a map of Beijing Urban area with markers at the place of these three sites.

3) In the validation study (Part 3): You must know that the reference wavelengths for the photometry is 500 nm (for satellite remote sensing, the reference wavelength is 550 nm). When you make a validation study against photometers, you must show results for this channel.

4) For your application study, confrontate your results to the expected results regarding AOD load, and compare it (at least qualitatively) to formewr studies

Do not give up! After this major changes, I would be very please to validate the acceptance of this publication of great interest!

In addition of these major changes expected, see below some minor changes and corrections that you could do along the manuscript to improve its quality.

NB: "L151-156" means for example: "Lines 151 to 156"

Good luck!

GENERALITIES

-> Please add a table with algorithms and parameters, this helps the understanding of the reader

ABSTRACT

L9-11" However, existing remote sensing methods mostly rely on the shortwave spectrum, which does not allow measurements at nighttime"

-> Wrong assumption. Today is lunar photometry very widespread (see hundreds of Aeronet stations... much more than MWR stations!). MWR measurements should be presented as a complement

-> Precise at which site you make this comparisons / measurements ---> Beijing Nanjiao Meteorological Observatory in China?

L17-18: "The algorithm demonstrates satisfactory performance, with strong agreements with lunar AOD retrievals." -> Give some statistical (RMSE or MBE or correlation ...) values to quantify the assumption "strong agreements".

L18-19: The results also reveal a distinct day-night cycle of AOD, with nighttime AOD typically higher than its daytime value" -> In general or for the station considered? Which station/site?

1. Introduction

L31-31:  Please regarding the challenge of extimating aerosol radiative forcing cite the error bars of the last IPCC reports. These are common worldwide references

L34: "optically" -> from its direct radiative impact

L37-34: You first need to explain that AOD is obtained inverting the Beer-Bougher-Lambert equation of spectral direct norml irradiance (DNI) attenuation, usually using a spectrometer or a spectroradiometer making a direct sun irradiance observation most monochromatically on a spectral channel as possible with a filter (photometer) or a narrow band (spectroradiometer).

-> This is the referene equation, method and instrumentation. Then you can describe other methods.

L42-43: "thus only daytime AOD can be obtained" -> This is wrong: There is stellar photometry with star photometers since the 90ies (mention it and cite references) and since 2013 with Cimel CE318T photometer (AERONET network instrument), lunar photometry (cite papers of Barreto et al., also 2013). You can poit out, that stellar photometry is not well widespread (only a few stations worldwide due to bulky facilities and complicate operational process) and that lunar photometry has restriction (only the half of the nights because of the cyclus, and mostly not complete nights because sun set and moon rise are not timely corresponding), but you cannot mention that there is no AOD measurement during he night possible with the standard well proved reference method that is photometry.

L55: OK now you mention lunar photometry. Good but 1) you forgot to mention solar photometry as first and reference method at the beginning of the ontroduction, and 2) you say now that we can monitor AOD during the night with photometers what is refuting your sentence of L42-43. Please restructure introduction: 1) Explain AOD as you did 2) Explain Beer-Lambert-Bougher equation andtalf about photometry: solar photometry (since the 80ies) sunphotometer has to be mention and the prio papers have to been cited. 3) Talk abouth photometry during the night (lunar and stellar) 4) explain the weaknesses/restrictions of lunar and stellar photometry that justify the use of new techniques for instance MWR that you will develop in this paper.

L63-65: Cite also older publications reagrding stellar photometry:

-> HERBER, A., THMASON, L.W., GERNANDT, H., LEITERER, U., NAGEL, D., SCHULZ, K.H., KAPTUR, J., ALBRECHT, T. and NOTHOLT, T., 2002, Continuous day and night aerosol optical depth observations in the Artic between 1991 and 1999. Journal of Geophysical Research, 107, p. 4097

-> LEITERER, U., NAEBERT, A., NAEBERT, T. and ALEKSEEVA, G., 1995, A new star photometer developed for spectral aerosol optical thickness measurements in Lindenberg. Contributions to Atmospheric Physics, 68, pp. 133–141.

L62 "eliminating lunar phase corrections" -> This sentence is confusing, we understand that you correct the lunar radiation what is not the case. Better withdraw it.

L72: Please precise that VIIRS is a satellite bassed instrument (before you were descrbing ground base instruments)

L84-85 if you compare day and night AOD, please mention Grassl et al. 2024 and her study of the Arcticc: Graßl, S., Ritter, C., Wilsch, J., Herrmann, R., Doppler, L., & Román, R. (2024). From Polar Day to Polar Night: A Comprehensive Sun and Star Photometer Study of Trends in Arctic Aerosol Properties in Ny-Ålesund, Svalbard. Remote Sensing, 16(19), 3725.

2. Data and Methods

2.1 Dataset

139-140 "Beijing Nanjiao Meteorological Observatory located in China (39.80°N, 116.47°E)" -> Please give more information about this site o the comparison and about the whol study, with maybe an adress, website, more precise coordinates, and some main reference (main publications or reports in English about this observatory).

Explain also in which area the site of the measurements are (urban, suburban, rural, flat lands, mountains, forests, green fields, ...)

L149 We use the data ranging from December 2019 to October 2020 -> Why so old data?

L151-156: Explain how much data (in proportions) you needed to flag out because of instrumental faults and calibration problems and environmental factors.

L157-159: "Notably, because the collocation between MWR and Level 2 sun photometer AOD products from the AERONET is already clear-sky data, there is no need to perform cloud screening on the MWR data." -> This is most of the time correct but not allways! You can be cloud free in the direction of the moon or the sun for AERONET data but if you look to zenith like the MWR does, there can be some clouds

L160: "Beijing-CAMS AERONET" -> Please precise the coordinates of the site, the distance to the site with the MWR instrument and the environment of th station

-> I guess it is at 26 km distance (regarding AERONET coordinates and coordinates), this is far away in an urban area for a comparison. Why didn't you set the MWR directly at Beijing-CAMS station for the comparative study?

Or at one of the three other AERONET station of Beijing?

At least you should make a spatial heterogeneity study showing how AOD differs from one Beijing AERONET site to the other during the 10 months comparison.

L163: Why don't you use the most common reference -> Please give the coordinates and the distance to AERONET station CAMS Beijing and also to the station of the MWR instrument.

L170: Why do you use ECMWF products, isn't it better directly train the MWR algorithm with the Radiosounde data.

+ Give references for ECMWF ERA-5

L172: "Beijing Meteorological Station": Please precise the coordinates of the site, the distance to the site with the MWR instrument and the environment of the station

Why don't you mention this profile measurements in the Figure 1 or in its caption? Don't you use them for the training?

L180: Give references for MonoRTM

2.2 Retrieval Algorithm

L188-189: "15 km spatial radius"? For ERA-5? in the way it is formulated: Last method/instrument mention is MWR, we understand that MWR has a radius of 15 km. It can only be the model. therefore please reformulate it clearly. By the way, sine the AERONET station and the MWR stations are more than 25 km away, there are not in the same model grid cell if this one has a radius of 15 km and centerred on one station = you obtain different model results for AERONET station and MWR station. How do you deal with this?

L197-198: Why don't you consider the most important (because reference) wavelegth of the AOD: 500 nm?

L196-212 + Figure 1: It is not clear: IS AERONET AOD used for the training or only for verification? Is it the same for day and night?

-> For me here there is not enough pyhisical explanation how the AOD impact the BT and therefore how the BTs measured/retrieved with MWR can be used as proxy to retrieve AOD. Just say "machine learning" and explain inputs/output of a black box is not enough.

L209: Subitely appears here another wavelength for the AOD: 550 nm. This is not a Photometer AOD (even if the AERONET product giuve it as computed value but not measured), this is the common (eg MODIS) satellite AOD channel... But you do not mention any satellite data. Is it the value given by the ECMWF ERA-5 model? Here also, be more precise and give a complete list of the parameter given by ERA-5 that you use.

L213-226 A daytime and a night time graphic with both AOD retrieved with this method (before and after filters/noise smoothing) and AERONET AOD for one day and for one night would be welcome

2.3 WRF-Chem simulations

L234-235: 00:00 UTC (not "0000")

-> Do you really only needed to make 3 Days+nights of WRF Chem simulation and then you have enough representative behaviour to estimate AOD impac on the BTs in order to retrieve AOD from the MWR BT? If yes, this needs to be justified.

I am afraid taht you only have hivernal conditions since you are only simulating 3 days in December.

Figure 3 -> Make an additonal map with a zoom on the region of Beojing and make crosses for the 3 sites (AERONET station, MWR station, radiosounde station).

3. Results

3.1 Model fitting and validation

L294-299 + Figure 4

-> I still do not understand why you do not consider reference wavelength 500 nm.

-> Please explain better what is the train set: When, what (which odel), how, etc... and how it differ to the test set

L298-299: "we will focus on results 299 at 440 nm in the following discussions" -> should be better 500 nm for AOD

Figure 4 + Figure 5: Is in y ("prediction")the MWR invertd product? If yes, write "MWR (prediction)" in y

L305-306: "existing shortwave-based algorithms (Levy et al., 2013)." -> This is a satellite validation explain it, if not, one can believe that it is also a MWR or model validation

L314-315: reference to Figure 6: Explain in the text what is the AOD difference between the lines MWR and Photometer.

L316-337 + figures 7 and 8

-> Again on the figures, if MWR then write "MWR Prediction" instead of only "Prediction"

-> Here also explain the differences between the training set and the test set: When, how, etc...

3.2 The diurnal cycle 

-> Precise again from when to when is this study? Same as test (Dec 2019 - Okt 2020) If yes: Is it representative of this region? It was COVID Lockdown time in urban area, many studies have shown a special Aerosol load in this time. You may consider a longer (many years) time range to make a more relevant study.

-> Also: I suggest you to slice the study in 3 months groups: Summer: JUN-JUL-AUG; Spring: MAR-APR-MAY, Winter: DEC-JAN-FEB and fall: SEP-OCT-NOV, this is surely interesting to look if the mainstream 24 hours evolution is the same in each season.

-> And again: The reference AOD wavelength is 500 nm. Therefore if you want to  show only one wavelength, then 500 nm please.

L351-352 The decrease in AOD and the low number of measurements (photometer night) has no correlation: We can have very few measurements but all of them high AOD

3.3 Physical interpret

Since the machine learning technique does not necessarily represent the physical relationship between aerosol loading and microwace radiances -> This is why YOU MUST HAVE HAD WRITE A COMPLETE PART EXPLAINING HOW YOUR METHOD MAKE THE BRIDGE BETWEEN AEROSOL AND MW RADIANCES

L403-408 and Figure 11: The variability of the temperature is so large, that the vertical profile pictures (a,d,b,e) have not relevance. (c) and (f) on the other side, are very interesting and show a large disagreement between WRF and observations, what should be commented and analyse.

L418-420: "The above observational evidence indicates that MWR estimate AOD by detecting the temperature and humidity profile differences caused by the presence of aerosols, further verifying the theoretical basis of our technique." -> 1) I do not agree maybe you have some season (winter) where Temperature are lower and aerosol different from other seasons are. Therefore you should slice your test analysis and graphics with seasons: JJA, SON, DJF. 2) Just saying that "Aerosol influence the BT observed with MWR" is a very poor description of your algorithm, far away for the expectations of a scientific publication about a new and innovative technique.

L421-441 + Figures 12&13 -> All this study has no relevance if you first do not prrove that aerosol load is not correlated to other events influencing the temperature statistic, like for instance the season. In many cities, in winter because of heating, the aerosol load is higher and the temperature of the atmosphere is lower, not because of aerosol cooling, but just because it is winter, the temperature is lower. On the other side, in some regions, in summer there are becuase of dryness more sand/earth aerosols and the temperature of the atmosphere is higher, not because of aerosol heating effect but because of summer. Therefore, making a one year statistic and saying "the temperature changes: T atmo -> BT change, it is because of the aerosols" is not a serious atmospheric science argumentation.

L442-449: This experiment embetters the poor argumentation that I was critising above. But nevertheless it is not enough. You have to deliver some keys how your algorithm is working: Give equation with the different BT that are used, and explaining how you exteract. Presenting your retrieval algorithm, and showing tests to validate the different steps (in-beetween phyiscal variables) computed by the algorithm.

4. Conclusions and discussions

-> Please mention also in the conclusion some statistical values about the agrrements between model results (WRF-Chem) and observations (MRW and/or AERONET) the same for the validation with comparison AERONET/MWR

-> Please concerning the analysis of AOD, fine and coarse mode, specify that it is specific to one site (Beijing Nanjiao Meteorological Observatory in China) and precise which datetime range? and the conditions during the measurement date range (season, polluted/ not polluted / hot, wet, warm, etc...)

FIGURES:

Figure 1: What do you do with the profile measurements with the radiosound mentionned at Line 172?

Figure 2 (Caption): "MWR Channel" -> "MWR Channels"

Figure 4,5: Why "prediction" and not "MWR"

Figure 5: Why focus on 440 nm? 500 nm is the standard value

Figure 10: Same coment as Figure 5: Why focus on 440 nm? 500 nm is the standard value

Figure 11: AOD at which wavelength?

Figure 12 & 13: Better do it with AOD at 500nm than 550 nm? 500 nm is the reference. 550 nm is only computed but not measured by AERONET.

Figure 14 (caption): Please write the hours HH:MM, example: 20:00 UTC (and not "2000 UTC")

Figure 14: Please on each map, make a cross where the AERONET station is and where the MWR station is