General comments:

The authors have developed a machine learning-based ATMS snowfall retrieval algorithm specifically for the high latitude environment. The main concept of the algorithm is to isolate snowfall signature in ATMS measurements by comparing observed brightness temperatures (TBs) with the corresponding simulated TBs under clear-sky conditions. The snowfall signature, expressed as the difference between the observed and simulated TBs, is important predictor in the ANN models for the detection and estimation of snowfall. Central to the retrieval of clear-sky TBs is the estimation of surface emissivity in high latitudes, including for several classes of sea ice and snow cover. The ATMS low frequencies are assumed to be less affected by cloud and are utilized to derive surface emissivity along with some model and ancillary variables. While it is not a new concept to derive snowfall/scattering signature from observations and simulated clear-sky measurements (see specific comments), the authors have conducted valuable research and employed machine learning techniques effectively to produce a viable algorithm for retrieving high-latitude snowfall. The validation study on the SWP and SSR retrievals demonstrate that the algorithm performs well within its designed limits, including reasonable performance under very cold and dry conditions. There are aspects of the algorithm that still require further developments, notably, more accurate estimation of surface emissivity for the retrieval of clear-sky TBs. However, the manuscript has sufficient merit to be considered for publication once the authors address the following comments.        

Specific comments

• Line 27: This approach has been used before so it’s not accurate to call it innovative. Zhao and Weng (2002, http://www.jstor.org/stable/26184983) retrieved ice cloud parameters by isolating ice scattering signature. The latter is derived from observed high frequency TBs and simulated cloud base (i.e. clear-sky) TBs. They calculated the over land cloud base high frequency TBs from low frequencies with the assumption that low frequency measurements are less affected by cloud scattering. Please modify the manuscript accordingly and cite Zhao and Weng’s paper.
• Line 89: Contrary to what’s stated here, Greenland and Antarctica show scattering year-round in window and water vapor sounding channels, and even in the low temperature sounding channels.
• Lines 116-119: While 2CSP is a well-recognized product and is not derived from radiative transfer modeling, it does include assumptions about snow microphysics, and uses optimal estimation to retrieve these parameters. The algorithm also uses a simplified radar reflectivity equation. Refer to the 2CSP ATBD at https://www.cloudsat.cira.colostate.edu/cloudsat-static/info/dl/2c-snow-profile/2C-SNOW-PROFILE_PDICD.P1_R05.rev0_.pdf. Please modify the text here accordingly.
• Line 181: How is the underestimation of heavy snowfall handled in training and validating the SWP and SSR models?
• Lines 191-192: Add the info on the dataset’s geographic area. Was the data filtered for high latitudes given the focus of this study?
• Lines 193-194: With a 15-min time window, the snow mass that ATMS detects in the atmosphere most likely is higher than the near-surface snow (SSR) observed by CPR (refer to You et al., doi: 10.1029/2019GL083426). This adds uncertainties to the SSR (and to a lesser degree to SWP). Suggest the authors run an experiment where ATMS data is collocated with CPR snowfall rate with a certain time lag (30-minute?), and compare the retrieved ATMS snowfall rate with what is presented in this manuscript.
• Line 273: Do the ANNs use environmental parameters? What are they?
• Line 282: Is there any noticeable discontinuity in the retrieved SWP and SSR between the different surface classes? Please add some discussion in the appropriate section.
• Line 290: While this is outside the scope of this study, is it possible to improve snow cover classification using ML approach? I’d like to get the authors’ comments on it.
• Line 350: Is the polarization effect on emissivity also neglected between viewing angles of 40 degree and 52.7 degree (the max ATMS viewing angle)? Need to state it if it’s the case.
• Line 403: Since high frequencies are more important for snowfall retrieval, need to discuss the impact of the significant uncertainties at these channels to retrieved SWP and SSR.
• Line 430: Logarithmic tangent function is not a common activation function. Please add a reference or explain what it is.
• Lines 435-436: Did the predictor set including TB_obs, TB_obs-TB_sim, and environmental variables give better result than the set only included the first two? If not, why? Is it because TB_sim also used the environmental variables being tested?
• Lines 444: Which 16 ATMS channels and how are they selected?
• Section 4.1: Some details about the validation data should be provided. Is the data from selected snowfall events used or from a time period? How many events were included and their geographic areas? How many data points were in the dataset etc.? The information is important because it provides the context for the performance metrics.
• Line 451: A large percentage of the snowfall appears to fall when T_2m is around the freezing point or higher. Snowfall under such conditions generally has different characteristics from snowfall in high latitudes which is the focus of this study. Add some discussion about the data distribution and its impact on the new snowfall algorithm.
• Line 471: Add HSS to Table 6.
• Table 5: Since the goal of this study is to retrieve snowfall in high latitude, it'd be informative to analyze how well the statistics represent the cold, dry and light snowfall verses the warm, moist, and heavier snowfall. Please add some quantitative analysis to show the performance of the snowfall representative of high latitude conditions.
• Line 487: Typically, high latitude snowfall is rather light. Does this result mean that the snowfall retrieval in high latitude is generally overestimated? Add some discussion here.
• Lines 555-558: See the comment on line 27.

Technical corrections

• Line 67: ‘last’, suggest to change to ‘new’ or ‘latest’. Also on line 137.
• Line 283: AutoSnow is a NOAA product.
• Line 327: give explicit definitions of POD, FAR, and HSS even though they are well known.
• Line 346: Give reference to the radiative transfer model, or add some information about the model.
• Line 362: Reference for the RTM?
• Line 397, the RMSE for ocean is 3.37 K in Table 2.

The High lAtitude sNowfall  Detection and Estimation aLgorithm for ATMS (HANDEL-ATMS): a new algorithm for snowfall retrieval at high latitudes

By Camplani et al., 

In this paper, a snowfall retrieval algorithm, the High lAtitude sNow Detection and Estimation aLgorithm for ATMS (HANDEL-ATMS), is described. This algorithm uses three years of coincidences of ATMS and CPR to retrieve snow water path (SWP), and surface snowfall rate (SSR) over high latitudes. The innovative principle in the algorithm development is the exploitation of the full range of ATMS channel frequencies to characterize the frozen background surface radiative properties at the time of the overpass to be able to better isolate and interpret the snowfall-related contribution to the measured multi-channel upwelling radiation.

The algorithm is a two-step algorithm that first detects snowfall and then estimates its rate. In the preprocessing step, the algorithm attempts to add features to the network by calculating the difference between actual TB observed by ATMS and the simulated TB over different surface types (that are labeled using previously published work by the co-authors) using the emissivity spectra of each surface type.

General comments. 

The text is a bit hard to follow. It is highly recommended that the authors make an effort to shorten it and make the language and the message more succinct. The quality of the figures can be significantly improved as well. There are a few important points that need to be cleared in the next revision. 

• The explanation of the inverse radiative transfer modeling is missing. Such an inversion can be significantly underconstrained and add additional uncertainty to the results.
• Please clarify upfront whether the estimated values of surface emissivities are used dynamically or statistically in the algorithm. Do they change in time or not?
• It will be helpful if the authors clarify why we need land surface classification for the algorithm. For example, there are multiple products for the detection of the presence of snow and sea ice dynamics using optical bands (every 30 minutes). These optical products can be more accurate than microwave classification schemes, in terms of the presence or absence of frozen surfaces. Why we should not use them? 
• From a methodological standpoint, the explanations of neural networks need to be improved. A the same time, the use of linear discriminant analysis seems outdated in light of the new deep-learning classification models. 
• While the paper focuses on different land surface types and sea ice ages, it is unclear how statistically significant the presented results are in Table 7. The number of training and testing samples needs to be clarified. 
• It would benefit the paper if the authors provide the entire confusion matrix of the detection of snowfall, including, recall, precision, and accuracy. 

Detail comments:

• Section 2.4 is long and has some generic explanations about for example neutral networks, which is not necessary at this time. It is recommended to shorten the text.
• The explanation of the architecture of the neural network is weak. First of all the networks use the Levenberg-Marquardt algorithm which is extremely old and is not being used in modern training of deep neural networks. Unlike algorithms like Adam, it is prone to get stuck in local minima and suffer from the vanishing gradient problem. 
• Line 424–445 It is unclear how the detection and estimation networks are implemented. What are the cost functions? This must be clarified. 
• Line 345-346: It is not well-described how the inverse radiative transfer model is used. What is the forward RT model?
• Lines 362-365: How emissivity is used for calculating the simulated TBs? It seems recursive to use the observations to estimate the emissivity and then use it for retrievals. Please clarify whether the used emissivities are dynamic or static. 
• Table 3: The parameters mentioned in the table are different than the ones mentioned in the text in lines 435-437.

Minor comments:

Line 273: It is better to mention all the variables that have been used for training the network here.

line 203-204: list of environmental and ancillary parameters is not presented in the dataset.

Line 356: “…for ocean and land respectively.”

Line 387: What is the used atmospheric radiative transfer model? Please spell out RTM.

Table 2: What is the accuracy represented here? The accuracy of PESCA for surface classification?

Line 489: Remove the dot at the beginning of the sentence.

Figure 1: The inputs of PESCA mentioned in this figure are not aligned with the original paper. For example, there exists no explanation for the low-frequency ratio and scattering coefficients.

Figure 6: No results are presented over sea ice.

Figure 10: Please mention that the shown green dots denote the CPR overpass.