Diagnosis of Antarctic Blowing Snow Properties Using MERRA-2 Reanalysis with a Machine Learning Model

Yuekui Yang aClimate and Radiation Laboratory, NASA Goddard Space Flight Center, Greenbelt, Maryland

Search for other papers by Yuekui Yang in
Current site
Google Scholar
PubMed
Close
https://orcid.org/0000-0002-1630-272X
,
Daniel Kiv bDepartment of Computer Science, University of Illinois Urbana–Champaign, Urbana, Illinois
cNASA Goddard Space Flight Center, Greenbelt, Maryland

Search for other papers by Daniel Kiv in
Current site
Google Scholar
PubMed
Close
,
Surendra Bhatta aClimate and Radiation Laboratory, NASA Goddard Space Flight Center, Greenbelt, Maryland
dGoddard Earth Sciences Technology and Research II, Morgan State University, Baltimore, Maryland

Search for other papers by Surendra Bhatta in
Current site
Google Scholar
PubMed
Close
,
Manisha Ganeshan aClimate and Radiation Laboratory, NASA Goddard Space Flight Center, Greenbelt, Maryland
dGoddard Earth Sciences Technology and Research II, Morgan State University, Baltimore, Maryland

Search for other papers by Manisha Ganeshan in
Current site
Google Scholar
PubMed
Close
,
Xiaomei Lu eLidar Science Branch, NASA Langley Research Center, Hampton, Virginia

Search for other papers by Xiaomei Lu in
Current site
Google Scholar
PubMed
Close
, and
Stephen Palm cNASA Goddard Space Flight Center, Greenbelt, Maryland
fScience Systems and Applications, Inc., Lanham, Maryland

Search for other papers by Stephen Palm in
Current site
Google Scholar
PubMed
Close
Open access

Abstract

This paper presents work using a machine learning model to diagnose Antarctic blowing snow (BLSN) properties with the Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2), data. We adopt the random forest classifier for BLSN identification and the random forest regressor for BLSN optical depth and height diagnosis. BLSN properties observed from the Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) are used as the truth for training the model. Using MERRA-2 fields such as snow age, surface elevation and pressure, temperature, specific humidity, and temperature gradient at the 2-m level, and wind speed at the 10-m level as input, reasonable results are achieved. Hourly blowing snow property diagnostics are generated with the trained model. Using 2010 as an example, it is shown that the Antarctic BLSN frequency is much higher over East than West Antarctica. High-frequency months are from April to September, during which BLSN frequency exceeds 20% over East Antarctica. For May 2010, the BLSN snow frequency in the region is as high as 37%. Due to the suppression by strong surface-based inversions, larger values of BLSN height and optical depth are usually limited to the coastal regions, wherein the strength of surface-based inversions is weaker.

For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Yuekui Yang, yuekui.yang@nasa.gov

Abstract

This paper presents work using a machine learning model to diagnose Antarctic blowing snow (BLSN) properties with the Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2), data. We adopt the random forest classifier for BLSN identification and the random forest regressor for BLSN optical depth and height diagnosis. BLSN properties observed from the Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) are used as the truth for training the model. Using MERRA-2 fields such as snow age, surface elevation and pressure, temperature, specific humidity, and temperature gradient at the 2-m level, and wind speed at the 10-m level as input, reasonable results are achieved. Hourly blowing snow property diagnostics are generated with the trained model. Using 2010 as an example, it is shown that the Antarctic BLSN frequency is much higher over East than West Antarctica. High-frequency months are from April to September, during which BLSN frequency exceeds 20% over East Antarctica. For May 2010, the BLSN snow frequency in the region is as high as 37%. Due to the suppression by strong surface-based inversions, larger values of BLSN height and optical depth are usually limited to the coastal regions, wherein the strength of surface-based inversions is weaker.

For information regarding reuse of this content and general copyright information, consult the AMS Copyright Policy (www.ametsoc.org/PUBSReuseLicenses).

Corresponding author: Yuekui Yang, yuekui.yang@nasa.gov

1. Introduction

It has been demonstrated by many studies that blowing snow (BLSN) plays a significant role in the Antarctic climate system. The impacts of BLSN include its effect on surface mass balance (e.g., Dingle and Radok 1961; Schmidt 1982; Scarchilli et al. 2009; Déry and Yau 2002; Lenaerts et al. 2012; Das et al. 2013; Palm et al. 2017) and on water vapor and radiation budget (Yamanouchi and Kawaguchi 1985; Barral et al. 2014; Yang et al. 2014; Palm et al. 2018b). The interaction between BLSN and lidar pulses also affects the measurements of spaceborne lidar altimeters, such as the Geoscience Laser Altimeter System (GLAS) on board the Ice, Cloud, and Land Elevation Satellite (ICESat) and the Advanced Topographic Laser Altimeter System (ATLAS) on board ICESat-2 (Duda et al. 2001; Yang et al. 2010, 2011).

Because of the remoteness and harsh conditions over the Antarctic ice sheet, ground-based observations of meteorological parameters, including BLSN, are very limited. Satellites remain the only feasible way to observe BLSN occurrences and properties on the continental scale. Using data from the Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observation (CALIPSO) mission, Palm et al. (2011) developed a method for BLSN detection. Results from the application of this method are provided as the CALIPSO Lidar Level 2 Antarctic BLSN product available at the NASA Langley Atmospheric Science Data Center (see the data availability statement below for data access information). Studies conducted using the product have shed light on the BLSN properties and its impacts on the local climate (e.g., Yang et al. 2014; Palm et al. 2017, 2018a; Ganeshan et al. 2022). Using these CALIPSO BLSN retrievals as truth for training, Yang et al. (2021) developed a machine learning model to study Antarctic BLSN storms (≥1000 km2 in area) with observations from the Moderate Resolution Imaging Spectroradiometer (MODIS) on board the Aqua satellite. However, both the CALIPSO and the MODIS BLSN detection methods have limitations. For MODIS, the Yang et al. (2021) method is only applicable to daytime observations, because a large portion of the information content is in the MODIS reflective channels; this is an issue since BLSN happens most frequently during the Antarctic winter months, when there is lack of sunlight. For CALIPSO, the observations are limited to a width of one pixel (∼333 m) and the satellite surface tracks only reach 82°S due the orbit inclination. However, we note that the Palm et al. (2011) method has also been successfully applied to ICESat-2 observations (Palm et al. 2021), which reach up to 88°S and provide a new spaceborne-lidar-based data record for continuous BLSN studies.

To better facilitate future investigations of the role of BLSN on the Antarctic surface mass balance and on the local climate system, it is desirable to have a gapless BLSN product. Such a product can be generated using numerical modeling based on physical parameterizations of BLSN (e.g., Déry and Yau 1999; Xiao et al. 2000; Lehning and Fierz 2008; Lenaerts et al. 2012; Wever et al. 2023). However, due to the usually simplified BLSN initiation assumptions, direct numerical model predictions may differ from observations. This paper adopts a different approach. Antarctic BLSN properties are diagnosed using NASA’s Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2), fields as input. A random forest–based machine learning model is developed with observations from CALIPSO for training, use of which achieves reasonable results. We note that the quality of the input data is a determining factor for the results. It has been shown that MERRRA-2 represents the Antarctic meteorological parameters skillfully, including wind, temperature, temperature gradient, and precipitation (Ganeshan and Yang 2019; Palerme et al. 2017). The remainder of the paper is organized as follows: section 2 describes the data sources and the machine learning model development. Model performance and results are shown in section 3; a summary and discussion are provided in section 4.

2. Data and method

a. Data

Four datasets are used in this study, all of which are publicly available (see the data availability statement below for data access link). The first is the CALIPSO BLSN product (Palm et al. 2011), which contains information on BLSN occurrence, height, and optical depth. This is the main source of the “truth” values for training the machine learning models. The second dataset is the CALIPSO Lidar Level 2 cloud-layer product, version 4.20 (Winker et al. 2009; Liu et al. 2019), which is used as a supplement to the truth. This product can provide information on whether the lidar signal is fully attenuated by clouds. If it is, then whether BLSN exists or not is unknown. We combine these two datasets to reclassify each data point as BLSN, non-BLSN, and unknown, following the procedure given in the next section. We note that even though we believe the CALIPSO BLSN product is well suited for training the machine learning models, and serves as the truth values, it is a remote sensing product derived from CALIPSO lidar backscatter measurements, which have uncertainties as all remote sensing products do (Palm et al. 2011). The third dataset used in the study is the Aqua MODIS Level 1b Collection 6.1 calibrated radiances (Toller et al. 2013), which are used for case analysis and for verification purposes.

The last dataset is the MERRA-2 reanalysis developed at the NASA Goddard Space Flight Center (Gelaro et al. 2017). MERRA-2 is based on the model run of the Goddard Earth Observing System, version 5 (GEOS-5) data assimilation system and includes two- and three-dimensional products. The data fields are at 0.5° latitude by 0.625° longitude resolution. For the three-dimensional products, model fields are given on 72 model levels. The two-dimensional products have hourly time intervals, and the three-dimensional fields have three-hourly intervals (Bosilovich et al. 2016). For this work, both products are explored to investigate the best parameter set for BLSN diagnosis with the machine learning model.

b. Machine learning model for BLSN diagnosis

The initiation and development of BLSN events can be affected by many factors, such as wind, temperature, pressure, and status of the surface snow. Studies have shown that machine learning models are efficient in combining information from numerous sources and making reasonable predictions (e.g., Jeppesen et al. 2019; Wang et al. 2020; Yang et al. 2021). For this study, we adopt the widely used random forest algorithms (e.g., Pedregosa et al. 2011; Cutler et al. 2012), which can handle both classification and regression. The random forest classifier is used for BLSN occurrence prediction and the random forest regressor for BLSN optical depth and height prediction.

Random forest is a supervised machine learning algorithm. Unlike conventional single-tree methods, the random forest algorithm builds an ensemble of decision trees (a forest) to reduce variance and boost prediction performance. A classifier makes the prediction based on the votes of all the trees. A random forest regressor fits the decision trees on subsamples of the dataset and uses averaging for the final prediction. For this work, the random forest algorithms are implemented using the Python-based Scikit-learn software package (Pedregosa et al. 2011). The steps of implementation include splitting the data into training and testing parts, creating the model, fitting the model, making a prediction, and model evaluation. To train the model, 80% of the training dataset is used and the other 20% is reserved for performance test.

Figure 1 gives the flowchart of the machine learning model for BLSN property diagnosis using MERRA-2 and CALIPSO data. As the first step, machine learning models need to be trained on a dataset of 〈input, truth〉 pairs. For this study, training datasets consist of the collocated MERRA-2 meteorological fields, that is, the “input,” and the “truth,” namely, BLSN properties from CALIPSO observations, including BLSN occurrence, optical depth, and height. To investigate the best MERRA-2 variable set for the machine learning models, we start with the 28 parameters given in Table 1. These variables included all essential information from the surface to the lowest four model levels. Except for the snow age (SnowAge) and the 2-m temperature gradient (TempGr), all others can be directly read from the MERRA-2 products. The snow age is derived as the number of days since last snowfall with precipitation amount of more than 0.5 mm day−1. The TempGr is calculated as
TempGr=(T71T2M)(h712),
where T71 and T2M are the temperatures at the lowest model level and at 2 m, and h71 is the height of the lowest model level above ground (model layer count starts from zero at the top). (We call the lowest model level the first model level hereinafter.) In the MERRA-2, for a nominal surface pressure of 1000 hPa, the height of the first model level corresponds to around 60 m above the surface and the height of the second model level corresponds to around 180 m above the surface. Over the Antarctic continent, these values are lower as the surface pressure remains below 1000 hPa over the continent during most times. The typical height range for the first model level is from 45 to 60 m and for the second model level is from 120 to 170 m. The values are lower in East Antarctica than in West Antarctica because of the significant elevation and lower atmospheric pressure of the plateau.
Fig. 1.
Fig. 1.

Flowchart for applying the random forest classifier and regressor to BLSN diagnosis with MERRA-2 data.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

Table 1.

MERRA-2 fields tested for best machine learning model performance. The eight parameters in boldface type are consistently ranked high in feature importance.

Table 1.

As mentioned above, based on the CALIPSO Lidar Level 2 cloud layer product and the CALIPSO BLSN product, a CALIPSO pixel (1-km resolution) is classified as BLSN, non-BLSN, or unknown. Non-BLSN means that surface is detected but no BLSN is found, while “unknown” means that surface signal is not detected; hence, we have no knowledge whether BLSN is present. Note that the CALIPSO lidar can penetrate optically thin clouds (usually optical depth <3) and possibly detect the BLSN underneath, but if clouds are optically thick (usually optical depth ≥3), then the lidar signal is fully attenuated, resulting in no surface return.

The process described above classifies each 1-km-sized pixel. To build the collocated MERRA-2–CALIPSO dataset, we group the CALIPSO pixels into 50-km segments and once again classify each segment into BLSN, non-BLSN, and unknown categories, similar to the approach used for classifying individual pixels. The classification of each segment is based on the following rules: 1) if ≥50% of the CALIPSO pixels within the segment are BLSN, then label it “BLSN”; 2) if ≥80% of the CALIPSO pixels within the segment are non-BLSN, then label it “non-BLSN”; 3) the rest are labeled “unknown.” For segments classified as BLSN, the average height and optical depth are recorded.

The 28 MERRA-2 fields used as input come from both two- and three-dimensional data products. For example, the 20 variables listed in the last three rows of Table 1 are all from three-dimensional data files. Since the time interval for the three-dimensional fields is 3 h, as compared with only 1 h for the two-dimensional fields, we first interpolate the variables selected from the three-dimensional data collection onto a 1-h time interval. Following this, the nearest-neighbor approach is used to collocate each CALIPSO segment with a corresponding MERRA-2 grid point. Only BLSN and non-BLSN data are used for the machine learning model training; the unknown category is not used.

After preprocessing the 〈input, truth〉 dataset as described above, the random forest models are then trained with 80% of the data while the other 20% are reserved for testing. The trained model is then applied to the MERRA-2 data over Antarctica for BLSN diagnosis (Fig. 1).

Models are trained on a monthly basis (i.e., a different model is built for each month). Here, we use October 2010 as an example.

A critical question is how important each of the 28 MERRA-2 fields is to the BLSN prediction. This can be investigated by comparing a random forest model quantity, namely, feature importance, which measures the contribution of different features to the model’s decision-making process (Yang et al. 2021). Figure 2 displays the normalized feature importance (summed to unity). It is found that the eight parameters listed in the first four rows of Table 1 (highlighted in boldface type) are consistently ranked high in feature importance. These parameters are referred to as baseline parameters throughout the paper. It makes sense that the most important fields include surface pressure, geopotential height, 10-m winds, 2-m air temperature and humidity, and snow age, which are common parameters used directly or indirectly for blowing snow determination in numerical modeling (Déry and Yau 1999; Xiao et al. 2000; Lehning and Fierz 2008; Lenaerts et al. 2012; Letcher et al. 2022; Wever et al. 2023). Among the parameters, the 2-m temperature gradient reflects boundary layer stability, which is linked to Antarctic boundary layer turbulence and mixing (e.g., Ganeshan and Yang 2018, 2019); hence, it is reasonable for it to be among the determining features for BLSN diagnosis. The results also show that although the 10-m winds in both directions are important, the wind in the meridional direction (V10M) is consistently ranked higher than that in the zonal direction (U10M) (Fig. 2). This is because katabatic winds, which are one of the major causes of Antarctic BLSN, blow mostly downslope from the Antarctic plateau largely along the longitudinal lines. Note that blowing snow occurrence is directly dependent on the wind speed and indirectly on the wind direction through its influence on the wind shear.

Fig. 2.
Fig. 2.

The normalized feature importance for the 28 MERRA-2 data fields tested in the random forest models for BLSN diagnosis using data from October 2010, showing results for (a) the random forest classifier for BLSN occurrence prediction and the random forest regressor for BLSN (b) height and (c) optical depth prediction. The dashed orange line is the 3.6% mark, which is the feature importance value if every feature contributes equally.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

As mentioned earlier, performance of the machine learning model predictions is evaluated with the reserved 20% of the dataset. For the random forest classifier, which is used for BLSN occurrence prediction, the evaluation adopts the confusion-matrix–based precision and recall metrics (Chawla 2009). Figure 3 shows the confusion-matrix diagram.

Fig. 3.
Fig. 3.

Confusion-matrix diagram.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

Based on the confusion matrix, precision, recall, and accuracy can be calculated as
Precision=TP/(TP+FP),
Recall=TP/(TP+FN), and
Accuracy=(TP+TN)/(TP+FP+TN+FN).
Precision gives the fraction of model-predicted BLSN points that are, indeed, BLSN in the truth dataset; recall is the fraction of the true BLSN points that are correctly identified by the model; and accuracy gives the overall correctness of the model prediction (Chawla 2009; Yang et al. 2021).

For the random forest regressor, which is used for BLSN height and optical depth prediction, we use mean absolute error, R2 score, and root-mean-square error (RMSE) for performance evaluation. Using these performance metrics, our tests with October 2010 data show that the aforementioned eight MERRA-2 baseline parameters are sufficient for BLSN property diagnosis. Adding more variables does not necessarily improve the model performance. Table 2 compares the performance of BLSN occurrence prediction when 8 versus 28 parameters are used as input. It is shown that although the recall rate is lower for the 8-parameter run, its precision is slightly higher than those of the 28-parameter run. Tables 3 and 4 compare the BLSN height and optical depth predictions. Both cases show that using the eight baseline parameters is better for the machine learning models. Hence, the eight parameters are adopted as the final selection. In the remainder of the paper, results are based on machine learning model runs using the eight parameters as input.

Table 2.

Comparison of random forest model performance for October 2010 BLSN occurrence prediction using 28 and 8 MERRA-2 fields.

Table 2.
Table 3.

Comparison of random forest model performance for October 2010 BLSN height prediction using 28 and 8 MERRA-2 fields.

Table 3.
Table 4.

Comparison of random forest model performance for October 2010 BLSN optical depth prediction using 28 and 8 MERRA-2 fields.

Table 4.

3. Model performance and results

Once the trained models are applied to the MERRA-2 data using the eight baseline parameters as input, they generate hourly BLSN property products, including BLSN occurrence, height, and optical depth. As with MERRA-2, the diagnosed BLSN fields have a spatial resolution of 0.5° latitude by 0.625° longitude resolution. This section analyzes the performance of the models.

a. Case analysis: One time step

Figure 4 demonstrates the performance of the random forest classifier at the time step level. Figure 4a is the model-predicted BLSN occurrence map at 0800:00 UTC 10 October. For comparison, Fig. 4b shows an image from MODIS on board the Aqua satellite taken from 0700:40 to 0700:45 UTC on the same day. The MODIS false-color image was constructed with 2.1 μm as both red and green and 0.85 μm as blue. This combination can make the BLSN area stand out, as seen inside the red oval in Fig. 4b (Yang et al. 2014). The BLSN area visible in the MODIS image also corroborates the CALIPSO observations, shown as the thick line on the image. Note that the CALIPSO observations are nearly simultaneous with those from the MODIS (the observation time difference is less than 2 min). The figure shows that the MODIS-observed BLSN area is well predicted by the machine learning model (Fig. 4a). Figure 5 shows the results from the random forest regressor models for BLSN height and optical depth for the same time step as shown in Fig. 4. Note that only the grid points that are diagnosed as BLSN by the classifier are fed into the height and optical depth prediction model.

Fig. 4.
Fig. 4.

(a) An example of BLSN diagnosis results based on MERRA-2 data over Antarctica at 0800 UTC 10 Oct 2010. Yellow is the predicted BLSN area. (b) An Aqua MODIS granule over Antarctica at 0740–0745 UTC 10 Oct 2010. The false-color image was generated with 2.1, 2.1, and 0.85 μm as red, green, and blue, respectively, so that BLSN area can be seen clearly. The pixels along the CALIPSO track (thick line in the middle) are classified as BLSN (yellow), cloudy (white), and clear (blue).

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

Fig. 5.
Fig. 5.

As in Fig. 4a, but for BLSN properties predicted with MERRA-2 showing (a) height and (b) optical depth.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

b. Case analysis: One month

Here, we continue to use October 2010 to analyze the performance of the machine learning models. Figure 6 compares the BLSN frequency maps for this month from CALIPSO observations (Fig. 6a) and the machine learning model prediction with MERRA-2 (Fig. 6b). The distribution pattern of each matches the other well. Model verification results given in Table 2 show that the accuracy, precision, and recall are 94.7%, 82.9%, and 71.7%, respectively.

Fig. 6.
Fig. 6.

BLSN frequency for October 2010 from (a) CALIPSO observations and (b) the machine learning model prediction with MERRA-2.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

Figure 7 compares the model-predicted BLSN height and optical depth distribution maps for October 2010 with the corresponding CALIPSO retrievals. The metrics, including the R2 score, the RMSE, and the mean absolute error, are given in Tables 3 and 4. Again, the model performed reasonably well. Differences between model results and observation are larger over West Antarctica, where the BLSN frequency is low. As compared with the CALIPSO observations, one prominent feature of the model results is that the coverage is gapless.

Fig. 7.
Fig. 7.

BLSN (a),(b) height and (c),(d) optical depth from (left) CALIPSO observations and (right) the machine learning model prediction with MERRA-2 for October 2010.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

c. Monthly model performance

As we did for October 2010, the training method and BLSN diagnosis procedure are applied to all other months of 2010. Figure 8 shows the performance of the models for each month. It is known that blowing snow frequency is higher in the Antarctic winter months (April to October) (Palm et al. 2011, 2018a). For these months, the precision and recall are greater than 70% (Fig. 8a). For most of the rest of the months (November–February), BLSN frequency over the Antarctic continent is very low (∼2% or lower; see the section 4 discussion); the recall of the models is lower as well. Model accuracy is greater than 90% across all months. In comparison with the winter months, the accuracy for November–February is even higher. The fundamental reason for a high accuracy for these months is that the true negative (TN) points dominate the field. Because of the extremely low BLSN frequency, if the model predicts every point as being no BLSN, the accuracy would still be very high [Eq. (4)]. On the other hand, since the model does not have enough BLSN data points to train on, it gives a very low false positive (FP) rate and high false negative (FN) rate, which lead to a high precision [Eq. (2)] and low recall [Eq. (3)].

Fig. 8.
Fig. 8.

(a) Accuracy, precision, and recall of BLSN occurrence prediction for each month of 2010. Also shown is a comparison of model predicted monthly mean BLSN (b) height and (c) optical depth with CALIPSO observations. The error bar gives the absolute error range.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

Figure 8b shows the verification of the model-forecasted BLSN height. The mean values of model prediction are very close to those of the CALIPSO observations. Mean absolute error is about 28.3 m for the whole year and slightly higher for the nonwinter months. The results of BLSN optical depth are given in Fig. 8c. Again, the model predictions and the CALIPSO observations match very well. The mean absolute error for the whole year is ∼0.06.

d. BLSN properties over the Antarctic continent

As seen from the analysis above, the machine learning approach adopted here can produce reasonable, gapless BLSN diagnosis on an hourly basis using the MERRA-2 data. The results can provide new insight in the properties of Antarctic BLSN. Figure 9 gives the 2010 monthly mean BLSN frequency (Fig. 9a), height (Fig. 9b), and optical depth (Fig. 9c) for East, West, and all of Antarctica, respectively. We use the −45° and 171° longitudes to separate East and West Antarctica (e.g., Ganeshan and Yang 2019). As can be seen from the figure, largely due to the strong katabatic winds originating from the Antarctic plateau, BLSN occurs much more frequently over East than West Antarctica. High-frequency months are from April to September, during which BLSN frequency exceeds 20% over East Antarctica. For May in 2010, the BLSN snow frequency in the region reached 37.1%. Average BLSN height and optical depth do not show significant monthly change during the high-frequency months; the differences between East and West Antarctica are small, as well. Ganeshan et al. (2022) showed that the stronger surface-based inversion over East Antarctica suppresses the vertical extension of BLSN; hence, although the surface wind is stronger over the region, the height and optical depth of BLSN are not necessarily larger than that of West Antarctica. We note that during summer months, the BLSN frequency is very low; hence, the uncertainty in the average height and optical depth is higher.

Fig. 9.
Fig. 9.

Monthly BLSN property statistics for 2010 from the machine learning model applied to MERRA-2 data for the entirety of Antarctica (black), for East Antarctica (red), and for West Antarctica (blue) for BLSN (a) fraction, (b) height, and (c) optical depth.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

Figures 1012 give the spatial distribution of BLSN frequency, height, and optical depth for each month of 2010. Again, BLSN frequency is much higher from April to September than in other months and is much higher for East Antarctica than for West Antarctica. It has been shown that the pattern of the high-BLSN-frequency areas match well with the strong katabatic wind field (e.g., Palm et al. 2018a; Yang et al. 2021). During the Antarctic summer months, however, the BLSN frequency is significantly lower; one reason is the weakening of the katabatic wind during the season (e.g., Palm et al. 2018a). We note that, as shown by Palm et al. (2018a), the BLSN spatial distribution is fairly consistent from year to year.

Fig. 10.
Fig. 10.

Average monthly BLSN frequency from the random forest classifier applied to MERRA-2 data for 2010.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

Fig. 11.
Fig. 11.

Average monthly BLSN height from the random forest regressor applied to MERRA-2 data for 2010.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

Fig. 12.
Fig. 12.

As in Fig. 11, but for BLSN optical depth.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

Unlike the BLSN frequency field, BLSN height and optical depth (Figs. 11 and 12) are not necessarily large over the areas with strong katabatic winds. Although high winds tend to erode the surface-based inversion and favor BLSN occurrence, using data from Dome C, Ganeshan et al. (2022) show that the turbulent layer depth remains shallow during wintertime despite higher wind speeds, due to the suppression effect of the stronger background temperature inversion. Thus, it is reasonable that higher BLSN layers and greater optical depth are predicted in regions where the temperature gradients are weaker, such as near the coast or over West Antarctica.

To further demonstrate the performance of the machine learning models, Fig. 13 compares the averaged BLSN properties for the high-frequency months (April–September 2010) from CALIPSO and MERRA-2. Among the properties, the model performs the best on BLSN fraction diagnosis. More differences exist between CALIPSO observations and modeled BLSN height and optical depth, especially over West Antarctica. One reason is that the BLSN frequency over West Antarctica is much lower than that over East Antarctica, which leads to fewer training samples.

Fig. 13.
Fig. 13.

Comparison of averaged BLSN properties for April–September 2010 between CALIPSO and MERRA-2 diagnosis, showing CALIPSO BLSN (a) fraction, (b) height, and (c) optical depth (OD) and (d)–(f) the corresponding MERRA-2 diagnosis from the machine learning models.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

4. Summary and discussion

BLSN plays an important role in the Antarctic climate system. It can extend hundreds of thousands of square kilometers in area and reach hundreds of meters in height. BLSN can significantly affect ice sheet mass balance and hydrological processes by redistributing surface mass and driving spatial and temporal variations in snow accumulation. This paper presents work using a machine learning model to diagnose BLSN properties with NASA’s MERRA-2 reanalysis. We adopt the random forest classifier for BLSN identification and the random forest regressor for blowing snow optical depth and height diagnosis. BLSN observations from CALIPSO are used as the truth for training the machine learning model. It is shown that with the MERRA-2 fields as input, such as snow age, surface elevation and pressure, temperature, specific humidity, and temperature gradient at the 2-m level, and wind speed at the 10 m level, promising results are achieved. With this method, we can generate gapless BLSN properties on an hourly basis, which was previously constrained by satellite orbital path and coverage. Using the year 2010 as an example, it is shown that the Antarctic BLSN frequency is much higher over East than West Antarctica. High-frequency months are from April to September, during which the BLSN frequency exceeds 20% over East Antarctica. For May 2010, the BLSN snow frequency in the region reached 37.1%. As observed by Ganeshan et al. (2022), BLSN height and optical depth can be suppressed by strong surface-based inversions, which are common over East Antarctica; hence, although characterized by strong katabatic wind and higher BLSN frequency, the height and optical depth of BLSN over East Antarctica, where surface-based inversions are strong, are not necessarily larger.

With this method, we can also conduct BLSN analysis on an hourly basis and learn about the details of Antarctic BLSN field distribution. Figure 14 gives an example of hourly BLSN property variation within a day over Antarctica. This information is important in understanding snow transport, sublimation, and surface mass balance. We note that although this work adopts MERRA-2 reanalysis as input for the machine learning models, other reanalysis datasets, such as ERA-5 (Hersbach et al. 2020), also perform well in representing Antarctic meteorological variables and can be used similarly for BLSN diagnosis (Gossart et al. 2019).

Fig. 14.
Fig. 14.

Hourly BLSN property example for 30 Jun 2010 from the machine learning model applied to MERRA-2 data for the entirety of Antarctica (black), for East Antarctica (red), and for West Antarctica (blue) for BLSN (a) fraction, (b) height, and (c) optical depth. The whisker gives the standard deviation of all BLSN points over Antarctica.

Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-23-0004.1

Many factors can still lead to uncertainties in the results of this method, including potential uncertainties in the machine learning algorithms, in the CALIPSO BLSN detection, and in the MERRA-2 reanalysis fields. However, our analysis shows that the machine learning model performs reasonably well. Our next step will involve applying the diagnosed BLSN fields for better Antarctic mass balance analysis.

Acknowledgments.

We thank three anonymous reviewers for their insightful comments. Funding support for this research is from NASA’s Modeling, Analysis, and Prediction (MAP) and CloudSat/CALIPSO Science programs, both managed by David Considine. Support from the NASA internship program and the Education Office of NASA Goddard Space Flight Center is also acknowledged.

Data availability statement.

All data used in this research are publicly available. The CALIPSO Lidar Level 2 cloud layer product, version 4.20, can be directly downloaded (https://opendap.larc.nasa.gov/opendap/CALIPSO/LID_L2_01kmCLay-Standard-V4-20/contents.html). The CALIPSO Lidar Level 2 Antarctic blowing snow product, version 1.00, can also be directly downloaded (https://opendap.larc.nasa.gov/opendap/CALIPSO/LID_L2_BlowingSnow_Antarctica-Standard-V1-00/contents.html). MERRA-2 data are available online (https://disc.gsfc.nasa.gov/datasets?project=MERRA-2), managed by the NASA Goddard Earth Sciences (GES) Data and Information Services Center (DISC). The Aqua MODIS Level 1b Collection 6.1 calibrated radiances data used in this research can also be found online (https://ladsweb.modaps.eosdis.nasa.gov/archive/allData/6/MYD021KM/).

REFERENCES

  • Barral, H., C. Genthon, A. Trouvilliez, C. Brun, and C. Amory, 2014: Blowing snow in coastal Adélie land, Antarctica: Three atmospheric-moisture issues. Cryosphere, 8, 19051919, https://doi.org/10.5194/tc-8-1905-2014.

    • Search Google Scholar
    • Export Citation
  • Bosilovich, M. G., R. Lucchesi, and M. Suarez, 2016: MERRA-2: File specification. GMAO Office Note 9 (version 1.1), 73 pp., https://gmao.gsfc.nasa.gov/pubs/docs/Bosilovich785.pdf.

  • Chawla, N. V., 2009: Data mining for imbalanced datasets: An overview. Data Mining and Knowledge Discovery Handbook, O. Maimon and L. Rokach, Eds., Springer, 875–886, https://doi.org/10.1007/978-0-387-09823-4_45.

  • Cutler, A., D. R. Cutler, and J. R. Stevens, 2012: Random forests. Ensemble Machine Learning, C. Zhang and Y. Ma, Eds., Springer, 157–175.

  • Das, I., and Coauthors, 2013: Influence of persistent wind scour on the surface mass balance of Antarctica. Nat. Geosci., 6, 367371, https://doi.org/10.1038/ngeo1766.

    • Search Google Scholar
    • Export Citation
  • Déry, S. J., and M. K. Yau, 1999: A bulk blowing snow model. Bound.-Layer Meteor., 93, 237251, https://doi.org/10.1023/A:1002065615856.

    • Search Google Scholar
    • Export Citation
  • Déry, S. J., and M. K. Yau, 2002: Large-scale mass balance effects of blowing snow and surface sublimation. J. Geophys. Res., 107, 4679, https://doi.org/10.1029/2001JD001251.

    • Search Google Scholar
    • Export Citation
  • Dingle, W. R. J., and U. Radok, 1961: Antarctic snow drift and mass transport. IAHS Publ., 55, 7787.

  • Duda, D. P., J. D. Spinhirne, and E. W. Eloranta, 2001: Atmospheric multiple scattering effects on GLAS altimetry. I. Calculations of single path bias. IEEE Trans. Geosci. Remote Sens., 39, 92101, https://doi.org/10.1109/36.898668.

    • Search Google Scholar
    • Export Citation
  • Ganeshan, M., and Y. Yang, 2018: A regional analysis of factors affecting the Antarctic boundary layer during the Concordiasi campaign. J. Geophys. Res. Atmos., 123, 10 83010 841, https://doi.org/10.1029/2018JD028629.

    • Search Google Scholar
    • Export Citation
  • Ganeshan, M., and Y. Yang, 2019: Evaluation of the Antarctic boundary layer thermodynamic structure in MERRA-2 using dropsonde observations from the Concordiasi campaign. Earth Space Sci., 6, 23972409, https://doi.org/10.1029/2019EA000890.

    • Search Google Scholar
    • Export Citation
  • Ganeshan, M., Y. Yang, and S. P. Palm, 2022: Impact of clouds and blowing snow on surface and atmospheric boundary layer properties over Dome C, Antarctica. J. Geophys. Res. Atmos., 127, e2022JD036801, https://doi.org/10.1029/2022JD036801.

    • Search Google Scholar
    • Export Citation
  • Gelaro, R., and Coauthors, 2017: The Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2). J. Climate, 30, 54195454, https://doi.org/10.1175/JCLI-D-16-0758.1.

    • Search Google Scholar
    • Export Citation
  • Gossart, A., S. Helsen, J. T. M. Lenaerts, S. V. Broucke, N. P. M. van Lipzig, and N. Souverijns, 2019: An evaluation of surface climatology in state-of-the-art reanalyses over the Antarctic Ice Sheet. J. Climate, 32, 68996915, https://doi.org/10.1175/JCLI-D-19-0030.1.

    • Search Google Scholar
    • Export Citation
  • Hersbach, H., and Coauthors, 2020: The ERA5 global reanalysis. Quart. J. Roy. Meteor. Soc., 146, 19992049, https://doi.org/10.1002/qj.3803.

    • Search Google Scholar
    • Export Citation
  • Jeppesen, J. H., R. H. Jacobsen, F. Inceoglu, and T. S. Toftegaard, 2019: A cloud detection algorithm for satellite imagery based on deep learning. Remote Sens. Environ., 229, 247259, https://doi.org/10.1016/j.rse.2019.03.039.

    • Search Google Scholar
    • Export Citation
  • Lehning, M., and C. Fierz, 2008: Assessment of snow transport in avalanche terrain. Cold Reg. Sci. Technol., 51, 240252, https://doi.org/10.1016/j.coldregions.2007.05.012.

    • Search Google Scholar
    • Export Citation
  • Lenaerts, J. T. M., M. R. van den Broeke, S. J. Déry, E. van Meijgaard, W. J. van de Berg, S. P. Palm, and J. S. Rodrigo, 2012: Modeling drifting snow in Antarctica with a regional climate model: 1. Methods and model evaluation. J. Geophys. Res., 117, D05108, https://doi.org/10.1029/2011JD016145.

    • Search Google Scholar
    • Export Citation
  • Letcher, T. W., S. L. LeGrand, and C. M. Polashenski, 2022: The Blowing Snow Hazard Assessment and Risk Prediction model: A Python-based downscaling and risk prediction for snow surface erodibility and probability of blowing snow. Engineer Research and Development Center (U.S.) Tech. Rep. ERDC/CRREL SR-22-1, 51 pp., https://doi.org/10.21079/11681/43582.

  • Liu, Z., and Coauthors, 2019: Discriminating between clouds and aerosols in the CALIOP version 4.1 data products. Atmos. Meas. Tech., 12, 703734, https://doi.org/10.5194/amt-12-703-2019.

    • Search Google Scholar
    • Export Citation
  • Palerme, C., C. Claud, A. Dufour, C. Genthon, N. B. Wood, and T. S. L’Ecuyer, 2017: Evaluation of Antarctic snowfall in global meteorological reanalyses. Atmos. Res., 190, 104112, https://doi.org/10.1016/j.atmosres.2017.02.015.

    • Search Google Scholar
    • Export Citation
  • Palm, S. P., Y. Yang, J. D. Spinhirne, and A. Marshak, 2011: Satellite remote sensing of blowing snow properties over Antarctica. J. Geophys. Res., 116, D16123, https://doi.org/10.1029/2011JD015828.

    • Search Google Scholar
    • Export Citation
  • Palm, S. P., V. Kayetha, Y. Yang, and R. Pauly, 2017: Blowing snow sublimation and transport over Antarctica from 11 years of CALIPSO observations. Cryosphere, 11, 25552569, https://doi.org/10.5194/tc-11-2555-2017.

    • Search Google Scholar
    • Export Citation
  • Palm, S. P., V. Kayetha, and Y. Yang, 2018a: Toward a satellite‐derived climatology of blowing snow over Antarctica. J. Geophys. Res. Atmos., 123, 10 30110 313, https://doi.org/10.1029/2018JD028632.

    • Search Google Scholar
    • Export Citation
  • Palm, S. P., Y. Yang, V. Kayetha, and J. P. Nicolas, 2018b: Insight into the thermodynamic structure of blowing-snow layers in Antarctica from dropsonde and CALIPSO measurements. J. Appl. Meteor. Climatol., 57, 27332748, https://doi.org/10.1175/JAMC-D-18-0082.1.

    • Search Google Scholar
    • Export Citation
  • Palm, S. P., Y. Yang, U. Herzfeld, D. Hancock, A. Hayes, P. Selmer, W. Hart, and D. Hlavka, 2021: ICESat-2 atmospheric channel description, data processing and first results. Earth Space Sci., 8, e2020EA001470, https://doi.org/10.1029/2020EA001470.

    • Search Google Scholar
    • Export Citation
  • Pedregosa, F., and Coauthors, 2011: Scikit-learn: Machine learning in Python. J. Mach. Learn. Res., 12, 28252830, https://doi.org/10.5555/1953048.2078195.

    • Search Google Scholar
    • Export Citation
  • Scarchilli, C., M. Frezzotti, P. Grigioni, L. De Silvestri, L. Agnoletto, and S. Dolci, 2009: Extraordinary blowing snow transport events in East Antarctica. Climate Dyn., 34, 11951206, https://doi.org/10.1007/s00382-009-0601-0.

    • Search Google Scholar
    • Export Citation
  • Schmidt, R. A., 1982: Properties of blowing snow. Rev. Geophys., 20, 3944, https://doi.org/10.1029/RG020i001p00039.

  • Toller, G., and Coauthors, 2013: Terra and Aqua Moderate-Resolution Imaging Spectroradiometer collection 6 level 1b algorithm. J. Appl. Remote Sens., 7, 073557, https://doi.org/10.1117/1.JRS.7.073557.

    • Search Google Scholar
    • Export Citation
  • Wang, C., S. Platnick, K. Meyer, Z. Zhang, and Y. Zhou, 2020: A machine learning-based cloud detection and thermodynamic phase classification algorithm using passive spectral observations. Atmos. Meas. Tech., 13, 22572277, https://doi.org/10.5194/amt-13-2257-2020.

    • Search Google Scholar
    • Export Citation
  • Wever, N., E. Keenan, C. Amory, M. Lehning, A. Sigmund, H. Huwald, and J. T. M. Lenaerts, 2023: Observations and simulations of new snow density in the drifting snow-dominated environment of Antarctica. J. Glaciol., 69, 823840, https://doi.org/10.1017/jog.2022.102.

    • Search Google Scholar
    • Export Citation
  • Winker, D. M., M. A. Vaughan, A. Omar, Y. Hu, K. A. Powell, Z. Liu, W. H. Hunt, and S. A. Young, 2009: Overview of the CALIPSO mission and CALIOP data processing algorithms. J. Atmos. Oceanic Technol., 26, 23102323, https://doi.org/10.1175/2009JTECHA1281.1.

    • Search Google Scholar
    • Export Citation
  • Xiao, J., R. Bintanja, S. J. Déry, G. W. Mann, and P. A. Taylor, 2000: An intercomparison among four models of blowing snow. Bound.-Layer Meteor., 97, 109135, https://doi.org/10.1023/A:1002795531073.

    • Search Google Scholar
    • Export Citation
  • Yamanouchi, T., and S. Kawaguchi, 1985: Effects of drifting snow on surface radiation budget in the katabatic wind zone, Antarctica. Ann. Glaciol., 6, 238241, https://doi.org/10.3189/1985AoG6-1-238-241.

    • Search Google Scholar
    • Export Citation
  • Yang, Y., A. Marshak, T. Varnai, W. Wiscombe, and P. Yang, 2010: Uncertainties in ice-sheet altimetry from a spaceborne 1064-nm single-channel lidar due to undetected thin clouds. IEEE Trans. Geosci. Remote Sens., 48, 250259, https://doi.org/10.1109/TGRS.2009.2028335.

    • Search Google Scholar
    • Export Citation
  • Yang, Y., A. Marshak, S. P. Palm, T. Varnai, and W. J. Wiscombe, 2011: Cloud impact on surface altimetry from a spaceborne 532 nm micro-pulse photon counting lidar: System modeling for cloudy and clear atmospheres. IEEE Trans. Geosci. Remote Sens., 49, 49104919, https://doi.org/10.1109/TGRS.2011.2153860.

    • Search Google Scholar
    • Export Citation
  • Yang, Y., S. P. Palm, A. Marshak, D. L. Wu, H. Yu, and Q. Fu, 2014: First satellite-detected perturbations of outgoing longwave radiation associated with blowing snow events over Antarctica. Geophys. Res. Lett., 41, 730735, https://doi.org/10.1002/2013GL058932.

    • Search Google Scholar
    • Export Citation
  • Yang, Y., A. Anderson, D. Kiv, J. Germann, M. Fuchs, S. Palm, and T. Wang, 2021: Study of Antarctic blowing snow storms using MODIS and CALIOP observations with a machine learning model. Earth Space Sci., 8, e2020EA001310, https://doi.org/10.1029/2020EA001310.

    • Search Google Scholar
    • Export Citation
Save
  • Barral, H., C. Genthon, A. Trouvilliez, C. Brun, and C. Amory, 2014: Blowing snow in coastal Adélie land, Antarctica: Three atmospheric-moisture issues. Cryosphere, 8, 19051919, https://doi.org/10.5194/tc-8-1905-2014.

    • Search Google Scholar
    • Export Citation
  • Bosilovich, M. G., R. Lucchesi, and M. Suarez, 2016: MERRA-2: File specification. GMAO Office Note 9 (version 1.1), 73 pp., https://gmao.gsfc.nasa.gov/pubs/docs/Bosilovich785.pdf.

  • Chawla, N. V., 2009: Data mining for imbalanced datasets: An overview. Data Mining and Knowledge Discovery Handbook, O. Maimon and L. Rokach, Eds., Springer, 875–886, https://doi.org/10.1007/978-0-387-09823-4_45.

  • Cutler, A., D. R. Cutler, and J. R. Stevens, 2012: Random forests. Ensemble Machine Learning, C. Zhang and Y. Ma, Eds., Springer, 157–175.

  • Das, I., and Coauthors, 2013: Influence of persistent wind scour on the surface mass balance of Antarctica. Nat. Geosci., 6, 367371, https://doi.org/10.1038/ngeo1766.

    • Search Google Scholar
    • Export Citation
  • Déry, S. J., and M. K. Yau, 1999: A bulk blowing snow model. Bound.-Layer Meteor., 93, 237251, https://doi.org/10.1023/A:1002065615856.

    • Search Google Scholar
    • Export Citation
  • Déry, S. J., and M. K. Yau, 2002: Large-scale mass balance effects of blowing snow and surface sublimation. J. Geophys. Res., 107, 4679, https://doi.org/10.1029/2001JD001251.

    • Search Google Scholar
    • Export Citation
  • Dingle, W. R. J., and U. Radok, 1961: Antarctic snow drift and mass transport. IAHS Publ., 55, 7787.

  • Duda, D. P., J. D. Spinhirne, and E. W. Eloranta, 2001: Atmospheric multiple scattering effects on GLAS altimetry. I. Calculations of single path bias. IEEE Trans. Geosci. Remote Sens., 39, 92101, https://doi.org/10.1109/36.898668.

    • Search Google Scholar
    • Export Citation
  • Ganeshan, M., and Y. Yang, 2018: A regional analysis of factors affecting the Antarctic boundary layer during the Concordiasi campaign. J. Geophys. Res. Atmos., 123, 10 83010 841, https://doi.org/10.1029/2018JD028629.

    • Search Google Scholar
    • Export Citation
  • Ganeshan, M., and Y. Yang, 2019: Evaluation of the Antarctic boundary layer thermodynamic structure in MERRA-2 using dropsonde observations from the Concordiasi campaign. Earth Space Sci., 6, 23972409, https://doi.org/10.1029/2019EA000890.

    • Search Google Scholar
    • Export Citation
  • Ganeshan, M., Y. Yang, and S. P. Palm, 2022: Impact of clouds and blowing snow on surface and atmospheric boundary layer properties over Dome C, Antarctica. J. Geophys. Res. Atmos., 127, e2022JD036801, https://doi.org/10.1029/2022JD036801.

    • Search Google Scholar
    • Export Citation
  • Gelaro, R., and Coauthors, 2017: The Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2). J. Climate, 30, 54195454, https://doi.org/10.1175/JCLI-D-16-0758.1.

    • Search Google Scholar
    • Export Citation
  • Gossart, A., S. Helsen, J. T. M. Lenaerts, S. V. Broucke, N. P. M. van Lipzig, and N. Souverijns, 2019: An evaluation of surface climatology in state-of-the-art reanalyses over the Antarctic Ice Sheet. J. Climate, 32, 68996915, https://doi.org/10.1175/JCLI-D-19-0030.1.

    • Search Google Scholar
    • Export Citation
  • Hersbach, H., and Coauthors, 2020: The ERA5 global reanalysis. Quart. J. Roy. Meteor. Soc., 146, 19992049, https://doi.org/10.1002/qj.3803.

    • Search Google Scholar
    • Export Citation
  • Jeppesen, J. H., R. H. Jacobsen, F. Inceoglu, and T. S. Toftegaard, 2019: A cloud detection algorithm for satellite imagery based on deep learning. Remote Sens. Environ., 229, 247259, https://doi.org/10.1016/j.rse.2019.03.039.

    • Search Google Scholar
    • Export Citation
  • Lehning, M., and C. Fierz, 2008: Assessment of snow transport in avalanche terrain. Cold Reg. Sci. Technol., 51, 240252, https://doi.org/10.1016/j.coldregions.2007.05.012.

    • Search Google Scholar
    • Export Citation
  • Lenaerts, J. T. M., M. R. van den Broeke, S. J. Déry, E. van Meijgaard, W. J. van de Berg, S. P. Palm, and J. S. Rodrigo, 2012: Modeling drifting snow in Antarctica with a regional climate model: 1. Methods and model evaluation. J. Geophys. Res., 117, D05108, https://doi.org/10.1029/2011JD016145.

    • Search Google Scholar
    • Export Citation
  • Letcher, T. W., S. L. LeGrand, and C. M. Polashenski, 2022: The Blowing Snow Hazard Assessment and Risk Prediction model: A Python-based downscaling and risk prediction for snow surface erodibility and probability of blowing snow. Engineer Research and Development Center (U.S.) Tech. Rep. ERDC/CRREL SR-22-1, 51 pp., https://doi.org/10.21079/11681/43582.

  • Liu, Z., and Coauthors, 2019: Discriminating between clouds and aerosols in the CALIOP version 4.1 data products. Atmos. Meas. Tech., 12, 703734, https://doi.org/10.5194/amt-12-703-2019.

    • Search Google Scholar
    • Export Citation
  • Palerme, C., C. Claud, A. Dufour, C. Genthon, N. B. Wood, and T. S. L’Ecuyer, 2017: Evaluation of Antarctic snowfall in global meteorological reanalyses. Atmos. Res., 190, 104112, https://doi.org/10.1016/j.atmosres.2017.02.015.

    • Search Google Scholar
    • Export Citation
  • Palm, S. P., Y. Yang, J. D. Spinhirne, and A. Marshak, 2011: Satellite remote sensing of blowing snow properties over Antarctica. J. Geophys. Res., 116, D16123, https://doi.org/10.1029/2011JD015828.

    • Search Google Scholar
    • Export Citation
  • Palm, S. P., V. Kayetha, Y. Yang, and R. Pauly, 2017: Blowing snow sublimation and transport over Antarctica from 11 years of CALIPSO observations. Cryosphere, 11, 25552569, https://doi.org/10.5194/tc-11-2555-2017.

    • Search Google Scholar
    • Export Citation
  • Palm, S. P., V. Kayetha, and Y. Yang, 2018a: Toward a satellite‐derived climatology of blowing snow over Antarctica. J. Geophys. Res. Atmos., 123, 10 30110 313, https://doi.org/10.1029/2018JD028632.

    • Search Google Scholar
    • Export Citation
  • Palm, S. P., Y. Yang, V. Kayetha, and J. P. Nicolas, 2018b: Insight into the thermodynamic structure of blowing-snow layers in Antarctica from dropsonde and CALIPSO measurements. J. Appl. Meteor. Climatol., 57, 27332748, https://doi.org/10.1175/JAMC-D-18-0082.1.

    • Search Google Scholar
    • Export Citation
  • Palm, S. P., Y. Yang, U. Herzfeld, D. Hancock, A. Hayes, P. Selmer, W. Hart, and D. Hlavka, 2021: ICESat-2 atmospheric channel description, data processing and first results. Earth Space Sci., 8, e2020EA001470, https://doi.org/10.1029/2020EA001470.

    • Search Google Scholar
    • Export Citation
  • Pedregosa, F., and Coauthors, 2011: Scikit-learn: Machine learning in Python. J. Mach. Learn. Res., 12, 28252830, https://doi.org/10.5555/1953048.2078195.

    • Search Google Scholar
    • Export Citation
  • Scarchilli, C., M. Frezzotti, P. Grigioni, L. De Silvestri, L. Agnoletto, and S. Dolci, 2009: Extraordinary blowing snow transport events in East Antarctica. Climate Dyn., 34, 11951206, https://doi.org/10.1007/s00382-009-0601-0.

    • Search Google Scholar
    • Export Citation
  • Schmidt, R. A., 1982: Properties of blowing snow. Rev. Geophys., 20, 3944, https://doi.org/10.1029/RG020i001p00039.

  • Toller, G., and Coauthors, 2013: Terra and Aqua Moderate-Resolution Imaging Spectroradiometer collection 6 level 1b algorithm. J. Appl. Remote Sens., 7, 073557, https://doi.org/10.1117/1.JRS.7.073557.

    • Search Google Scholar
    • Export Citation
  • Wang, C., S. Platnick, K. Meyer, Z. Zhang, and Y. Zhou, 2020: A machine learning-based cloud detection and thermodynamic phase classification algorithm using passive spectral observations. Atmos. Meas. Tech., 13, 22572277, https://doi.org/10.5194/amt-13-2257-2020.

    • Search Google Scholar
    • Export Citation
  • Wever, N., E. Keenan, C. Amory, M. Lehning, A. Sigmund, H. Huwald, and J. T. M. Lenaerts, 2023: Observations and simulations of new snow density in the drifting snow-dominated environment of Antarctica. J. Glaciol., 69, 823840, https://doi.org/10.1017/jog.2022.102.

    • Search Google Scholar
    • Export Citation
  • Winker, D. M., M. A. Vaughan, A. Omar, Y. Hu, K. A. Powell, Z. Liu, W. H. Hunt, and S. A. Young, 2009: Overview of the CALIPSO mission and CALIOP data processing algorithms. J. Atmos. Oceanic Technol., 26, 23102323, https://doi.org/10.1175/2009JTECHA1281.1.

    • Search Google Scholar
    • Export Citation
  • Xiao, J., R. Bintanja, S. J. Déry, G. W. Mann, and P. A. Taylor, 2000: An intercomparison among four models of blowing snow. Bound.-Layer Meteor., 97, 109135, https://doi.org/10.1023/A:1002795531073.

    • Search Google Scholar
    • Export Citation
  • Yamanouchi, T., and S. Kawaguchi, 1985: Effects of drifting snow on surface radiation budget in the katabatic wind zone, Antarctica. Ann. Glaciol., 6, 238241, https://doi.org/10.3189/1985AoG6-1-238-241.

    • Search Google Scholar
    • Export Citation
  • Yang, Y., A. Marshak, T. Varnai, W. Wiscombe, and P. Yang, 2010: Uncertainties in ice-sheet altimetry from a spaceborne 1064-nm single-channel lidar due to undetected thin clouds. IEEE Trans. Geosci. Remote Sens., 48, 250259, https://doi.org/10.1109/TGRS.2009.2028335.

    • Search Google Scholar
    • Export Citation
  • Yang, Y., A. Marshak, S. P. Palm, T. Varnai, and W. J. Wiscombe, 2011: Cloud impact on surface altimetry from a spaceborne 532 nm micro-pulse photon counting lidar: System modeling for cloudy and clear atmospheres. IEEE Trans. Geosci. Remote Sens., 49, 49104919, https://doi.org/10.1109/TGRS.2011.2153860.

    • Search Google Scholar
    • Export Citation
  • Yang, Y., S. P. Palm, A. Marshak, D. L. Wu, H. Yu, and Q. Fu, 2014: First satellite-detected perturbations of outgoing longwave radiation associated with blowing snow events over Antarctica. Geophys. Res. Lett., 41, 730735, https://doi.org/10.1002/2013GL058932.

    • Search Google Scholar
    • Export Citation
  • Yang, Y., A. Anderson, D. Kiv, J. Germann, M. Fuchs, S. Palm, and T. Wang, 2021: Study of Antarctic blowing snow storms using MODIS and CALIOP observations with a machine learning model. Earth Space Sci., 8, e2020EA001310, https://doi.org/10.1029/2020EA001310.

    • Search Google Scholar
    • Export Citation
  • Fig. 1.

    Flowchart for applying the random forest classifier and regressor to BLSN diagnosis with MERRA-2 data.

  • Fig. 2.

    The normalized feature importance for the 28 MERRA-2 data fields tested in the random forest models for BLSN diagnosis using data from October 2010, showing results for (a) the random forest classifier for BLSN occurrence prediction and the random forest regressor for BLSN (b) height and (c) optical depth prediction. The dashed orange line is the 3.6% mark, which is the feature importance value if every feature contributes equally.

  • Fig. 3.

    Confusion-matrix diagram.

  • Fig. 4.

    (a) An example of BLSN diagnosis results based on MERRA-2 data over Antarctica at 0800 UTC 10 Oct 2010. Yellow is the predicted BLSN area. (b) An Aqua MODIS granule over Antarctica at 0740–0745 UTC 10 Oct 2010. The false-color image was generated with 2.1, 2.1, and 0.85 μm as red, green, and blue, respectively, so that BLSN area can be seen clearly. The pixels along the CALIPSO track (thick line in the middle) are classified as BLSN (yellow), cloudy (white), and clear (blue).

  • Fig. 5.

    As in Fig. 4a, but for BLSN properties predicted with MERRA-2 showing (a) height and (b) optical depth.

  • Fig. 6.

    BLSN frequency for October 2010 from (a) CALIPSO observations and (b) the machine learning model prediction with MERRA-2.

  • Fig. 7.

    BLSN (a),(b) height and (c),(d) optical depth from (left) CALIPSO observations and (right) the machine learning model prediction with MERRA-2 for October 2010.

  • Fig. 8.

    (a) Accuracy, precision, and recall of BLSN occurrence prediction for each month of 2010. Also shown is a comparison of model predicted monthly mean BLSN (b) height and (c) optical depth with CALIPSO observations. The error bar gives the absolute error range.

  • Fig. 9.

    Monthly BLSN property statistics for 2010 from the machine learning model applied to MERRA-2 data for the entirety of Antarctica (black), for East Antarctica (red), and for West Antarctica (blue) for BLSN (a) fraction, (b) height, and (c) optical depth.

  • Fig. 10.

    Average monthly BLSN frequency from the random forest classifier applied to MERRA-2 data for 2010.

  • Fig. 11.

    Average monthly BLSN height from the random forest regressor applied to MERRA-2 data for 2010.

  • Fig. 12.

    As in Fig. 11, but for BLSN optical depth.

  • Fig. 13.

    Comparison of averaged BLSN properties for April–September 2010 between CALIPSO and MERRA-2 diagnosis, showing CALIPSO BLSN (a) fraction, (b) height, and (c) optical depth (OD) and (d)–(f) the corresponding MERRA-2 diagnosis from the machine learning models.

  • Fig. 14.

    Hourly BLSN property example for 30 Jun 2010 from the machine learning model applied to MERRA-2 data for the entirety of Antarctica (black), for East Antarctica (red), and for West Antarctica (blue) for BLSN (a) fraction, (b) height, and (c) optical depth. The whisker gives the standard deviation of all BLSN points over Antarctica.

All Time Past Year Past 30 Days
Abstract Views 741 741 0
Full Text Views 858 858 33
PDF Downloads 546 546 16