Performance Analysis of the Temperature and Humidity Profiles Retrieval for FY-3D/MWTHS in Arctic Regions

Abstract

The special geographical location of the polar regions increases the difficulty of modeling surface emissivity, thus the physical retrieval algorithms of the temperature and humidity profiles for microwave radiometers mainly focus on the regions between 60°S and 60°N. In this paper, the deep neural networks (DNN) and long short-term memory (LSTM) models are first implemented to retrieve atmospheric temperature and humidity profiles in real time from FY-3D/MWHTS in Arctic regions and are compared with the physical retrieval algorithm. The hyperparameters of the machine learning models are determined using the grid search and 10-fold cross-validation. Results show that, compared with the physical retrieval algorithm, the retrieval accuracies of the atmospheric temperature and humidity profiles of the DNN and LSTM models in June 2021 are higher over sea ice, and the maximum retrieval accuracies are improved by about 3.5 K and 42%. Over land, the retrieval accuracies of the atmospheric temperature profiles for the DNN and LSTM models in June 2021 are improved by about 5 K. The retrieved humidity results for these two models are not compared with the physical retrieval algorithm, which fails for the humidity profile retrieval over land. In addition, the retrieval results of the DNN-based and LSTM-based models using the independent validation data in February, April, and September are also evaluated over different surface types. The RMSEs of the retrieved temperature profiles for the two models are within 4 K, except for the near-surface, and the humidity profiles are within 25%, except for in February. The temperature profiles in September and the humidity profiles in February are somewhat reduced compared to other months because of the highly variable emissivity properties in autumn and winter. Overall results show that the machine learning method can well-evaluate the retrieval capability of FY-3D/MWHTS of the atmospheric temperature and humidity profiles in Arctic regions.

1. Introduction

The Arctic has a significant effect on global climate change as it is one of the important cold sources in the world [1,2]. The extent of Arctic snow and sea ice cover is shrinking with global warming [3]. Particularly, the Arctic region has warmed at about twice the rate of the entire Northern Hemisphere over the past few decades [4,5]. Moreover, climate change in the Arctic not only has a significant impact on climate regulation in the middle and low latitudes but will also have an uncertain influence on trends in the global climate system in the future [6,7]. Atmospheric temperature and humidity profiles are indispensable parameters for Arctic climate monitoring; they can initialize and evaluate numerical weather forecast models and be used for short-term weather warnings [8].

Compared with radiosonde and ground-based measurements, satellite microwave remote sensing has the great advantage of wide space coverage [9,10,11]. Satellite microwave remote sensing can penetrate clouds for detection and accomplish all-weather detection in comparison to hyperspectral infrared detectors [11,12,13]. Thus, the satellite microwave radiometer can obtain observational data covering the whole Arctic [14]. The physical retrieval method is realized by modeling the atmospheric radiative transfer process and inverting the radiative transfer equation, which is a common retrieval method for atmospheric temperature and humidity profiles with microwave radiometers. Liu et al. [15] used a one-dimensional variational (1DVAR) algorithm to simultaneously retrieve atmospheric temperature, humidity, and cloud-water profiles through the observational data of the advanced microwave sounding unit (AMSU). The results show that, with the scattering model, the radiative transfer model can significantly improve the retrieval accuracy of temperature, humidity, and cloud water. He et al. [16,17] established a 1DVAR system to retrieve the atmospheric temperature and humidity profiles from FY-3C/MWHTS measurements. It is shown that the retrieved temperature and humidity profiles can improve the accuracy of the forecast profile, whereafter the authors in the work [18] proposed a deviation correction method to correct the FY-3C/MWHTS-observed brightness temperature. It is found that the accuracy of the physical retrieval algorithm increases with the brightness temperature correction effect. These studies demonstrate that the physical retrieval algorithm has a high level of precision, but the modeling process is complex. Meanwhile, the performance of the physical retrieval algorithm depends on the forward model, of which the accuracy is limited by the near-surface parameters. Due to the special geographical location of the Arctic and the complex weather conditions, accurate near-surface information is hard to obtain, which increases the difficulty of modeling surface emissivity [12,19]. Although Mathew et al. in [20] utilized AMSU data to retrieve polar temperature profiles by the physical method, the retrieval error of near-surface atmospheric parameters is relatively large because of the influence of surface emissivity. Based on the observations from the ground-based scanning radiometer, which operates at millimeter wavelengths, Cimini et al. [21] utilized the 1DVAR retrieval method to retrieve the temperature and humidity profiles in the Arctic. The retrieved temperature and humidity profiles present as significantly improved compared with the NWP background profiles. At present, the measurement data from microwave radiometers in polar regions are still rarely applied.

The statistical method retrieves the atmospheric temperature and humidity profiles by fitting the nonlinear relationship between the microwave radiometer observed data and the atmospheric profile information, which can avoid the modeling of the forward model. Karbou et al. [22] used a neural network to retrieve the temperature and humidity profiles under clear sky over land from AMSU observation data and classified land surface types. The retrieved temperature and relative humidity profiles are satisfactory regardless of the vegetation types and atmospheric conditions. Gangwar et al. [23] utilized a neural network to retrieve temperature profiles over the ocean and land from AMSU observation data and conducted the classified retrieval of temperature profiles in different latitudes according to the characteristics of the surface types. The retrieval error over the ocean is smaller than that over the land. Gu et al. [24] analyzed the performance of FY-3C/MWHTS and implemented a statistical retrieval method to retrieve atmospheric parameters. Results show that FY-3C/MWHTS improves the comprehensive detection ability of atmospheric temperature and humidity profiles. He et al. [25] retrieved atmospheric temperature and humidity profiles from FY-3D/MWHTS data based on the deep neural networks (DNN) and provided a valuable pathway for retrieving reliable atmospheric temperature and humidity profiles. This research indicates that the machine learning method does not need to calculate the surface emissivity and simulate the radiative transfer process and performs well in dealing with the nonlinear relationships between satellite observations and atmospheric temperature and humidity profiles. This may make up for the poor performance of the atmospheric temperature and humidity profile retrieval of the physical retrieval algorithm in polar regions. However, there is a lack of research on retrieving atmospheric temperature and humidity profiles from microwave radiometer observations in polar regions using machine learning methods.

Although the accurate estimation of the atmospheric temperature and humidity profiles from satellite microwave radiometers in Arctic regions is helpful to improve the initial field of the assimilation system [26,27], the retrieval algorithms of microwave radiometers mainly consider the region within 60°N to 60°S because of the influence of surface emissivity. Thus, it is urgent to study the potential of microwave radiometry in the retrieval of Arctic atmospheric temperature and humidity profiles. In this paper, based on the FY-3D/MWHTS measurements, the DNN and long short-term memory (LSTM) models are implemented to retrieve temperature and humidity profiles in real time in Arctic regions by using the obtained hyperparameters. Then the results are compared with the physical retrieval algorithm. The organization of this paper is as follows. Section 2 introduces the data used in this paper and the data preprocessing process. Section 3 describes the model and the retrieval method for this paper. In Section 4, the retrieval results are analyzed and discussed. Section 5 contains the conclusions.

2. Data

2.1. Data

The datasets used in this paper include the FY-3D/MWHTS-observed data, ERA5 reanalysis data, and National Centers for Environmental Prediction (NCEP) reanalysis and forecast data. These datasets are summarized in

Table 1. Summary of the datasets used for retrieval of atmospheric temperature and humidity profiles.

Data Source	Variable	Time Range
FY-3D/MWHTS	Brightness temperatures Land–sea mask	January to December 2020 and February, April, June, and September 2021
ERA5 reanalysis data	Temperature Relative humidity Specific humidity 2 m temperature 2 m dewpoint temperature Surface pressure Skin temperature 10 m v wind component 10 m u wind component Sea ice cover	January to December 2020 and February, April, June, and September 2021 Temporal resolution 3 h
NCEP reanalysis data	Temperature Relative humidity	February, April, June, and September 2021 Temporal resolution 6 h
NCEP forecast data	Temperature Relative humidity	June 2021 Temporal resolution 6 h

. The geographical ranges of these data include the land area of 60°N–75°N and 90°E–120°E, and the sea ice and mixed ice–water area of 75°N–90°N and 120°W–180°W, respectively.

2.1.1. MWHTS Observations

The MWHTS belongs to the satellite-borne microwave radiometer. It is one of the important payloads of the FY-3D satellite. The MWHTS-observed data can be obtained every 102 min while FY-3D is in polar orbit operation [11]. The MWHTS uses a channel with a central frequency point at the 183.31 GHz water-vapor absorption line to vertically detect atmospheric humidity. It is sensitive to clouds and precipitation with a resolution of 16 km at the nadir. The channel with a central frequency point at 118.75 GHz of MWHTS is sensitive to atmospheric temperature with a horizontal resolution of 29 km at the nadir. The FY-3D/MWHTS-observed data are provided by Level 1 data on the website of the National Satellite Meteorological Center (http://satellite.nsmc.org.cn, accessed from January to December 2020 and February, April, June, and September 2021).

2.1.2. ERA5 Reanalysis Data

ERA5 is the fifth-generation ECMWF reanalysis data, which are available starting from 1950. At present, the ERA5 replaces the previous ERA-Interim [28]. As a large upgrade over ERA-Interim, ERA5 has a higher spatial-temporal resolution, which can obtain the hourly estimates of atmospheric variables at a horizontal resolution of 31 km and can be used to build and validate retrieval systems or assimilation systems. To estimate more accurate atmospheric conditions, these more historical observations are used for producing the ERA5 data [29]. At present, the ERA5 data assimilate multiple data sources, such as satellites, ground observatories, soundings, etc., combined with the atmospheric radiative transfer for television and infrared orbiting satellite operational vertical sounder model, and continuously revise the results, and it is a comprehensive dataset. The reanalysis data are provided by the ECMWF ERA5 website (https://cds.climate.copernicus.eu, accessed from January to December 2020 and February, April, June, and September 2021) and are collected as sample dataset containing profile parameters and the surface parameters over land, sea ice, and mixed ice–water [30]. The spatial resolution and temporal resolutions are 0.25° × 0.25° and 3 h, respectively. The profile parameters are 37 layers unevenly from 1 hPa to 1000 hPa.

2.1.3. NCEP Reanalysis and Forecast Data

The NCEP reanalysis data and 6 h forecast data were provided by the NCEP website (https://rda.ucar.edu (accessed on 1 December 2021)). The forecast data were collected from June 1 to June 30, 2021. The reanalysis data were collected from February, April, June, and September 2021. The spatial and temporal resolutions for reanalysis and forecast data are 0.25° × 0.25° and 6 h, respectively. The reanalysis and forecast dataset includes the same parameters with the atmospheric temperature and humidity profiles. The profile data of NCEP include 31 layers from 1 hPa to 1000 hPa and are interpolated to the specified pressure layer (1–1000 hPa: 37 layers) using the cubic spline interpolation method, which is consistent with the stratification of the ERA5 reanalysis data. After matching, the NCEP forecast data in the matched dataset are used as the initial profiles for the physical retrieval algorithm. The NCEP reanalysis data are used as the validation data.

2.2. Data Preprocessing

The quality control of MWHTS-observed data is carried out to eliminate invalid data before data matching. Then, the MWHTS-observed data, ERA5 reanalysis data, and NCEP reanalysis and forecast data are matched using different temporal and spatial resolutions. Figure 1 summarizes the data preprocessing process. The matching principle between the MWHTS observed data and the ERA5 and NCEP data is that the time difference is not less than 30 min, and the spatial difference is less than 0.1°. Considering that there is a large amount of sea ice and mixed ice–water and that the natural background of land is more complex in the Arctic marine region, we divided these matching data into sea ice, mixed ice–water, and land through the land–sea mask data of MWHTS and the sea ice cover data of ERA5. After completing the matching steps, a total of 771,623, 1,508,152, and 1,183,313 samples were matched over sea ice, mixed ice–water, and land for the whole year of 2020, respectively. To establish more stable and reliable models, 10-fold cross-validation (10-CV) was implemented by calling up the GridSearchCV function. Therefore, the above matched datasets were divided into ten subsets, of which nine subsets were used for model training and one subset for testing to establish retrieval models over sea ice, land, mixed ice–water, and mixed surface, respectively. Meanwhile, GridSearchCV further randomly divided the whole year of 2020 into an 80% training dataset and 20% test dataset to adjust hyperparameters. Finally, we fit the model to the entire set of training data with the best parameters. To verify the generalization performance of the established retrieval models, we also collected data over sea ice, land, mixed ice–water, and mixed surface in February, April, June, and September 2021, respectively, as the independent validation datasets.

3. Algorithm and Experiment Design

3.1. Deep Neural Networks

DNN is a neural network model with multiple hidden layers, including the forward propagation algorithm and the backpropagation algorithm [31]. The loss function is calculated in the forward propagation, and the backward propagation is an optimization process. The model parameters can be optimized and updated by reducing the loss function value using the gradient descent method. The neural network layer inside DNN can be divided into three types: input layer, hidden layer, and output layer. These layers are connected by full connection layers [32].

In this paper, a temperature and humidity profile retrieval model based on DNN is established. The accuracy of the DNN model is mainly affected by the activation function, hidden layer number, hidden layer neurons, learning rate, and optimizer. By using the GridSearchCV method to automatically adjust parameters, a set of optimal parameter combinations is found so that the model has the best effect under this set of parameters. Here, we select the ReLU function as the activation function of the hidden layer to overcome the gradient disappearance problem. The parameters of hidden layers number, hidden layer neurons, and batch size are determined by the GridSearchCV method. The same and the different numbers of neurons in each hidden layer are searched from 300 to 600 at intervals of 50 neurons. Results show that the same number of neurons in each hidden layer has better performance than that of a different number of neurons. The DNN model is finally selected as a six-layer network structure, including one input layer, four hidden layers, and one output layer, and the number of neurons in each hidden layer is 400. The learning rate is 0.0001. The optimizer chooses Adam. Figure 2 shows the neural network structure of the established DNN model, which can simultaneously retrieve the atmospheric temperature and humidity profiles. The input parameters are FY-3D/MWHTS brightness temperature data. The reference truth of the DNN-based model is the atmospheric temperature and humidity profiles of the ERA5 reanalysis data.

3.2. Long Short-Term Memory

LSTM is an improved model of recurrent neural network (RNN), hence inheriting the strengths of the RNN model [33]. Compared with RNN, it has long-term memory functions and can avoid the gradient disappearance problem. Different from DNN, the hidden layer of LSTM contains storage units with memory functions, which can effectively process the time series data. The forget gates, input gates, and output gates of LSTM are used to control the degree of memory and the forgetting of previous and current information [34].

In this paper, a retrieval model of temperature and humidity profiles based on LSTM is also established. The key parameters of the LSTM model are the loss function, activation function, dropout function, optimizer, and the number of hidden layer neurons. The purpose of the loss function is to calculate the amount that the model should seek to minimize during training and finally set the loss function as the mean square error (MSE). Activation functions are mainly used to give the input or output of a group of input nodes. Common ones such as Tanh and Sigmoid are nonlinear activation functions. Tanh is used to adjust the value flowing through the network with the value range of [−1, 1]. The Sigmoid function, unlike Tanh, which helps the network update or forget data, has a value range of [0, 1]. We utilized the GridSearchCV method to optimize the model hyperparameters. After comparison, the activation function is set to Tanh. The dropout function provides an effective method for the approximate combination of various neural network structures in exponential form. It can prevent over-fitting by randomly ignoring selected neurons during the training period. The dropout probability is set to 0.2. The optimizer selects Adam. The number of neurons in the hidden layer increases from 30 to 100, and the search is performed with intervals of 5 neurons. Through the experiment, we found that the best result is the same number of 50 neurons in each hidden layer. Figure 3 shows the neural network structure of the established LSTM model.

3.3. The Physical Retrieval Algorithm

Saunders, R. [35] indicated that the calculation accuracy of simulated brightness temperatures using the radiative transfer for television and infrared orbiting satellite operational vertical sounder (RTTOV) can meet the application requirements. Thus, the RTTOV and the 1DVAR algorithm are used as the forward model and retrieval algorithm, respectively, in this paper. The 1DVAR algorithm is mainly composed of two parts. One is the RTTOV model used to generate simulated brightness temperature, and another is to minimize the cost function [36,37,38]. In the process of the 1DVAR algorithm, we first input the initial atmospheric profiles to the RTTOV model to produce simulations brightness temperature, and then conduct an iterative process to fit RTTOV simulations to MWHTS observations by adjusting the initial atmospheric profiles. Figure 4 shows the procedure to establish the 1DVAR retrieval system.

First, the initial values of the atmospheric temperature and humidity profiles in the dataset are used as inputs into the RTTOV model to calculate the simulated brightness temperature. The initial state variables come from the collocated NCEP 6 h forecast temperature and humidity profiles. Second, we establish a DNN-correction model (four-layer network structure) to correct the observation bias between the observed brightness temperature from MWHTS and the simulated brightness temperature from the RTTOV model. The building process of the DNN-correction model is the same as in Section 3.1. The input of the DNN-correction model is the observed brightness temperature by MWHTS, and the output is the bias between observations and simulations. Based on the DNN-correction model, the bias of the observations for MWHTS can be corrected. Third, the observation-error covariance matrix generated by the errors of the forward models and instrument and background covariance matrix describing the atmospheric state variables is calculated, respectively. Reference [35] gives the details calculation process of these two covariance matrixes. Finally, the corrected brightness temperature, two covariance matrixes, background profiles, and initial atmospheric profiles are input into the 1DVAR retrieval system to obtain the retrieval results.

4. Results Analysis and Discussion

In this section, based on the MWHTS observations, the temperature and humidity profile retrieval results in Artic regions using the DNN model, LSTM model, and physical retrieval algorithm are presented. To better ensure the generalization ability of the established DNN and LSTM models, we use the fully independent validation dataset in June 2021 to compare the performance of these two models with the physical retrieval algorithm. The ERA5 data are used as the reference data to evaluate the retrieval accuracies of the retrieved atmospheric temperature and humidity profiles for three retrieval methods. Reference [39] compared the near-surface temperature data of ERA5 and ERA-Interim with the measured data of 41 weather stations to evaluate the performance of ERA5 in Antarctica. In general, ERA5 can effectively reflect the temperature changes of Antarctica, and it is a powerful tool that can play an important role in exploring Antarctic climate change when in situ observations are sparse. In polar regions, the ECMWF data is generally used as the real values of temperature and humidity profiles to compare with the retrieved results. For example, reference [20] compared the retrieved temperature profiles from AMSU with ECMWF data. Reference [40] estimated the retrieval accuracies of the temperature and pressure profiles, as well as the refractive index of tropospheric dry air and water vapor in the entire Arctic Circle, from radio occultation data, which also used the ECMWF data as reference data. Reference [41] utilized the observations from the high-spectral resolution infrared atmospheric sounder for the FY-3D to retrieve the atmospheric temperature and humidity profiles in the Arctic region, which also compared the retrieved results with the ERA5. Thus, we follow the general validation method and also use ERA5 data as the reference data to verify and compare the retrieval results. Meanwhile, we also add the NCEP independent validation data in February, April, June, and September 2021 to access the quality of the DNN-based model and LSTM-based model over different surface types. The following evaluation indexes are considered to analyze the performance of these two established models and physical retrieval algorithm:

$$R=\frac{{\displaystyle {\sum }_{i=1}^n\left(x_i-\overline{x}\right)}\left(y_i-\overline{y}\right)}{\sqrt{{\displaystyle {\sum }_{i=1}^n{\left(x_i-\overline{x}\right)}^2}{\displaystyle {\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}}$$

(1)

$$RMSE=\sqrt{\frac{1}{n}{\displaystyle {\sum }_{i=1}^n{\left(x_i-y_i\right)}^2}}$$

(2)

$$bias=\frac{{\displaystyle {\sum }_{i=1}^n{x}_i-y_i}}{n}$$

(3)

where n represents the number of profiles; x_i denotes the retrieved atmospheric parameters from the MWHTS observed data; y_i represents the true value of atmospheric parameters.

4.1. Comparison of Retrieval Results over Sea Ice

Considering the distribution characteristics of peak weight function heights for MWHTS channels, in this paper, the analysis of temperature retrieval results is from 50 hPa to 1000 hPa and the relative humidity retrieval results are 225 hPa to 1000 hPa. Figure 5 shows the density scatter plots of the retrieval results of the DNN and LSTM models versus the ERA5 based on the validation dataset in June 2021 over sea ice, and the black lines represent the diagonal y = x, N is the samples of test data, and R represents the correlation coefficient. As can be seen, the temperature retrieval results of the DNN and LSTM models are concentrated on the diagonal, and the correlation coefficients are as high as 0.99 for these two models. The bias of the retrieved temperatures for both are 0.01 K, and the root mean squares errors (RMSEs) are 1.29 K and 1.21 K, respectively. Although the correlation coefficients between the retrieval values and validation values of relative humidity are still high, the retrieval results are relatively dispersed compared with the temperature. The humidity retrieval bias and RMSE are −0.15% and 11.03%, respectively, for the DNN-based model, and −0.02% and 10.50%, respectively, for the LSTM-based model.

In addition, to evaluate the generalization capability of the established DNN and LSTM models, we use the fully independent validation dataset in June 2021 to retrieve the atmospheric temperature and humidity profiles for MWHTS over sea ice. The output results of the DNN and LSTM models are compared with the physical retrieval algorithm, which is shown in Figure 6. It should be noted that to further validate the reliability of the retrieval results for the three retrieval methods, we also add the comparison between the retrieved temperature and humidity profiles and NCEP reanalysis data in Figure 6. The temperature retrieval results between 50–1000 hPa are presented on the left, and the humidity retrieval results between 225–1000 hPa are presented on the right. As can be seen, the trends of the retrieved temperature and humidity profiles over sea ice compared with NCEP are almost consistent with those compared with ERA5, especially for the retrieved humidity profiles. The RMSEs of the retrieved temperature profiles for the three methods near 900 hPa are slightly lower in the NCEP comparisons than in the ERA5 comparisons. On the contrary, the RMSEs of the retrieved temperature profiles for three methods between 350 hPa and 700 hPa are slightly higher in the NCEP comparisons than in the ERA5 comparisons.

Compared to the ERA5 and NCEP data, the retrieval accuracies of the temperature and relative humidity profiles for the DNN and LSTM models are basically better than the physical retrieval algorithm in the whole atmosphere. The RMSEs of the retrieved temperature profile of the DNN model, the LSTM model, and the physical retrieval algorithm are within 4 K, 3.5 K, and 6 K, respectively. Between 200–950 hPa, the RMSEs of the DNN and LSTM models are reduced by more than 0.45 K and the maximum reduced by 2.62 K compared to the physical retrieval algorithm. Especially, at the near-surface, the RMSEs of these two models can be maximum improved by about 3.55 K. The temperature retrieval accuracy of the DNN model is better than that of the LSTM model at around 900 hPa, and the RMSE is reduced by 0.58 K. While from 200 hPa to 400 hPa, the RMSE of the LSTM model is better, and the retrieval accuracy is improved by 0.4 K. The RMSEs of the retrieved humidity profile of the DNN model, LSTM model, and physical retrieval algorithm are within 5–25%, 5–20%, and 10–60%, respectively. The RMSEs of the retrieved relative humidity profile of the DNN and LSTM models are improved by about 20% compared to the physical retrieval algorithm in the range of 650–950 hPa. In particular, retrieval accuracy can be improved by about 42.6% at 450 hPa. The RMSE of the LSTM model is improved by about 1–3% between 250–450 hPa over that of the DNN model. Overall, the performance of the LSTM model is slightly better than that of the DNN model for the humidity profile retrieval over sea ice in June 2021.

Based on the independent validation dataset, Figure 7, Figure A1 and Figure A2 (see Appendix A) show the retrieval results of the temperature profile (left) and relative humidity profile (right) over sea ice in February, April, and September 2021, respectively. As can be seen, the trends of the retrieved temperature profiles over sea ice compared with NCEP are almost consistent with those compared with ERA5 except for the 75–150 hPa and 800–900 hPa of September. However, the accuracies of the retrieved humidity profiles compared with NCEP are higher than those compared with ERA5 between 75–150 hPa, which are opposite, between 500–900 hPa. Especially for April and September, compared with NCEP and ERA5, the maximum differences in the accuracies for humidity profiles can reach 5%. This is because the ERA5 is used as the reference of the truth; the retrieval results of the two models will be closer to ERA5. The RMSEs of the retrieved temperature profiles for the two models are basically within 3 K, except for near-surface in February and at 50 hPa in April, and the retrieved humidity profiles are within 15–25% between 300–1000 hPa. The retrieved temperature profile in February is impacted by the surface emissivity; the RMSE is up to 5.4 K at 975 hPa compared to NCEP. In general, the retrieved results of the temperature profiles for the DNN-based model in three months are superior to the LSTM-based model, except for 225–650 hPa in April, but it is the opposite for the retrieved humidity profiles; the retrieval results of the LSTM-based model are better than those of the DNN-based model apart from 350–550 hPa. Combined with Figure 6, the accuracies of the retrieved temperature and humidity profiles for the two models in April and June are better than those in February and September. Especially in February, the snow cover increases and the volume scattering in the snow is strong, which impact the retrieval accuracy of the two models at the near-surface [42].

4.2. Comparison of Retrieval Results over Land

As in Figure 6, Figure 8 shows the density scatter plots of the retrieved temperature and humidity profiles from the validation dataset in June 2021 over land. It shows that the correlation coefficients of the temperature retrieval results for the DNN and LSTM models can reach 0.99 and are concentrated on the diagonal. The bias of the retrieved temperature of the DNN model and LSTM model are both 0.02 K, and the RMSEs are 1.97 K and 1.83 K, respectively. The density scatter plots of humidity retrieval results are more dispersed than those of temperature. The humidity retrieval bias of the DNN and LSTM models are 0.16% and −0.14% and the RMSEs are 13.70% and 13.10%, respectively.

Figure 9 shows the retrieval results of the temperature profile (left) and relative humidity profile (right) over land in June 2021. The trends of the retrieved temperature and humidity profiles over land compared with NCEP are almost consistent with those compared with ERA5. Compared to the ERA5 and NCEP data, the retrieved temperature and humidity profiles of the DNN and LSTM models over land are better than the physical retrieval algorithm in the whole atmosphere. The RMSEs of the retrieved temperature profiles for the DNN and LSTM models are both within 4 K except for 800–900 hPa, but within 9 K for the physical retrieval algorithm. Compared with the physical retrieval algorithm, the RMSEs of the DNN and LSTM models are reduced by about 5 K and 4 K between 175–225 hPa and between 400–650 hPa, respectively. In the range of 200–500 hPa, the temperature retrieval accuracy of the LSTM model is improved by 0.1–0.5 K compared with the DNN model. It is worth noting that the special geographical location of the Arctic regions increases the difficulty of modeling surface emissivity over land; there still are larger errors between simulation brightness temperatures from RTTOV and observation brightness temperatures from MWHTS after bias correction. However, the 1DVAR retrieval algorithm requires that the bias between the observed brightness temperature and the simulated brightness temperature is unbiased and agrees with the Gaussian distribution; the existence of the larger bias will lead to retrieval failure. In this paper, for the atmospheric humidity profile retrieval over land, the physical retrieval algorithm does not converge under a larger observation bias and fails to obtain the humidity profile information from observations (the green lines of the right sub-graph of Figure 9). Here, we do not compare the retrieved humidity results for the DNN and LSTM models with the physical retrieval algorithm. The RMSEs of the retrieved humidity profile for the DNN and LSTM models are all between 15–25% compared with ERA5 and NCEP. Between 400–950 hPa, the humidity retrieval accuracy of the LSTM model is better than that of the DNN model, with a maximum improvement of 1.4%. In the range of 300–400 hPa, the LSTM and DNN models almost have the same RMSE. These results indicate that the performance of the LSTM model is slightly better than that of the DNN model over land in June 2021. Combined with Figure 6, the temperature retrieval accuracy of the DNN and LSTM models over land is worse than that over sea ice between 175–1000 hPa, but the humidity profile is slightly better than that over sea ice between 500–800 hPa.

Figure 10, Figure A3 and Figure A4 (see Appendix A) show the retrieval results of the temperature profile (left) and relative humidity profile (right) over land. Compared to ERA5 and NCEP, the RMSE of the retrieved temperature and humidity profiles for the DNN-based model and LSTM-based model over land is higher than those over sea ice in the three selected months. The trends of the retrieved temperature profiles over land compared with NCEP are almost consistent with those compared with ERA5, except around 900 hPa in February and September and 200–250 hPa in September. Between 225–350 hPa, the RMSE of the retrieved humidity profiles compared with NCEP are lower than those compared with ERA5, and opposite trends appear in the range of 450–700 hPa. The maximum differences of the accuracies for humidity profile up to 5% at 850 hPa in September when compared with NCEP and ERA5. The RMSEs of the retrieved temperature profiles for the two models are within 4 K between 75–750 hPa in the selected months and reached 5 K in the range of 800–950 hPa in February and September. Except for 300–400 hPa in February and 225 hPa in September, the RMSEs of the retrieved humidity profiles are within 10%–25%. The retrieved temperature profiles are impacted by the surface emissivity in February and September; compared to NCEP and ERA5, the RMSE is up to 7.7 K at 1000 hPa in February and basically above 3 K between 200–1000 hPa in September. In general, the retrieved temperature profiles for the DNN-based model are superior to the LSTM-based model except for 200–500 hPa in April and 200–350 hPa in September. However, for the retrieved humidity profiles, the retrieval results of the LSTM-based model are better than those of the DNN-based model. Combined with Figure 9, the accuracies of the retrieved temperature and humidity profiles for the two models over land in April and June are also better than those in February and September. Especially in February, the emissivity of the land surface is highly variable because of the snow cover, and the retrieval accuracies of the two models are impacted at the near-surface [42]. The retrieval trends of the two models are almost consistent compared with NCEP and ERA5, including the retrieval of the temperature and relative humidity profiles.

4.3. Retrieval Results over Mixed Ice–Water

The physical retrieval algorithm is affected by the surface emissivity, which is difficult to simulate by the RTTOV over mixed ice–water. Therefore, Figure 11, Figure 12, Figure A5 and Figure A6 (see Appendix A) show the retrieval results of the temperature profile (left) and relative humidity profile (right) over mixed ice–water. The trends of the retrieved temperature and humidity profiles over mixed ice–water is almost consistent with those over sea ice, but the retrieval accuracies are worse than those over sea ice because there is the presence of different surface types within one sensor footprint, which impacts the retrieval results. The RMSEs of the retrieved temperature profiles for the two models compared with NCEP are lower than those compared with ERA5 at 900 hPa in April and September. The RMSEs of the retrieved humidity profiles for two models between 500–900 hPa are slightly higher in the NCEP comparisons than in the ERA5 comparisons in February, April, and June. Especially the maximum differences in the accuracies for temperature and humidity profiles can reach about 1 K and 4% compared with NCEP and ERA5, respectively, in February and September. Overall, the RMSEs of the retrieved temperature profile for the two models are within 4 K, apart from the retrieval result for the LSTM-based model in February. The stability of the DNN-based model for temperature profile retrieval is better than the LSTM-based model. For the humidity profile retrieval, the RMSEs for the two models are basically within 5–25%, apart from in February. The LSTM-based model shows better performance, and the RMSEs are improved by 0.1–2.1% over those of the DNN-based model between 300–900 hPa in February and April. The accuracy is maximum improved by about 5% for the LSTM-based model over that for the DNN-based model in April. Compared to other months, the accuracies of the retrieved temperature and humidity profiles for the two models are still lower in February.

4.4. Retrieval Results over Mixed Surface

The temperature and humidity profiles over the mixed surface (including sea ice, land, and mixed ice–water) are difficult to obtain by the physical retrieval algorithm. Here, we still only analyze the retrieval results of the DNN and LSTM models using the validation dataset. Figure 13, Figure 14, Figure A7 and Figure A8 (see Appendix A) show the retrieval results over the mixed surface. The trends of the retrieved temperature and relative humidity profiles over the mixed surface are also basically consistent with those over sea ice and mixed ice–water. Due to the relatively worse performance of the DNN and LSTM for the temperature profile retrieval over land, the RMSEs of the retrieved temperature and humidity profiles for the DNN-based model and LSTM-based model with respect to ERA5 and NCEP over mixed surface are higher than those over sea ice. The trends of the retrieved temperature profiles over mixed surface compared with NCEP are almost consistent with those compared with ERA5 besides the near-surface of September. Between 225–350 hPa, the RMSE of the retrieved humidity profiles compared with NCEP are lower than those compared with ERA5 besides June, and opposite trends appear in the range of 450–700 hPa. The maximum differences of the accuracies for humidity profile up to 2.9% at 450 hPa in June when compared with NCEP and ERA5. The RMSEs of the retrieved temperature profiles for the two models are within 4 K in April and September and within 5 K in February and June. Except for in February, the RMSEs of the retrieved humidity profiles are within 10–25%. The retrieved temperature profiles in February and September are impacted by the surface emissivity; the RMSE is up to 5 K between 800–100 hPa compared to NCEP and ERA5. The retrieved temperature profiles for the DNN-based model are superior to the LSTM-based model between 450–950 hPa in February, June, and September, and the RMSE is maximum improved by 1.1 K around 850 hPa in February, but for the retrieved humidity profiles, the retrieval results of the LSTM-based model are better than those of the DNN-based model except for the 750–950 hPa of February. The RMSEs of the LSTM-based model are improved by 1.1–3.9% and 0.7–3.5% over those of the DNN-based model in April and September, respectively. Overall, the accuracies of the retrieved temperature and humidity profiles for the two models over the mixed surface in April and June are also better than those in February and September.

By comparing the results from Figure 6, Figure 7, Figure 8, Figure 9, Figure 10, Figure 11, Figure 12, Figure 13, Figure 14, Figure A1, Figure A2, Figure A3, Figure A4, Figure A5, Figure A6, Figure A7 and Figure A8 (see Appendix A), it can be seen that the accuracies of the retrieved temperature profiles for the DNN-based model are better than for the LSTM-based model in most cases, but the retrieved humidity profiles have the opposite results. The trends of the retrieved temperature and humidity profiles for the two models are almost consistent, except for the near-surface, whether compared to ERA5 or NCEP. The RMSEs of the retrieved temperature profiles for the two models only around 900 hPa are different in the NCEP and ERA5 comparisons. However, for the retrieved humidity profiles, the RMSEs for the two models are slightly higher in the NCEP comparisons than in the ERA5 comparisons in most cases. For different surface types, the retrieval accuracies of the temperature profiles for the two models are relatively poor at near-surface in February. Apart from the near-surface, mostly the RMSEs of the retrieved temperature profiles in September are higher than in other months. The RMSEs of the retrieved temperature profiles for the two models are within 4 K, except for the near-surface, and the humidity profiles are within 25% except for in February. The temperature profiles in September and humidity profiles in February are somewhat reduced compared to other months because of the highly variable emissivity properties.

5. Conclusions

Different from the physical retrieval algorithm, machine learning methods can avoid the modeling of surface emissivity. Through learning the internal relationships between measurements and retrieval parameters, the atmospheric temperature and humidity profiles can be obtained by the machine learning method. Considering that it is difficult to model the surface emissivity in the polar regions, in this paper, we establish two machine learning models to retrieve atmospheric temperature and humidity profiles in Arctic regions. To optimize the model performance, we utilize the grid search and 10-CV method to determine the hyperparameters of the DNN and LSTM models. Then, the retrieval performances of the two models are compared with the physical retrieval algorithm under all-weather conditions over different surface types, and the performances of the two models are also evaluated using independent validation data for different seasons by comparing with NCEP and ERA5 data. The capability of the temperature and humidity profile retrieval in the Arctic region using MWHTS observations is demonstrated.

To better evaluate the generalization performance of these two established machine learning models, we used the independent validation dataset collected in June, February, April, and September 2021 as the representatives of summer, winter, spring, and autumn, respectively, to assess the quality of the retrieval results for the temperature and humidity profiles. Taking June as an example, we evaluate the performance of the two models by comparing the retrieval results with the physical retrieval algorithm. Results show that the trends of the retrieved temperature and humidity profiles for two machine learning models and physical retrieval algorithms over sea ice and land in the NCEP comparisons are almost consistent with those in the ERA5 comparisons. Compared to ERA5 and NCEP, the RMSEs of the temperature and humidity profiles of these two models are better than those of the physical retrieval algorithm over sea ice and land. The RMSEs of the retrieved temperature and humidity profiles over sea ice can be improved by more than 0.4 K and 20%, respectively, between 650–950 hPa. The RMSEs maximum is improved by 3.55 K at 1000 hPa and 42.6% at 450 hPa, respectively. Over the land, the RMSEs of the retrieved temperature profiles for these two models can improve up to 5 K at 200 hPa. The retrieved humidity results for these two models are not compared with the physical retrieval algorithm, which fails to retrieve the humidity profile over land. Overall, the retrieved temperature profile accuracy over sea ice is higher than that over land.

In addition, considering that the surface emissivity is difficult to model over mixed ice–water and mixed surface, we only analyze the temperature and humidity profile retrieval performance of the DNN and LSTM models by comparing them with ERA5 and NCEP data. For different surface types, the retrieval accuracies of the temperature profiles for the DNN-based model are better than the LSTM-based model in most cases, but there are opposite results for the retrieved humidity profiles. The trends of the retrieved temperature and humidity profiles for the two models over different surface types compared with NCEP are almost consistent with those compared with ERA5, except for the near-surface. The retrieval accuracies of temperature profiles for the two models in different seasons over different surface types are within 4 K, except for near-surface, and the humidity profiles are within 25% except for in February. Overall, the temperature profiles in autumn and humidity profiles in winter are somewhat reduced compared to other seasons because of the highly variable emissivity properties. In the future, to improve the temperature and humidity profile retrieval accuracies of the machine learning method, we will further develop the representativeness of the training samples by dividing the MWHTS observations into clear sky and cloudy sky observed data, respectively.