Precipitation Cloud Identification Based on Faster-RCNN for Doppler Weather Radar

Precipitation clouds are visible aggregates of hydrometeor in the air that floating in the atmos-phere after condensation, which can be divided into stratiform cloud and convective cloud. Different precipitation clouds often accompany different precipitation processes. Accurate identification of precipitation clouds is significant for the prediction of severe precipitation processes. Traditional identification methods mostly depend on the differences of radar reflectivity distribution morphology between stratiform and convective precipitation clouds in three-dimensional space. However, all of them have a common shortcoming that the radial velocity data detected by Doppler Weather Radar has not been applied to the identification of precipitation clouds because it is insensitive to the convective movement in the vertical direction. This paper proposes a new method for precipitation clouds identification based on deep learning algorithm, which is according the distribution morphology of multiple radar data. It mainly includes three parts, which are Constant Altitude Plan Position Indicator data (CAPPI) inversion for radar reflectivity, Radial projection of the ground horizontal wind field by using radial velocity data, and the precipitation clouds identification based on Faster-RCNN. The testing result shows that the method proposed in this paper performs better than typical existing algorithms in terms of accuracy rate. Moreover, this method boasts great advantages in running time and adaptive ability.


Introduction
In recent decades, the precipitation clouds identification based on ground-based weather ra-dar detection data, has been widely used in radar quantitative precipitation estimation, weather modification, and aviation meteorology. The capability of microwaves to penetrate cloud and rain has placed the weather radar in an unchallenged position for remotely surveying the atmosphere [1][2][3]. Doppler weather radar, referred as a greatest tool for detecting weather processes, can not only detect the distribution of precipitation clouds in the atmosphere, but also detect the move-ment trend of precipitation clouds [4][5][6].
In earlier research, the zero-layer bright band is mostly used as the basis for the identification of precipitation cloud types. For the region with the phenomenon of the zero-layer bright band, the precipitation clouds type defaults to stratiform cloud; otherwise it is convective cloud [7]. The limitation of these methods is that the zero-layer bright band is visible only when the stratiform cloud precipitation develops to maturity. Houze et al. [8] proposed a method to distinguish the precipitation cloud types by using rainfall gauge measurement data. When the precipitation intensity exceeds a certain fixed threshold, the precipitation cloud is considered as convective cloud; otherwise, it is stratiform cloud. The limitation of this method is that it can only determine the center of convective cloud, and it is easy to misjudge the precipitation area of convective cloud with weak precipitation intensity nearby. Churchill et al. [9] determine the convective cloud center through the radar reflectivity threshold. Moreover, precipitation clouds within a fixed radius of the convective cloud center are all convective clouds by default, otherwise they are stratiform clouds.
Steiner et al. [10] pointed out the irrationality of using reflectivity threshold and influence radius to determine convective cloud region. The reflectivity threshold of the convective cloud center should be determined according to the reflectivity distribution of the region near the convective cloud center, and the influence radius should be calculated by the reflectivity of the convective cloud center. The limitation of these methods is that they do not take into account the influence of the zero-layer bright band, which leads to misdiagnosis of stratiform clouds as convective clouds.
Biggerstaff et al. [11] analyzed the distribution of radar reflectivity corresponding to stratiform and convective cloud in three-dimensional space, and applied it to the identification of precipitation clouds. This method takes into account the influence of zero-layer bright band and achieves good recognition results. Therefore, this method has been widely used in the United States. In Reference [12], the authors proposed a method to distinguish the precipitation cloud types based on the difference of radar reflectivity distribution morphology between convective and stratiform precipitation clouds. In this paper, six preparative reflectivity-morphological parameters are presented, which are composite reflectivity and its horizontal gradient, echo top height associated with 35 dBZ reflectivity and its horizontal gradient, vertically integrated liquid water content and its density. Moreover, the precipitation clouds are identified by using the fuzzy logic algorithm based on the distribution of each reflectivity-morphological parameter. The accuracy of these methods depends on the reliability and quantity of the selected parameter. However, the increment of identification parameters leads to the multiplication of computation, which causes the great reduction of the efficiency of the algorithm.
All of the above methods are based on the reflectivity data, which are detected by Doppler weather radars. However, the radial velocity data detected by Doppler Weather Radar has not been applied to the identification of precipitation clouds. It is due to the radial velocity of the radar is the velocity of the precipitation particle relative to the radar, which is not sensitive to the convection movement in the vertical direction [13,14]. In the present study, we propose a new method for precipitation clouds identification based on the deep learning algorithm. Firstly, the adaptive Barnes interpolation method is used to interpolate the radar scanning volume data to obtain the reflectivity CAPPI data at multiple altitudes. Secondly, the radial velocity data of the ground horizontal wind field is inversed by using the radial velocity data detected by Doppler weather radar at minimum elevation. Finally, the Faster-RCNN model is applied to identify the precipitation clouds based on the result above mentioned. This paper is organized as follows: Section 2 describes the Datasets for deep learning training and testing. Section 3 presents the new method for precipitation clouds identification based on Faster-RCNN algorithm. Section 4 evaluates the proposed method by comparing the identification results with the traditional method. Section 5 concludes this study.

Data
In this research, we collected a large amount of precipitation data from the CINRAD-SA dual-polarization Doppler weather radar, which is located in Guangzhou, Guangdong, China (23°01′N, 113°35′E). The radar characteristics are shown in Table 1.  In this paper, Faster-CNN algorithm is adopted to identify the precipitation clouds, which is a target detection algorithm based on convolution neural network (CNN). The accuracy of the final identification result depends on the number of sample sets in the training process. In order to obtain high accuracy of precipitation clouds identification, this paper constructs a rich sample set, which including 6400 training samples and 800 test samples. Table 2 summarizes the datasets used in this paper. As shown in Table 2, each sample is a matrix with a size of 400 400 7  cells, and it can be divided into three parts, which are reflectivity data ( Z ) after CAPPI inversion; the radial velocity data (V ) and spectral width data (W ) of the ground horizontal wind fields, which are inversed by using the radar detection data at minimum elevation.

The CAPPI Inversion of Reflectivity
The rainfall measurement of dual polarization weather radar is based on the basis that the raindrops in the process of falling are not spherical but have an oblate spheroidal shape, and the shape is determined by the equivalent diameter and axis ratio. The axis ratio b/a , the ratio of the major axis length to minor axis length, is related to the diameter e D of an equivalent-volume spherical raindrop as [9].
The stratiform cloud is caused by the vertical rising movement of the air in a wide range.
Moreover, the rising velocity is uniform and often less than the falling velocity of raindrops. In generally, stratiform cloud is larger in horizontal scale, thinner in vertical thickness, and has flat tops [15,16]. During the radar detection, the characteristics of stratiform cloud in reflectivity data can be summarized as follows. a) The reflectivity data are uniformly distributed and have a diffuse shape. In the process of large-scale stratiform cloud precipitation, there may be multiple reflectivity centers, but the value of reflectivity center generally does not exceed 30dbZ. b) The horizontal and vertical gradients of reflectivity data are small. c) It is often accompanied by the phenomenon of the zero-layer bright band when surface precipitation particles are liquid.
The convective cloud is caused by the air vertical movement due to the instability of the atmosphere. Its horizontal scale is small, and the vertical convection is strong, often accompanied by storm, rainstorm, hail, and other disastrous weather [17]. During the radar detection, the characteristics of convective cloud in reflectivity data can be summarized as follows. a) The reflectivity value of the convective cloud is larger than stratiform cloud, and the central reflectivity value can reach above 50dbZ. b) The horizontal and vertical gradients of reflectivity data are large. In china, the main operational radar of China Meteorological Administration is Doppler weather radar, which can not only obtain the distribution information of precipitation particles in atmosphere, but also provide the information of the atmospheric wind field. During the Doppler weather radar operation process, VCP21 working mode is a common choice for meteorological service. It can complete a volume scan in 6 minutes, and each volume scan consists of 9 elevations.
For each elevation, the Plane Position Indicator (PPI) scan mode is adopted. As shown in Fig. 1, the radar antenna captures precipitation data in the atmosphere in a fixed elevation, omni-directional scanning mode, and in polar coordinates (radar-centered) with different color scales to indicate the magnitude and direction of the values. This scanning method is called PPI scanning. In other words, radar can acquire the information of precipitation particles in a three-dimensional space in 6 minutes, which provides guarantee for the real-time prediction in weather operation. In the process of identifying the precipitation cloud, in order to describe the vertical and horizontal characteristics of the precipitation cloud system more intuitively, the PPI data of all elevations need to be processed through spatial interpolation to obtain the information of the precipitation particles at a certain altitude, namely CAPPI data.

Fig. 2 An illustration of Adaptive Barnes interpolation.
Adaptive Barnes interpolation is a commonly used interpolation method, which is mainly used in discrete point interpolation with sparse distribution [18,19]. As shown in Fig. 2, where 18 ff are values detected by radar in 8 range bin, 0 f is the point that needs to be interpolated.
The interpolation method can be expressed as Where, k  is the weight of k f , which can be defined as Where, In equation 2, r K , K  and K  are smoothing parameters of radial distance, azimuth and elevation respectively, which can be adjusted to achieve different smoothing effects. In this paper, The resolution of radar data in radial distance, azimuth and elevation is different during interpolation. In the radial direction, the distance between adjacent range bins is fixed. In the direction of azimuth, the distance between adjacent range bins increases with the increase of radial distance. Moreover, in the elevation direction, the distance between adjacent range bins increases with the increase of radial distance. Therefore, different interpolation points have different dependencies in three directions, so it is necessary to set smoothing parameters according to the dependencies in three directions [34]. Compared with the traditional interpolation method for CAPPI inversion, the Adaptive Barnes interpolation method has better smoothing effect. The inputs of the deep learning algorithms are mostly the data after standardization. Before using the Faster-RCNN model to identify the precipitation clouds, the paper needs to conduct a standardized processing on the reflectivity data. Fig. 3(a) is the CAPPI distribution diagram of reflectivity data at a height of 2 km, which corresponds to a reflectivity data matrix with a size of 400 400  cells. Fig. 3(b) is the distribution diagram of reflectivity matrix after standardization, which can be summarized as Where, X is the input to the normalized function, and X is the output after standardization.

Radial projection of ground horizontal wind field
As the name suggests, convective cloud have strong convective phenomena in the vertical direction, which should be an important basis for judging convective cloud. However, the velocity data detected by the Doppler weather radar is the relative velocity of cloud particles in the radial direction, and it is not sensitive to the convective motion in the vertical direction, so it is difficult to apply this feature to the identification of precipitation clouds. Convective clouds usually consist of one or more convective cells, each of which has a horizontal scale ranging from several kilometers to tens of kilometers. In radar reflectivity maps, convective cells are usually marked by a tight reflectivity region or a strong updraft that causes deep convection [20]. As shown in Fig. 4, the evolution process of convective cells usually includes three stages: growth stage, mature stage, a b and disappearing stage [21].  of radar, same scale cyclones were simulated in four directions of the radar station, which are projected into the radial velocity map corresponding to the 0.5° elevation. Fig. 6(a) shows the PPI distribution of radial velocity corresponding to the convergence wind field. As shown by the cyclone due north, the radial velocity of the left region is negative, and the right side is positive.
The radial velocity of the convergence field gradually increases from the outside to the inside, and suddenly becomes zero at the center of the convergence field, which is mainly caused by the uplift of the convergent airflow. Moreover, the convergence field has obvious zero velocity line at the boundary of positive and negative velocity. In addition, the PPI distribution of radial data corresponding to the divergence wind field is shown in Fig. 6(b).

Fig. 6
The PPI distribution of the radar radial velocity data corresponding to a convergence and b divergence As mentioned above, in the growth and maturity stages of convective cells, the ground wind field is accompanied by convergence and divergence, which has obvious distribution characteristics in the radial velocity data detected by the radar at the minimum elevation.
Moreover, the minimum elevation of Doppler weather radars is usually less than 0.5°, and the vertical component of the radial velocity detected by the minimum elevation is close to zero.
Therefore, the radial velocity detected by the minimum elevation can be regarded as the projection of horizontal wind field on the radar radial direction. In order to maintain the consistency of the radial velocity data and the reflectivity CAPPI data in spatial distribution, this paper projected the radial velocity data into the horizontal surfaces, and took it as the radial velocity data of the ground horizontal wind field.

a b
The spectral width of the radial velocity data is an important Doppler weather radar data product, which reflects the dispersion of the radial velocity data in a range bin. Moreover, it is proportional to the variance of the velocity vectors of each scatterer in the range bin. In the convergence and divergence field, the distribution of velocity vectors in the range bin is usually chaotic, so it usually has a large spectral width. Therefore, the spectral width of the radar radial velocity data should be applied to identify the precipitation clouds. In this paper, the radial velocity data and spectral width data corresponding to the ground wind field is obtained by the projection of the radar detection data at the minimum elevation. As shown in Table 2, the size of radar radial data and spectral width data corresponding to the ground wind field is 400 400  .

The identification of precipitation clouds base on Faster-RCNN
In recent years, deep learning has developed rapidly, and the target detection method based on deep learning has achieved good results. References [21] to [33] describe the achievements of deep learning in target detection. The deep learning algorithm has strong self-learning ability and can extract features from the original input data, which reflect the structural information of the input data. Therefore, compared with the traditional target detection algorithm, the target detection methods based on deep learning have more advantages in accuracy and efficiency. Moreover, the emergence of GPU provides a guarantee for the real-time performance of deep learning. In this paper, the Faster-RCNN model is adopted to identify the precipitation clouds, and its block diagram is shown in Fig. 7.

Fig. 7 The block diagram of precipitation clouds identification based on Faster-RCNN
Before identifying precipitation clouds, this paper preprocesses radar data, including CAPPI inversion of reflectivity, and radial projection of ground horizontal wind field, which are mentioned above. Faster-RCNN model is composed of Feature extraction, Region Proposal Networks (RPN) and Classification, which are described below.
feature extraction. In this algorithm, multiple 33  convolution layers and 22  pooling layers are used to extract features from the original data, and the performance of the model is improved by increasing the number of layers in the network. Moreover, the series of multiple small convolution kernels can be equivalent to a large scale convolution kernel, but the number of parameters in the network will be greatly reduced, and the nonlinear expression ability of the network will be enhanced. In this paper, the VGG16-Net is applied to extract features from radar data, and the architecture of the VGG16-Net used in this paper is shown in Fig. 8.  Table 3. Finally, the output layer is a Feature Map with a size of 25 25 512  , which reflects the spatial structure characteristics of the input data.  Fig. 9. Fig. 9 The architecture of RPN The detailed explanation of the RPN in Fig. 9 can be seen below: 1. The input data is a 25 25 512  feature map, which is obtained by VGG16-Net.
2. To generate region proposals, we slide a small network over the input data. This network is fully connected to a 3 × 3 spatial window of the 25 25 512  feature map. Moreover, each sliding window is mapped to a 512-d vector.
3. At each sliding-window location, we simultaneously predict k region proposals (where 9 k  ). The 9 proposals are parameterized relative to 9 reference boxes, called anchors.
Moreover, 9 anchors are generated by using 3 scales and 3 aspect ratios. (2) 5. The outputs are 300 optimal region proposals, which are selected from a large number of region proposals by using the Non-Maximum Suppression method [35].

Classification
The architecture of classification is shown in Fig. 10. Each optimal region proposal is mapped to the 25 25 512  feature map to obtain the corresponding feature expression. This feature expression is divided into 77  sub-blocks according to horizontal and vertical directions, and then each sub-block is maximally sampled to obtain a feature map with a size of 77  . This process is called ROI Pooling [33]. iterations was 10,000 times.
2. To train the detection network based on (1), the number of iterations was 10,000 times.
3. Fixed feature extraction module, fine-tuned parameters of non-shared layer in RPN network, and the number of iterations was 5000 times.
4. Fine-tuning the parameters of the non-shared layer in the detection network, and the number of iterations is 5000 times.
In this paper, the experimental environment with: NVIDIA Tesla P40 GPU, CUDA 9.1, Ubuntu 16.04, memory 24GB. Moreover, the development platform is Python+Tesorflow. The training set and testing set are described in Table 2. The test results show that the precision of identification results is 96%.

Results and discussions
In this paper, the performance of the precipitation clouds identification method is evaluated by the data detected in a mixed cloud precipitation process by using the CINRAD-SA dual-polarization Doppler weather radar which is upgraded in China. In order to verify the reliability of the identification method, we adopt the identification method proposed in Reference [12] to identify this case, and find that the identification results of the two methods are basically consistent. The reflectivity CAPPI data are inverted by using the Adaptive Barnes interpolation method, which is introduced in section 3.1. Fig. 11(a)  The data resolution is 500×500 meters, and the distance between adjacent circles is 25 km. Fig.   11(b) is the radial velocity distribution map of the ground horizontal wind field. Fig.11(c) is the spectral width data distribution map of the ground horizontal wind field. Each Figure contains two parts, the original distribution map on the left, and the standardized distribution map on the right.

Fig. 11
The pre-processed radar data corresponding to 2342UTC, a CAPPI distribution map of the reflectivity at 2-6 Km altitudes, b Radial velocity distribution map of the ground horizontal wind field, c Spectral width data distribution map of the ground horizontal wind field.
As shown in Fig. 11(a), the red area on northeast of the radar station is a heavy precipitation area. The reflectivity value of this area reaches over 40dbZ, and the reflectivity center reaches 60dbZ. In the horizontal direction, the radar reflectivity decreases gradually with the outward direction of the reflectivity center, and the reflectivity value at 20 kilometers away from the reflectivity center decreases to 30dbZ. It reflects that the reflectivity of this region has a large gradient in the horizontal direction. In the vertical direction, the radar reflectivity reaches its maximum value at the height of 3km. As the height increases or decreases, the reflectivity will gradually decreases, which reflects that the reflectivity of this region also has a large gradient in the vertical direction.
As shown in Fig. 11(b), the left side of the radar station is southeast wind, and the right side is southwest wind. It is accompanied by slight convergence. As described in section 3.2, it can be inferred that this time is the middle and late period of convective cloud precipitation. The radial velocity has obvious abrupt change in the heavy precipitation area above mentioned. Moreover, the radial velocity decreases gradually when it closes to the precipitation center. In Fig. 11(c), the a b c spectral width data of this heavy precipitation area gradually increases when it closes to the precipitation center. Combined with the characteristics above mentioned, the precipitation cloud corresponding to the heavy precipitation region should be convective cloud. Fig. 12(a) shows the identification result of Faster-RCNN model. It can be seen that the identification results are basically consistent with the analyses above.
Reference [12] proposes six preparative reflectivity morphological parameters for identifying precipitation clouds, includes composite reflectivity and its horizontal gradient, echo top height associated with 35 dBZ reflectivity and its horizontal gradient, vertically integrated liquid water content and its density, which are based on the differences of radar reflectivity distribution morphology between convective and stratiform clouds. Moreover, the fuzzy logic method is applied to identify the precipitation cloud based on the distribution law of each parameter. This identification method has been widely used in China. We use this method to identify the precipitation clouds at 2342UTC, and the identification results are shown in Fig. 12  The identification results of the two methods are compared, and the comparison results are shown in Fig. 12(c). The red box corresponds to the convective cloud box in Fig. 12(a). Moreover, the red area corresponds to the convective cloud in Fig. 12(b). It can be seen that the identification results of the two methods are basically consistent. In addition, the above two methods are used to identify the precipitation clouds corresponding to the 12 continuous volume scan data. The identification results are shown in Fig. 13. b a c Fig. 13 The identification results of precipitation clouds corresponding to 2342UTC-0054UTC As shown in Fig. 13，the identification results of the two methods are basically consistent.
Moreover, the testing set in Table 2 is used in the evaluation of the two identification methods, and the evaluation results are shown in Table 4. It can be seen that the traditional methods have higher recall rate, but poor accuracy. However, the new identification method proposed in this paper has great advantages in efficiency and accuracy.

Conclusions
With continued threat from severe convection weather, the identification of precipitation clouds is crucial to weather forecast and aviation meteorology. Traditional identification methods utilize some parameters calculated from radar reflectivity to identify the precipitation clouds, which are based on the differences of radar reflectivity distribution morphology between strati-form and convective clouds in three-dimensional space. These methods have a common shortcoming that the radial velocity data detected by Doppler Weather Radar has not been applied to the identification of precipitation clouds because it is insensitive to the convective movement in the vertical direction. However, the convective movement in the vertical direction is an important feature for precipitation clouds identification. Section 3.2 draws the conclusion that the ground horizontal wind field is often accompanied by convergence and divergence during the growth and mature stages of convective cells. Moreover, the radial velocity detected by the Doppler weather radar at minimum elevation can be regarded as the projection of horizontal wind field on the radar radial direction. Therefore, the radial velocity data detected by the Doppler weather radar at minimum elevation can be used to the identification of precipitation clouds.
In this paper, a new method for precipitation clouds identification was described, which makes full use of the data products of Doppler weather radar, including reflectivity, radial velocity and spectrum width of radial velocity. Firstly, the adaptive Barnes interpolation method is used to interpolate the radar scanning volume data to obtain the reflectivity CAPPI data at multiple altitudes. Secondly, the radial velocity and the spectral width of the horizontal wind field on the ground are respectively obtained by the projection of the radar detection data at the minimum elevation, which reflect the formation and development of convective motion. Finally, the Faster-RCNN algorithm was applied to identify the precipitation clouds based on the data above mentioned. Experiments to evaluate the proposed method showed that the new method has great advantages in efficiency and accuracy compared with the traditional methods.
This study is the first one that reports applying the radial velocity data detected by the Doppler weather radar to the identification of precipitation clouds. Moreover, the deep learning algorithm was applied to identify the precipitation clouds. Compared with traditional identification methods, this method achieves better results in efficiency and accuracy. However, the training process of Faster-RCNN model usually requires large labeled data and powerful computational resources. It is obvious that the training samples described in section 3.2 are insufficient. Future work should be focused on collecting more samples for training and testing of the Faster-RCNN model. Moreover, Future work should minimize the running time of the precipitation clouds identification system by optimizing the internal algorithms.