
The data used for establishing different deep learning models are extracted from the global 1° × 1° NCEP final (FNL) analysis data during the period 2010–14. All data are available 4 times a day (0000, 0600, 1200, 1800 UTC), providing global scenarios of weather. After the deep learning model is established, the 1° × 1° forecast data from GFS are used for forecasting SCW.
In order to accelerate the training process and improve the predictive accuracy of the deep network, we first select a set of predictors among all variables in the FNL analysis data. The predictors contain all major environmental conditions that are favorable for SCW events, which include basic meteorological elements such as pressure, temperature, geopotential height, humidity, and wind, as well as a number of convective physical parameters that can reflect water vapor, atmosphere instability, and uplift conditions (Tian et al., 2015). For example, most unstable convective available potential energy (MUCAPE), precipitable water (PWAT), convective inhibition (CIN), convergence, and wind shear are such physical parameters. To account for the geographical differences between various regions, we also use elevation, longitude, and latitude in our model. In total, 144 predictors were selected to describe environmental characteristics of SCW (Table 1) and all those predictors were extracted from the FNL analysis data.
Feature Level (hPa) Multilevel variable T (temperature)
H (geopotential height)1000, 925, 850,
700, 600, 500,
400, 300, 200WS (wind speed)
WD (wind direction)
W (vertical wind speed)TDD (temperature dew point difference)
Q (specific humidity)
VAPFLUXDIV (water vapor flux divergence)PV (potential vorticity)
TMPADV (temperature advection)
SITASE (potential pseudoequivalent temperature)
DIV (divergence)
VOR (vorticity)
VORADV (vorticity advection)Singlelevel convective parameter MUCAPE (most unstable convective available potential energy)
BLI (best lift index)
CIN (convective inhibition)
DCAPE (downdraft convective available potential energy)
K (Kindex)
LI (lift index)
Z0 (altitude of 0°C)
Z20 (altitude of 20°C)
PWAT (precipitable water)
SHIP (significant hail parameter)
SHEAR1 (0–1km wind shear)
SHEAR3 (0–3km wind shear)
SHEAR6 (0–6km wind shear)
SI (Showalter index)
TT (total index)Others Elevation
Longitude
LatitudeTable 1. Selected predictors for SCW forecasting with the deep learning model

The SCW events, in particular the HR, hail, and CG, are very rare. On average, the maximum number of days per year with thunderstorm, HR, hail, and CG is less than 110, 13, 5, and 13, respectively. Thunderstorm and HR seem to share somewhat similar spatial pattern. They both occur most frequently in South China. However, only thunderstorm while no HR is observed in West China. The spatial pattern of hail and CG also appears to be similar with hotspots of both types of events concentrated in Tibet (Sun et al., 2014).
Observations of thunderstorms, HR, hail, and CG, which were used to label the predictors, were obtained from the severe weather observation dataset of NMC (Zheng et al., 2013). The thunderstorm observations consist of lightning location data collected by the National Lightning Location Network (NLLN) of China. It has been installed with groundbased advanced time of arrival and direction system of cloudtoground lightning detection sensors, reaching 394 in operation in 2016, covering most of China. According to relevant studies (Xia et al., 2015; Yang et al., 2015), the lightning location accuracy of the whole network is approximately 300 m, the detection rate is larger than 80%, and the average detected radius of a sensor is approximately 300 km. A thunderstorm is recorded if at least one lightning strike is observed by the NLLN. The HR data consist of observations of hourly rainfall no less than 20 mm. Rainfall is measured by automatic rain gauges at 2420 nationallevel weather stations (NWSs) and more than 20,000 automatic weather stations. The hail and CG observations are from observer reports, and are available 24 h a day at the 2420 NWSs in mainland China.

The deep learning algorithm for SCW forecasting (Fig. 1) includes three major steps. First, the training and testing datasets are collected. Second, a deep learning network is constructed, trained, and tested. Third, the trained network is implemented for forecasting.

The weather forecasting can be regarded as a twocategory classification problem, i.e., 0 indicates that the event will not happen, and 1 indicates that it will happen. To feed the deep learning network with sufficient spatial information of climate variables, our model input is set to be 144 observed climate variables over a square patch with dimension L × L centered at each SCW event grid. These L × L × 144 data arrays are labeled by either 1 or 0 depending on whether SCW occurs at its center grid or not. These labeled data arrays form either the training or testing samples. The choice of L serves as a balance of tradeoff between computational efficiency and model performance. Some preliminary experiments suggested that L = 7 is an optimal choice for this purpose.
Since the NWP system yields gridded fields and the SCW observations are sitebased data, the observations need to be remapped to the NWP grids first. If an observed SCW event occurred within a radius, R, of the grid point, the grid point is marked by 1, indicating that the event occurred at this grid point. Otherwise, the grid point is marked by 0. Considering that the SCW often occurs on a mesoγ scale, R is set to be 20 km in this study. Note that if R is set too small, there will be too many missing forecasts; while if R is too large, there will be too many false alarms.
Compared to nonSCW events, the SCW is a highimpact and lowprobability event. Therefore, positive samples (marked as “with SCW”) are far fewer than negative samples (marked as “without SCW”). This reflects a typical sample set imbalance (Krawczyk, 2016). To remedy this issue, positive samples are replicated to balance the positive and negative samples in the training sets. This process is called oversampling (Buda et al., 2018). Oversampling is unnecessary for test sets though, as the test sets are mainly used to assess the performance of the trained models. We therefore constructed test sets without oversampling to assess the performance of our trained models under the real positive–negative sample ratio.
Two independent datasets were constructed based on SCW observations and NCEP FNL analysis data for the period of March–October during 2010–14. One contains a sample of 50 days that were selected by randomly choosing one day in each month from March to October of 2010–14, and is considered as the test set. The other constructed dataset contains all the remaining positive and negative samples (4,582,577 thunderstorm samples, 3,609,185 HR samples, 1,468,158 hail samples, and 1,488,531 CG samples) and is treated as the training set.

As mentioned above, the prediction of SCW can be regarded as a classification task with binary categories. A deep learning network for classification is therefore constructed for this purpose. Among various deep learning networks, CNN is a class of deep and feedforward artificial neural networks, which has been successfully applied to many fields, especially image and video recognition (LeCun and Bengio, 1995; Krizhevsky et al., 2012). CNN algorithm can effectively extract twodimensional (2D) features, reduce the number of model parameters, and accelerate the training speed by utilizing receptive fields and weights sharing (LeCun and Bengio, 1995). We constructed deep 2D CNN classification models for binary classification and trained them to predict thunderstorms, HR, hail, and CG.
Our CNN consists of convolution, fully connected layer, and the Softmax classifier. The input of a 2D CNN requires a data array with a format of height × width × channel (channel corresponds to predictors here). Due to the 144 predictors selected for each patch, our input is a threedimensional (3D) array with dimensions 7 × 7 × 144. As mentioned above, these predictors represent the environmental conditions favorable for SCW events.
The core of our processing was carried out by a feedforward stack of five convolutional layers (C1 to C5), followed by one fully connected layer that outputs class scores. Each channel of the five convolutional layers was obtained by convolving the channels of the previous layer with a bank of linear 2D filters such as summing, adding a bias term, and applying a pointwise nonlinearity, as follows：
$${ X}_n^l = {\mathop{\rm Re}\nolimits} {\rm{LU}}\left({{ b}_n^{\left(l \right)} + \sum\nolimits_{k = 1}^K {{ W}_n^{\left({k,l} \right)}} *{ X}_{n  1}^{\left(k \right)}} \right), \;l \in \left\{ {1, \cdots,5} \right\},$$ (1) where ReLU(x) = max(0, x) is the rectified linear unit activation function. The symbol “*” denotes the twodimensional convolution operation. The matrices
${{ W}_n^{\left({k,l} \right)}}$ represent the filters of layer n, and${{ b}_n^{\left(l \right)}}$ the bias for feature map l. Note that a feature map${ X}_n^l$ is obtained by computing a sum of K convolutions of the feature maps from the previous layer.The fifth layer was followed by a fully connected layer with 128 neurons, which transformed the 3D array (
${{ X}^5} $ ) into a onedimensional (1D) array (${{\bar { X}}^5}$ ). Then${{\bar { X}}^5}$ was processed by a linear and fully connected layer to compute the class scores S_{c} with c = 0 or 1 as follows:$${S_{\!\!c}} = \sum\nolimits_{i = 1}^{128} {\bar { X}_i^5} \cdot { W}_{ci}^6 + b_c^6.$$ (2) Finally, we applied the Softmax classifier function to class scores to obtain a properly normalized probability distribution (p), which could be interpreted as a posterior distribution of the two classes given the input
${X^0} $ and the network parameters W and b:$${p_c} = {\rm{P}}\left({{\rm class} = c{X^0},{ W},{ b}} \right) = \frac{{\exp \left({{S_{\!\!c}}} \right)}}{{\exp \left({{X_0}} \right) + \exp \left({{X_1}} \right)}},c = \left\{ {0,1} \right\},$$ (3) where W = {W^{1}, ···, W^{6}} is the set of weights, and b = {b^{1}, ···, b^{6}} is the set of biases.
The structure of the deep CNN is shown in Fig. 2. The input for the model was a 7 × 7 × 144 data array. After the data were imported, they were convolved by five convolution layers that have 256, 256, 512, 256, and 128 filters, respectively. The convolution kernel size was 2 × 2, and a valid mode was applied to every convolutional layer. Because the height and width of the input patch (7 × 7) were small, no pooling layer was utilized. The output was then passed through a fully connected layer with 128 neurons, and the 3D array was transformed into a 1D output array. Finally, the classified probability was calculated by the Softmax classifier function. The number of parameters used in this CNN model was 1,647,362.
Figure 2. Structure of the deep CNN algorithm for SCW forecasting, including five convolutional layers, a fully connected layer, and the Softmax classifier.
We optimized the network parameters by minimizing an L2regularized crossentropy loss function. For optimization, we used the ADAM algorithm (Kingma and Ba, 2015), an algorithm for firstorder gradientbased optimization of stochastic objective functions, which kept track of the first and secondorder moments of the gradients and was invariant to any diagonal rescaling of the gradients. We set the learning rate at 10^{−4} and kept all other parameters to their default values recommended by Kingma (Perol et al., 2018).

After the optimal forecasting model was established, we now make forecast based on the 144 predictors obtained from GFS. To accommodate to the format of the input to the SCW forecasting model, the 144 predictors were normalized and transformed into an M × 7 × 7 × 144 array (M is the number of forecast samples). Probabilistic forecasts were then produced again by the Softmax classifier. Here, probabilistic forecasts were preferred mainly because the mesoγ scale SCW is difficult to observe, especially for the hail and CG, and thus it is challenging to fully understand the phenomena (Cintineo et al., 2014).
Deep CNN training is computationally intensive. Compared to the usually small number of logical CPUs (central processing unit), the GPU (graphics processing unit) used in CNN training is a huge computational matrix with thousands of compute cores. GPUs are able to support parallel computing which is crucial for deep learning because it greatly accelerates the training process (Sanders and Kandrot, 2010). The NVIDIA CUDA (Compute Unified Device Architecture) library and NVIDIA GeForce 1080 Ti graphics chip were utilized in our training and forecasting processes. Tests showed that 0–72h forecasts (at 6h intervals) at 1° × 1° resolution over mainland China can be completed in 3 min, which makes the forecasts practical for operation.

We chose four skill scores to measure the performance of the forecasts: threat score (TS), equitable threat score (ETS), probability of detection (POD), and false alarm rate (FAR), which are defined as follows:
$${\rm POD} = \frac{h}{{{h + m}}},\qquad\qquad\qquad\qquad\qquad\;\;\;\;$$ (4) $${\rm FAR} = \frac{f}{{{h + f}}},\qquad\qquad\qquad\qquad\qquad\;\;\;\;\;\;\;\quad$$ (5) $$ \!\!\!\!\!\! \!\!\! \!\!\! \!\!\! {\rm TS} = \frac{h}{{{h + m + f}}},\qquad\qquad\qquad\qquad\qquad$$ (6) $$\begin{aligned} & {\rm{ETS}} = \frac{{h  {h_{\rm random}}}}{{{h + m + f  {h_{\rm random}}}}}, \\ & \quad {h_{\rm random}} = \left({h + f} \right) \times \left({h + m} \right)/\left({h + m + f + c} \right), \end{aligned} $$ (7) where h is the number of hits, m is the number of missing forecasts, f is the number of false forecasts, and c is the number of correct negatives.
Although the above scores are typically used for evaluating the deterministic forecasts, they can also be used for evaluating the probabilistic forecasts by thresholding the probabilistic forecasts and turning them into deterministic forecasts. Thus, we used those four scores to evaluate the deep CNN forecasting results. After exploring different thresholds for each prediction, we found that the most effective probabilistic threshold values for thunderstorms, HR, hail, and CG were 0.5, 0.5, 0.9, and 0.9, respectively.

In order to compare the performance of deep CNN to that of traditional algorithms, we report the classification performance of various algorithms in Table 2 using the HR test set (592 positive samples and 14,049 negative samples) as well as the hail test set (149 positive samples and 14,492 negative samples). The input of traditional ML algorithms is the 144 predictors at each individual SCW event grid. Note that the logistic regression (LR), RF, SVM, and multilayer perceptron (MP) algorithms are from the scikitlearn package (Pedregosa et al., 2011), and we used GridSearchCV function to conduct exhaustive search over specified parameter values for each classifier.
SCW Algorithm POD FAR ETS TS Heavy rain (HR) LR 0.515 0.570 0.285 0.306 RF 0.499 0.531 0.300 0.319 SVM 0.509 0.543 0.297 0.317 MP 0.526 0.562 0.294 0.314 Deep CNN 0.536 0.504 0.328 0.347 Hail LR 0.178 0.933 0.044 0.051 RF 0.182 0.916 0.054 0.061 SVM 0.185 0.922 0.051 0.058 MP 0.192 0.925 0.050 0.057 Deep CNN 0.213 0.892 0.070 0.077 Table 2. Skill scores for different algorithms using an HR test set (592 positive samples and 14,049 negative samples) and a hail test set (149 positive samples and 14,492 negative samples). The input of CNN is 7 × 7 × 144 data arrays formed by 144 predictors over 7 × 7 patches centered at each SCW event grid, while the input of traditional ML algorithms is 144 predictors at each individual SCW grid
It can be seen from Table 2 that different algorithms have different classification skill. The performance of RF, SVM, and MP is similar in terms of ETS and TS, and is slightly better than that of LR.
Compared to traditional algorithms, the deep CNN model has a deeper network architecture and more model parameters. Owing to the increased complexity, deep CNN’s training time is much longer than simpler algorithms. However, the added complexity can take into account the spatial features of SCW occurrence, leading to improved forecasting performance. As shown in Table 2, deep CNN achieved best performance among all the algorithms for HR and hail forecasting in terms of the four skill scores. Because the comparison results for thunderstorm and CG are quite similar to those for HR and hail respectively, we omitted the results for thunderstorm and CG in Table 2.
In summary, the above results showed that deep CNN algorithm outperforms traditional machine learning algorithms in SCW forecasting over China.

On 21 September 2017, thunderstorms, CG, and hail occurred over a large area in northern China. Meanwhile, a large area in southern China suffered from thunderstorms and HR. The SCW observations and forecasts for this case are shown in Fig. 3.
Figure 3. Forecasts and observations of (a) thunderstorm, (b) HR, (c) hail, and (d) CG on 21 September 2017. Blue shades are objective forecasts, black points are observations of SCW events, and red lines are meteorologist’s subjective forecasts. Only forecasts on land were evaluated. The table in each image lists the number of hits, misses, and false alarms made by either human forecasters (HF) or deep learning (DL) algorithm.
Figure 3 clearly shows that the deep CNN algorithm has a good forecasting skill for thunderstorms, hail, and CG in northern China. Most of the occurrences in the forecasting area were successfully identified based on our forecasts. Moreover, the deep CNN algorithm also appeared to be skillful on forecasting thunderstorms and HR in southern China.
The meteorologist’s forecasts seemed quite different from the objective forecasts. We can see that with the meteorologist’s forecasts there were a large number of false alarms for the thunderstorm, and lots of missing HR, hail, and CG events. The TSs of the deep CNN forecasts of thunderstorms, HR, hail, and CG are 0.48, 0.41, 0.13, and 0.46, respectively, while those of the meteorologist’s forecasts are only 0.40, 0.25, 0.08, and 0.33, respectively.
In summary, for the above typical SCW case, deep CNN forecasts demonstrated much better performance than the forecasters’ forecasts.

The false forecast of SCW that occurred in northeastern China on 2 August 2015 was selected to investigate the reason for false forecast. For this SCW, the forecasts indicated that HR would occur over a large area in the eastern Inner Mongolia, southwestern Heilongjiang, Jilin, and Liaoning provinces. However, observations showed that false alarms of HR were issued over most of the above regions except for eastern Inner Mongolia, central Heilongjiang, southern Liaoning, and some other areas. In order to identify the reason for the false forecasts, two important parameters for HR forecasting, i.e., PWAT and Kindex, were selected to compare the forecasts and observations.
The GFS forecasts at 1400 Beijing Time (BJT) on 2 August 2015 (issued at 2000 BJT 1 August 2015) indicated that PWAT in Heilongjiang, Jilin, and Liaoning provinces would exceed 50 mm and the Kindex would exceed 40°C, suggesting good environmental conditions for HR events. Meteorologists also predicted the occurrence of HR events in the same area and during the same time period. Based on the GFS forecasts, an HR warning was issued over the above regions.
However, a comparison between the NWP forecast fields and the analysis fields indicates big differences between the NWP predictions and observations. Figure 4a shows that the observed PWAT values were significantly lower than the forecasted values, and the predicted values of PWAT in central Jilin and Liaoning were about 8 mm higher than the observed values. In addition, the predicted Kindices were also 3–7°C higher than the observations as seen in Fig. 4b.
Figure 4. Comparison between the NWP forecasts and FNL analyses from NCEP at 1400 BJT 2 August 2015 in terms of (a) PWAT and (b) Kindex. Contours are HR probabilistic forecasts, black dots indicate HR observations, and shaded areas indicate differences between GFS forecasts and FNL analyses, i.e., GFS forecasts minus FNL analyses.
The above results indicate that the water vapor condition and the distribution of Kindex in the NWP forecasts favored the occurrence of HR events. For this reason, a false alarm of SCW was issued. Since the deep CNN algorithm took the incorrect NWP forecasts as its input, it therefore also predicted that HR events would occur over a large area in Northeast China.

During 0200–1400 BJT on 9 July 2015, the deep CNN algorithm predicted that HR events would occur only in central Sichuan Province and its surrounding areas, while observations indicate that HR events actually occurred over a much larger area in eastern Sichuan. As a result, this was a case of missed forecasts.
In order to analyze the causes for the missed forecasts, Fig. 5 shows a comparison between the GFS forecast fields and observations. At 0800 BJT 9 July 2015, one of the SCW conditions in Sichuan was reflected in PWAT, which was about 50 mm, implying an adequate water vapor condition. Meanwhile, the Kindices were 30–35°C, indicating an instability in the atmosphere; the best lift index (BLI) was within −1 to 1, and the energy condition was relatively weak; and the vertical velocity at 500 hPa was from −20 to 10 × 10^{−2} Pa s^{−1}, suggesting a weak dynamic lifting. Overall, the GFS forecasted fields suggested that the SCW environmental conditions were only favorable for the development of weak convection.
Figure 5. As in Fig. 4, but over eastern Sichuan Province, China at 0800 BJT 9 July 2015.
In general, there were obvious differences between the FNL analysis fields and the forecast fields. According to the FNL analysis data, at 0800 BJT 9 July 2015, the PWAT values exceeded the forecasted values by 2–8 mm while the Kindices exceeded the forecasted values by 1–6°C. Although the forecast missed the SCW, surprisingly other features of the forecasts demonstrated a more favorable condition for the development of SCW. For example, BLI values were less than −2, the convective available potential energy (CAPE) values were more than 600 J kg^{−1}, and the welldeveloped dynamic uplifting was found at 500 hPa and even higher levels. Significant differences between the forecasts and analyses are displayed in Fig. 5, which shows that the analysis values of PWAT and Kindex are much larger than the forecasted values in eastern Sichuan where the forecast missed the HR event.
In summary, the GFS forecast fields suggested that the environmental conditions were only favorable for the development of weak convection. In contrast, the analyses indicated that the true environmental conditions were actually very favorable for the development of SCW. The meteorologists also failed to predict the HR events in the eastern part of Sichuan. We found the reason that the deep CNN algorithm missed forecasting the SCW event was because the NWP forecasts as input were only favorable for weak convection.

In order to overall evaluate the performance of the deep CNN algorithm, the forecasts of SCW from April to September in 2015, 2016, and 2017 by both the CNN algorithm and meteorologists were reported (Fig. 6).
Figure 6. Performance diagram of forecasts by the deep CNN algorithm and meteorologists from April to September of 2015, 2016, and 2017 for (a) thunderstorm, (b) HR, (c) hail, and (d) CG. Dashed lines represent bias scores with labels on the outward extension of the line, while labeled solid contours are TS. The red, green, and blue lines indicate the performance of CNN in 2015, 2016, and 2017, respectively. The red, green, and blue triangles indicate the forecast performance by meteorologists in 2015, 2016, and 2017, respectively. The squares indicate the performance of CNN with the determined probability thresholds, which were 0.5, 0.5, 0.9, and 0.9 for thunderstorms, HR, hail, and CG, respectively.
Table 3 shows that the deep CNN algorithm significantly improved the forecasts of all kinds of SCW when compared to forecaster’s forecasts.
SCW Year POD(DL) POD(HF) FAR(DL) FAR(HF) TS(DL) TS(HF) ETS(DL) ETS(HF) HR 2015 0.546 0.338 0.532 0.523 0.337 0.247 0.292 0.211 2016 0.513 0.318 0.515 0.481 0.332 0.246 0.289 0.212 2017 0.589 0.342 0.558 0.466 0.338 0.264 0.290 0.229 Thunderstorm 2015 0.771 0.758 0.480 0.581 0.451 0.370 0.372 0.277 2016 0.760 0.762 0.486 0.560 0.442 0.387 0.363 0.297 2017 0.763 0.731 0.491 0.545 0.440 0.390 0.360 0.303 Hail 2015 0.211 0.147 0.896 0.971 0.075 0.025 0.071 0.021 2016 0.245 0.176 0.913 0.968 0.069 0.028 0.065 0.023 2017 0.249 0.107 0.943 0.981 0.049 0.016 0.044 0.012 CG 2015 0.310 0.149 0.895 0.916 0.085 0.057 0.074 0.047 2016 0.289 0.121 0.892 0.916 0.085 0.052 0.075 0.044 2017 0.246 0.104 0.905 0.920 0.074 0.047 0.063 0.039 Table 3. Evaluation of deep CNN (DL) forecasts and human forecasts (HF) from April to September of 2015, 2016, and 2017 (for 12h forecasts initialized at 0800 BJT)
The average TS of HR forecasts by deep CNN algorithm was 0.336, which showed an increase of 33.2% over the TS of 0.252 for subjective forecasts. It is worth mentioning that the miss rate (1 − POD) of forecaster’s forecasts was larger than the FAR, which means that a large number of heavy rainfall events were missed. In contrast, the opposite was true for forecasts by the deep CNN algorithm; that is, fewer missing forecasts were made by CNN. Thus, the deep CNN algorithm provides valuable guidance for meteorologists in their operational forecasts.
The deep CNN algorithm has stable performance in thunderstorm forecasting. The average TS of forecasts by the algorithm exceeded 0.44 each year, which is on average 16.1% higher than the score of meteorologist’s forecasts. The deep CNN algorithm had a higher POD and lower FAR, implying more reliable forecasting.
The performance diagrams are given in Fig. 6. Forecasts of hail and CG were greatly improved by the deep CNN algorithm in comparison to meteorologist’s forecasts. The average TS of hail and CG forecasts by the algorithm were 0.064 and 0.081, respectively. This indicates 178% and 55.7% improvements to the meteorologist’s forecast scores of 0.023 and 0.052. The deep CNN algorithm also had better performance in terms of POD and FAR. Because hail and CG are incompletely observed due to their local distribution, the forecasts from both the algorithm and meteorologists had high FAR value.
In summary, deep CNN algorithm showed higher capability in all four types of SCW forecasting. Compared to the forecaster’s forecasts, the deep CNN forecasts showed notable improvements in both qualitative and quantitative evaluations, suggesting that the algorithm has much better overall performance. Nevertheless, there are still inadequacies in the algorithm as it issues too many false alarms of hail and CG, a problem also found in meteorologist’s forecasts. Thus, a key goal of improving the algorithm in the future will be reducing the FAR of forecasts.
It is worth mentioning that we also pinned down to the performance of deep learning at hotspots, and we found that the deep learning has better forecasting capability at hotspots than in other areas. For example, the TS of HR is 0.30 in North China in 2017, while 0.39 in South China where hotspots are observed. Similar pattern is also observed for other types of SCW events.
Feature  Level (hPa)  
Multilevel variable  T (temperature) H (geopotential height)  1000, 925, 850, 700, 600, 500, 400, 300, 200 
WS (wind speed) WD (wind direction) W (vertical wind speed)  
TDD (temperature dew point difference) Q (specific humidity) VAPFLUXDIV (water vapor flux divergence)  
PV (potential vorticity) TMPADV (temperature advection) SITASE (potential pseudoequivalent temperature) DIV (divergence) VOR (vorticity) VORADV (vorticity advection)  
Singlelevel convective parameter  MUCAPE (most unstable convective available potential energy) BLI (best lift index) CIN (convective inhibition) DCAPE (downdraft convective available potential energy) K (Kindex) LI (lift index) Z0 (altitude of 0°C) Z20 (altitude of 20°C) PWAT (precipitable water) SHIP (significant hail parameter) SHEAR1 (0–1km wind shear) SHEAR3 (0–3km wind shear) SHEAR6 (0–6km wind shear) SI (Showalter index) TT (total index)  
Others  Elevation Longitude Latitude 