Recognition of Organizational Morphology of Mesoscale Convective Systems Using Himawari-8 Observations

Mesoscale convective systems (MCSs) are a substantial contributor to flash flooding, hailstorms, wind gusts, and tornadoes (Maddox et al. ^[1]; Moore et al. ^[2]; Schumacher and Johnson ^[3]; Houze ^[4]). They are usually composed of cumulonimbus clouds and can cause an extensive range of precipitation of thousands of square kilometers. Since the last century, much in-depth work has been done on the initiation and development mechanisms of mesoscale convective systems, and scientists have consistently found that the size, organizational mode, and propagation speed of MCSs are crucial factors in determining their ability to produce heavy precipitation (Parker et al. ^[5]; Schumacher and Johnson ^[6]; Storm et al. ^[7]; Gallus et al. ^[8]).

Houze once stated that "MCSs exhibit a variety of cloud and precipitation structures" and "when the individual cumulonimbus clouds and/or lines of cumulonimbus group together in these cloud systems, additional phenomena appear" (Houze ^[4]). These remarks suggest that the research on the morphology of MCS organization, which is linked to large-scale forcing and microphysical processes, would be vital to improving the understanding of mechanisms of the initiation, development, and propagation of convective systems.

Currently, a variety of MCS organizational modes that can cause extremely heavy precipitation have been found worldwide. For example, Houze et al. ^[9] discovered that the MCSs responsible for heavy rainfall in Oklahoma, United States, typically had a linear structure. Pettet and Johnson (hereinafter PJ03) documented that the leading stratiform rain (LS) and the parallel stratiform rain (PS) patterns were two dominant modes in causing flooding ^[10]. Schumacher and Johnson ^[6] investigated the morphology of the MCSs that were associated with severe precipitation events in the United States from 1999 to 2001. They found that the MCS in these precipitation events were generally depicted as two archetypes: the training line/adjoining stratiform and the back-building/quasi-stationary systems. On the basis of radar observations, Gallus et al. ^[8] (hereinafter G08) classified the morphology of the convective system organization related to severe weather hazards in North America into nine categories: individual cells (IC), a cluster of cells (CC), broken squall lines (BL), nonlinear convective systems (NL), trailing stratiform (TS), squall lines without stratiform rain (NS), bow echoes (BE), leading (LS) and line-parallel stratiform rain (PS). Jirak et al. ^[11] investigated hundreds of MCSs during the warm season in the central United States using satellite and radar data. According to the study, the organization of MCSs as defined by satellite characteristics can be classified into mesoscale convective complexes (MCCs), persistent elongated convective systems (PECSs), meso-β circular convective systems (MβCCSs), and meso-β elongated convective systems (MβECSs). Nowadays, thanks to the improved global observation capabilities, extensive studies have been conducted on the MCSs all over the globe, including some regions in China. For example, Wang and Cui ^[12] provided insights into the MCS organization modes in the middle and lower reaches of the Yangtze River. They recognized MCS with a TS pattern as the most common producer of heavy precipitation in this area. Zheng et al. ^[13] categorized MCSs in central East China into seven different morphologies. They also found that among them, MCSs with the TS and/or PS modes had longer lifespans. Chen et al. ^[14] documented the occurrence of a linear convective system with a length of more than 300 km during heavy rainfall over coastal regions of South China. Later, Li et al. ^[15], Xue et al. ^[16], and Zhang et al. ^[17] studied convective system organizational structure in South China using long-term radar composite reflectance data. They found that the organizational structure of the convective system in South China is different from that of central East China due to differences in synoptic environments. The most common convective organizational structure mode that causes severe weather in South China is linear structure. However, different weather types may be associated with specific linear structure modes. For example, Xue et al. ^[16] found that the storms with LS and PS modes are dominant in producing extremely intense rainfall. Zhang et al. ^[17] proposed that TS and BE modes are associated with short-duration heavy rainfall and severe convective winds in South China. Zhang et al. ^[18] examined the organizational archetypes for the heavy-rainfall-producing MCSs in eastern China. It was shown that instead of linear convective systems, multiple MCSs and non-linear convective systems are the two most common patterns in this region. The morphology of MCS organization is generally affected by large-scale forcing. Since there may be differences in the configurations of circulations in particular areas, it is not surprising that there is significant variability in the organizational structures of extreme-rain-producing MCSs. The characteristics of MCSs that are summarized in particular regions may not be universally applicable worldwide. So far, the organizational modes for the MCSs in Northwest China have received little attention compared with the other areas around the globe.

Northwest China, including Shanxi, Gansu, Ningxia, Qinghai, and Inner Mongolia, is located on the northeastern slope of the Qinghai-Tibet Plateau, which features a semi-arid-semi-humid climate. The weather here is often complex and changeable, which poses challenges in accurately forecasting and providing timely weather alerts (Wang et al. ^[19]; Wang et al. ^[20]; Yang et al. ^[21]). Due to the limited observation conditions in the past, there have been few studies on the storm morphology over this area. Unlike South and East China, Northwest China experiences weaker monsoon moisture surges, but it is more affected by cold air outbreaks over mid-high latitudes and the Tibetan Plateau. Whether such differences in environmental conditions are responsible for the distinct convective organizational pattern is worth further exploration.

A mesoscale convective system is mainly composed of convective and non-convective zones, in which the convective zone comprises single or multiple convective cells, and the non-convective part is formed by stratiform precipitation regions and non-precipitating anvil clouds. Therefore, the morphology of MCS organization can be depicted by distinguishing between convective and non-convective regions in a mesoscale convective system. Currently, the classification can be made based on radar and satellite data. Due to the different principles of these two types of remote sensing instruments, the classification methodology and criteria related to each of them are quite different (Jirak et al. ^[11]).

So far, most of the documented organization archetypes of MCSs have been recognized and generalized using radar observations. The idea of this method is based on the relationship between the concentration of precipitation particles and radar reflectivity (Bluestein and Jain ^[22]; Bluestein et al. ^[23]; Blanchard ^[24]; Loehrer and Johnson ^[25]). Since radar echoes represent scattered precipitation particles, a high value in reflectivity can indicate dense precipitation particles in the column, which are highly correlated to the intensity of upward motion. Therefore, radar observations offer a distinct advantage in understanding the structure of precipitation systems. However, the reliability of MCS morphology identification based on radar data is often diminished in mountainous regions or locales with limited radar coverage due to geographical limitations that impede comprehensive radar observation.

In terms of space coverage, satellites have a clear advantage over radar. They can obtain continuous spatiotemporal variations about an MCS. Traditional satellite-defined MCS organization is determined by the correlation between cloud optical properties and the radiation temperature revealed by the long-wave infrared channel. For developing convective clouds, due to the obstruction of the tropopause, most water vapor within convective cores will be frozen and diverged to form broad cloud shields after being rapidly lifted to the upper troposphere. The optical thickness and coldness of these broad cloud shields typically result in a wide area of bright white in the satellite's infrared images. Non-convective clouds can blur the embedded convective cores on infrared images due to their similar brightness temperatures to the cloud anvils. Consequently, the delineation of MCS organization by satellite is not solely reliant on infrared brightness temperature. Instead, it is a comprehensive classification that also takes into account the size, shape, coldness, and duration of the cloud tops as observed by the satellite.

Due to the use of different classification criteria, it is currently impossible to compare the organization modes defined by satellite and radar data (Jirak et al. ^[11]). For MCSs that occur in regions with sparse radar stations, it would be difficult to understand the fine organization and the underlying mechanism of MCS systems. To fill this gap, the present study aims to create an objective satellite-based MCS segmentation and classification method that adheres to the category criteria based on radar observations. Then, we will apply this method to characterize the organizational morphology of a typical heavy-rainfall-producing MCS in Northwest China to expand our understanding of the organizational structure of mesoscale convective systems in areas with limited radar observation.

The structure of the subsequent sections of this article is outlined as follows. The data and method used in this paper are presented in Section 2. A preliminary validation is given in Section 3, followed by an application of the algorithm to identify the organizational structure of a typical MCS event in Northwest China. A synthesis of findings and conclusions are provided in Section 4.

The MCS segmentation and classification algorithm built using satellite observations was mainly based on the Advance Himawari Imager (AHI) onboard Himawari-8. Himawari-8 is the initial Japanese geostationary meteorological satellite of its new generation. The AHI onboard is an ABI-class (ABI: Advanced Baseline Imager) imager. Currently, it has 16 channels, with three visible channels, three near-infrared channels, and ten infrared channels. The general horizontal resolution of sub-satellite points for the infrared channel is 2 km, and a full disk observation can be obtained every 10 minutes. Such high-resolution and spatio-temporally continuous observational data can be beneficial in recognizing the morphological structure and evolution of a convective system.

The Himawari-8 AHI data used in this study were obtained from the Japan Aerospace Exploration Agency (JAXA) Himawari Monitor (P-Tree System). A subset of four infrared channels and a channel combination defined by two AHI infrared channels were selected to build the algorithm. Table 1 shows the physical bases of these channels.

This study used radar composite reflectivity data from central East China during July 1–10, 2020, with a 6-minute interval and a spatial resolution of 1 km, to validate the algorithm. The data were selected based on the following two factors. First, the terrain in central East China is relatively flat, and the radar echoes are less affected by signal blockage issues. Furthermore, S-band radar is the primary radar in this region, and it provides reliable data quality. Second, mesoscale convective systems are more active in July, which means there is a high chance of occurrence for various types, which can help better test algorithm performance. Central East China covers the range of 27°–40°N, 108°–123°E, including Jiangsu, Zhejiang, Anhui, Fujian, Jiangxi, Shandong, and Shanghai, approximately 1300 km ×1500 km. The focus of this article was meso-β to meso-α scale convective systems, with a spatial scale of 20 to 1000 km and a temporal scale of one hour to one day. During July 1–10, 2020, 50 MCSs were collected in this region. All the data from these cases' entire life cycle were utilized for evaluation.

Currently, the primary radar data types used to distinguish the morphological structure of MCSs are the level-Ⅱ mosaic of the base reflectivity data (Hilgendorf and Johnson ^[28]; Hane et al. ^[29]) and composite reflectivity data (Schumacher and Johnson ^[6]; Parker and Johnson ^[30]). The latter is made up of the maximum value at which the radar receives echoes reflected by clouds at different heights. Compared to base reflectivity, composite reflectivity can better reflect information within storm systems. Severe convection or intense precipitation is often associated with strong echoes. Due to its enhanced capabilities, composite reflectivity is used more in studying MCSs. In order to facilitate a comparison with satellite-based MCS classifications, we adopted a classical radar-based MCS classification scheme proposed by Parker and Johnson ^[30] (hereinafter PJ00). PJ00 defined MCS as a continuous echo area with echoes over 20 dBZ, inclusive of a line or area of convection (≥40 dBZ). Within this area, stratiform precipitation corresponds to the continuous echo zone between 20–40 dBZ, while the convective and deep convective zones correspond to echo zones over 40 dBZ and over 50 dBZ, respectively.

To test the efficacy of the algorithm presented in this study, we applied the algorithm to recognize the organizational structure characteristics of an MCS that occurred during an extremely heavy rainfall process on August 21, 2016, in the eastern foothills of Helan Mountains, Ningxia Province. Fig. 1 shows the topography of the area where the rainstorm occurred. The subgraph reveals the locations of three ground observation stations in the rainfall center. These stations can only collect accumulated rainfall for a interval of 5 minutes.

Figure 1.  Topography (units: km) of the studied domain and the locations of observation stations (the red dots in the sub-graph) in the eastern foot of Helan Mountains in Ningxia.

Cumulus clouds develop into cumulonimbus clouds through the development of organized updrafts and downdrafts within the cloud, a transition that typically signifies the onset of convection (Müller et al. ^[31]). The convective core is associated with regions of intense upward motion, which causes the cloud tops to always appear to be rapidly bubbling, as demonstrated in the three-dimensional (3D) model in Fig. 2a. In contrast, those over the non-convective area with weak updrafts are relatively flat (Houze ^[4]). Given this, the identification of organizational morphology could be considered as distinguishing actively growing convective cores in cumulonimbus clouds.

Figure 2.  (a) A typical 3D shape of a rapidly growing convective cloud top; (b) a schematic diagram of a real T_b observed in infrared or WV channels and its corresponding Gaussian distributions for a rapidly growing convective cloud top.

During the convection, a great deal of water vapor soar upward by strong updrafts and eventually condense into cloud water droplets aloft. More radiation being emitted from the water vapor above the cold cloud shield is observed by satellite due to the optical thickening of the clouds. Water vapor tends to emit radiation at cooler (warmer) temperatures in the troposphere (stratosphere) as temperature decreases or increases with height in the troposphere (stratosphere). Thus, before the observed cloud top reaches the tropopause, a low value in the brightness temperature (T_b) of water vapor channels can indicate dense precipitable water particles in the column (Müller et al. ^[32]). In that case, the developing convective cores corresponding to strong vertical updrafts may be expected to have local minimum values in the water vapor (WV) channels. In an ideal situation, as shown in Fig. 2a, the T_b around the convective cells will exhibit an inverted Gaussian distribution with the local minimum in the center (solid green line in Fig. 2b). That is, the T_b at the center of the convective core is the lowest in all directions. Away from it, the T_b will gradually increase. In other words, if the T_b of the center point and its radial points within a window satisfy this two-dimensional Gaussian distribution, it always indicates that the T_b at the center point is a local minimum and is observed at a higher height than surrounding pixels.

As shown in Table 1, there are three water vapor channels on Himawari-8 AHI, channel 8 (6.2 μm), channel 9 (6.9 μm), and channel 10 (7.3 μm). With the peak weighting function levels at ~375, ~450, and ~600 hPa, respectively (Zou et al. ^[33]), these channels can detect water vapor variations in the upper, middle, and lower troposphere. Once the T_b values within a window in the three water vapor channels approximate Gaussian distributions, the T_b minimum points within the window may be candidates for actively growing convective cores. The candidate points are then tracked to measure the variation trend. If the T_b values at the candidate point have a continuous cooling trend, the grid point and its neighboring grid points in the window are identified as a convective core (Lee et al. ^[34]). The detailed procedure, as shown in Fig. 3, includes three parts, which are described in the following sub-sections.

Figure 3.  A flowchart describing the detection of rapidly growing cloud tops in cumulonimbus clouds.

Before identifying organization morphology, cumulonimbus cloud targets have to be separated. This paper first employed the 240 K threshold for the 11.2-μm infrared channel to filter out thick cumulus objects (Zhang et al. ^[35]). To assume the selected thick cumulus objects to be the cumulonimbus clouds, two criteria were further checked: (1) a continuous area over 5000 km² and (2) the expansion speed of 100 km² h^–1, which were introduced by Pope et al. ^[36]. Those not satisfying the criteria were marked as non-cumulonimbus clouds.

Active growing convective cores can be detected by observing the features that these cloud tops typically exhibit, such as rapid cooling due to strong updrafts. To distinguish between growing convective clouds and non-convective clouds, a measure of the temporal variations of the brightness temperature in the water vapor channels can be applied (Müller et al. ^[32]; Lee et al. ^[34]). In this paper, an 11×11-pixel window was set to cover convective cores smaller than 20 km in diameter. The T_b values within the window were normalized to fit a typical Gaussian distribution. The normalization adopted the maximum-minimum method, which first calculated the maximum T_b and minimum T_b values within the 11×11-pixel window and subtracted the maximum value from each pixel to obtain T_b'. The T_b' of each pixel was then divided by the maximum and minimum brightness temperature differences. As shown in Fig. 2b, the black line represents the T_b matrix of a given channel, and the green curve represents the Gaussian matrix that corresponds to it. If two matrices were similar, the T_b matrix can be considered to have a Gaussian distribution, and the minimum T_b within the window may be the candidate growing top.

The channels used for detecting growing convective cloud tops were selected from three WV channels (Table 1). Moreover, the brightness temperature difference (BTD) between 6.2 μm and 11.2 μm was calculated to aid in the identification of extremely vigorous convective cores. According to the physical bases of the two channels, the 6.2-μm channel reflects the water vapor content within the upper troposphere, while the 11.2-μm channel indicates cloud top height. The combination of these two channels can be utilized to estimate whether a cloud has grown to reach or above the tropopause. Generally, the value of BTD between 6.2 μm and 11.2 μm is negative for the clouds lower than the tropopause height. It can increase to near zero or positive values when the cloud tops reach or exceed the tropopause. To objectively select the threshold, multiple tests were conducted on the threshold combination of BTD and the difference between T_b and Gaussian matrices. The threshold selection is primarily determined by the accuracy of flow identification, as it considers the impact on subsequent convective identification. The probability of detection (POD) and accuracy (ACC) statistics were calculated by comparing radar-recognized convection. The equations for POD and ACC are as follows:

where TP represents radar and satellite detected as convective; TN denotes both radar and satellite assigned as non-convective. FP represents radar recognized as non-convective, while satellite recognized as convective. FN denotes radar identified as convective, while satellite missed.

Table 2 shows the accuracy comparison using different threshold combinations. It can be seen that the thresholds of –3 K for the BTD and 10 K for differences between T_b and the Gaussian matrices had the best probability for detecting convective among the test sets. Based on the results, pixels whose differences between T_b and the Gaussian matrices were smaller than 10 K and BTD larger than –3 K were set as growing cloud tops candidates for determining the convective cores in the next step. The pixels filtered out were assigned as stratiform and anvil regions.

The deep convective zone is often connected to the most intense vertical movement within the cloud, which is the core area within the system and the key to maintaining the entire convective system. The T_b of a deep convective core usually has the characteristic of continuous cooling with time (Zinner et al. ^[37]). Based on this, this paper tracks the growing cloud tops selected in the previous step and calculated the brightness temperature changes of these pixels in the past 30 minutes. The dense inverse search-based (DIS) method proposed by Kroeger et al. ^[38] was adopted to track candidate objects. The DIS method used a reverse search and gradient descent scheme to estimate the target's movement. By integrating the ideas of sparse and dense optical flow schemes, tracking efficiency and accuracy can be significantly improved. The prerequisite for this method is that the tracked MCS has relatively small changes in 10 minutes, even during the merging and splitting stages. If the cloud clusters in the two consecutive images have the same area overlap, they are considered to be the same MCS. In this study, the growing cloud pixels were tracked for 30 consecutive minutes in three WV channels and the 11.2-μm channel, with an average interval of 10 minutes. If the averaged temperature change rates at each channel during 30 minutes were negative, the candidate cloud pixels and their neighbors were determined to be a deep convective zone.

Since deep convective zones are always surrounded by moderate convection, the moderate convective areas can be derived from the spatial expansion of deep convective zones. To do so, the identified intense convective zones and non-convective areas, including stratiform and anvil regions, were first labeled as 2 and 1, respectively, to form a segmentation image. Then, use a 5×5 dilate kernel given in Eq. 3 to scan the segmentation image and compute the maximum pixel value overlapped by the kernel (Gonzalez and Woods ^[39]). If the value of the center pixel in the 5×5 window was smaller than the maximum, it would be replaced by the maximum value. Finally, all the pixels with values of 2 were classified as convective regions. The convective system as a whole is thus divided into deep convective, convective, and non-convective parts.

To verify the credibility of the results obtained by this algorithm, radar-based MCS classification results during July 1–10, 2020, were adopted in the qualitative and quantitative validation.

Qualitative validation was conducted based on three cases that occurred on July 1, 3, and 7, 2020. The MCS that occurred in northern Hunan on July 1, 2020, caused hailstorms, short-term heavy rainfall, and thunderstorms (Fig. 4). Radar and satellite observations indicate that the highest echoes for this system were above 50 dBZ (Fig. 4a), and the lowest T_b was near 200 K (Fig. 4c). A linear arrangement was present in the structure of this MCS derived from both types of data (Figs. 4b and 4d). The convection was strong at the eastern and western ends but relatively weak in the middle of the systems. The satellite-based MCS results were more prominent in terms of these features, with convection almost unseen over this place.

Figure 4.  Comparison of radar and satellite-based algorithms for segmenting a candidate MCS at 00:00 UTC on July 1, 2020. (a) Radar composite reflectivity, (b) radar-based MCS structure segmentation, (c) brightness temperature (color shadings are T_b≤240 K) of 11.2-μm channel observed by the Himawari-8 AHI, and (d) satellite-based MCS structure segmentation.

The second case was an MCS that affected the areas from Jiangxi to Zhejiang provinces on July 3. The system brought over short-term heavy rainfall of over 50 mm h^–1 to local areas of Jiangxi and Zhejiang provinces during the period of 05:00–06:00 UTC. From the radar reflectivity image and the classification result (Figs. 5a and 5b), it can be seen that the convective area with strong echoes was mainly concentrated on the northern edge of the system, with a linear distribution. In this instance, the stratiform precipitation area was bigger than the convective zone by more than ten times, which shows LS structural characteristics. However, this structural feature was difficult to see on the infrared cloud image (Fig. 5c). By using the traditional method of identifying convective and deep convective regions based on infrared images, deep convective regions with brightness temperatures below 221 K (–52℃) account for almost half of the entire MCS area. Instead of being linear, its shape was elliptical. These distribution characteristics were markedly different from the MCS structure observed by radar. However, on the MCS classification map obtained using the algorithm in this study (Fig. 5d), not only the area ratio but also the position of the convective zone relative to the stratiform precipitation area were closer to the radar results.

Figure 5.  Same as Fig. 4 but for the candidate MCS at 05:00 UTC on July 3, 2020.

The third case was a squall line process in Anhui Province on the afternoon of July 7. A short-term intense rainfall of 87.1 mm h^–1 occurred in central Anhui at 09:00 UTC as a result of this squall line. The radar echo of the MCS was higher than 50 dBZ in Figs. 6a and 6c, but the brightness temperature on the infrared image was not as low as those in the first two examples. On the radar classification map, the morphological structure of the MCS was depicted with convective areas arranged in a linear pattern from northwest to southeast. By comparison, although satellites identified fewer convective zones than radar, their distribution patterns were similar.

Figure 6.  Same as Fig. 4 but for the candidate MCS at 09:00 UTC on July 7, 2020.

Quantitative validation was first performed by verifying the accuracy of detection of the convective area. Validation was conducted from July 1 to July 10, 2020. Besides POD and ACC, the false alarm rate (FAR) was also calculated, which is defined as:

Deep convection and convective zones were combined to form convective zones for validation. In the calculation, the POD was 0.65, FAR was 0.39, and ACC was 0.94. These accuracy levels were comparable to those of some deep-learning algorithms. For example, Duan et al.^[40] used Himawari-8 observations to reconstruct the radar reflectivity of convective storms using deep learning. According to their evaluation, the POD for the reflectivity over 35 dBZ (indicating convection and deep convection) was 0.45, while FAR was 0.67 and ACC was 0.97. GOES-R observations were used by Hilburn et al.^[41] to suggest a deep-learning method for estimating radar reflectivity. Their accuracy in identifying convective areas was no higher than 0.65, and the false alarm rate was around 0.67. It is possible that a parallax problem in satellite observations is the cause of a moderate POD but a relatively high FAR. As claimed by Di et al. ^[42], the parallax is more severe for high cloud top heights. When cloud heights reach over 10 km, the position shift caused by parallax can reach 50 km, which is around 25 pixels for a 2 km resolution.

Since we were more focused on the structural features, it was necessary to evaluate the similarity between the MCS organizational structures depicted by the two different observational data in addition to assessing the pixel-level accuracy. The Structural Similarity Index Measure (SSIM), an objective image evaluation tool (Zhou et al. ^[43]), was used here. The range of this measure is between 0 and 1, and the closer it comes to 1, the more similar the two structures are. The equation for SSIM is given as follows:

where $\mu_x$ and $\mu_y$ are the mean values of the organizational structure of an MCS recognized by satellite and radar, and $\sigma_x$ and $\sigma_y$ represent the corresponding variance values, $\sigma_{xy}$ is the covariance of the structure images obtained by satellite and radar, C₁ and C₂ are constants, taking as 1×10^–4 and 3×10^–4. We tested SSIMs for 50 MCS samples. Fig. 7 shows the distribution of SSIM obtained from 50 examples, with a median value of 0.58 and mean value of 0.6, indicating that the organizational structure features of MCS identified by satellites are generally similar to those of radar. The three cases depicted above had SSIMs of 0.58, 0.68, and 0.98, respectively.

Figure 7.  Violin plot of the structure similarity index measure between morphological structure recognized by satellite and radar for the selected 50 MCSs. The numbers on the top, middle, and bottom are the maximum, median, and minimum values of SSIM.

The MCS structure obtained by the algorithm in this study was comparable to that identified by radar, as demonstrated by the validation. Afterwards, we applied it to recognize an MCS in Northwest China.

The selected case was an extreme rainstorm on August 21, 2016, over the southern part of Inner Mongolia and the Ningxia Helan Mountains. It began in the afternoon and developed into severe precipitation at midnight. The process lasted 14 hours to achieve a maximum cumulative rainfall of 231.8 mm and an hourly rainfall intensity of 82.5 mm h^–1, respectively. The extremely heavy rain occurred in Helan, Huaxuechang, and Baisikou at the foot of the Helan Mountains, approximately 20 km east of the main peak of Helan Mountains (as shown in Fig. 1).

As shown in Fig. 8, heavy precipitation exceeding 20 mm h^–1 mainly occurred between 12:00 UTC and 18:00 UTC, during which the intensity of the rainfall fluctuated. Huaxuechang experienced the greatest rainfall intensity of any location during this period, reaching 82.5 mm h^–1 at 17:00 UTC on August 21.

Figure 8.  Time series of the hourly rainfall (units: mm h^–1) recorded at the Huaxuechang and Baisikou stations around the rainfall center.

The subtropical high expanded abnormally westward during the process. At 00:00 UTC on August 21, the western ridge point of subtropical high extended to around 100°E, with the northern edge between 35°–38°N. At the same time, at 700 hPa, a low-pressure system developed in the northeast of Qinghai strengthened and constantly moved eastward. In the wind field, the low-level wind was accelerating, a southeasterly jet was formed along the southwest side of the subtropical high between 700 hPa and 850 hPa, and the rainstorm area appeared near the convergence line at the front of the jet (not shown).

The description above indicates that the rainstorm process was divided into three stages: initiation (10:00–13:00 UTC), development (14:00–17:00 UTC), and weakening phase (18:00 UTC on August 21 to 00:00 UTC on August 22). The convective cloud structures from 10:30 UTC to 13:30 UTC with an hour interval are shown in Fig. 9. The isolated cell structure at the initiation stage of the convective system is visible in the figures. At 10:30 UTC, the embedded convective core was discovered for the first time in the northwest part of the cloud (Fig. 9a). The convective core area was only 10 km² at this time, despite the cloud's minimum T_b reaching –65℃ (~208 K). By 11:30 UTC, the coverage of the convective core had expanded by nearly ten times. The minimum T_b at the cloud top was close to –71℃ (~202 K) at this time. The outflows caused by intense updrafts resulted in a significant expansion of the cloud top area. Since 12:00 UTC, the number of convective cores within the MCS was increasing. At 12:30 UTC, the previously occurring convective cores weakened, and a new one appeared to the east. The convective system's transition from its initiation to rapid development was marked by a significant increase in the number of convective cores in clouds. At 13:30 UTC, a dozen convective cores appeared. Meanwhile, the continuous area of the cloud top with a T_b value below –65℃ reached ~40, 000 km², showing a nearly circular structure (Fig. 9d). From west to east, there were multiple convective cores, and the easternmost one lied above the Helan Mountains in Ningxia. Due to the cores' close location, the cloud anvils around them overlapped with each other aloft, thus creating a more extensive range of continuous cloud shields.

Figure 9.  Himawari/AHI 11.2-μm brightness temperature imagery (units: K, T_b≤240 K) along with the convective and intense convective areas (black thick dashed contours) on the mesoscale convective system in Northwest China at (a)10:30 UTC, (b) 11:30 UTC, (c)12:30 UTC and (d) 13:30 UTC on August 21, 2016.

Figure 10 displays the progression of the convective system's morphology over a 10-minute period during the vigorous development stage. At this stage, the convective system was almost stationary due to its slow movement. Along the axis of the cloud, there was a distribution of convective cores that quickly dissipated in the local area. The average life cycle of a single convective core was ~20 minutes, with the longest-lived at the western end and the shortest-lived at the eastern end of the cloud system. When convective cores were built forward and backward on the eastern and western sides of the cloud, the convective cores within the entire cloud system exhibited a linear distribution. At 15:50–16:00 UTC, a few nascent convective cells to the west of the convective cloud rapidly developed (Figs. 10c and 10d). At around 16:10 UTC, they were connected to the main cloud system (Figs. 10e–10h).

Figure 10.  Same Fig. 9, but for the storm at (a) 15:30 UTC, (b) 15:40 UTC, (c) 15:50 UTC, (d) 16:00 UTC, (e) 16:10 UTC, (f) 16:20 UTC, (g) 16:30 UTC, and (h) 16:40 UTC on August 21, 2016, when it was at the development stage.

After 18:00 UTC, the precipitation intensity weakened significantly, indicating that the MCS had entered a weakening period. As illustrated in Fig. 11, the organizational structure of MCS during this period tended to have the following two characteristics: (1) the coverage of the cloud anvil of the convective cloud reached nearly 300, 000 km², while the intense convective areas indicated by convective cores accounted for no more than 2%; (2) convective cores became very dispersed and irregularly arranged.

Figure 11.  Same as Fig. 9, but for the storm at (a) 18:30 UTC, (b) 19:00 UTC, (c) 19:30 UTC, and (d) 20:00 UTC on August 21, 2016, when it was at the dissipation stage.

To verify the credibility of the mesoscale convective system structure identified in the "0821" process, we compared the results with the Dual-frequency Precipitation Radar (DPR) carried on the Global Satellite Precipitation Program (GPM). The GPM program is an innovative satellite observation venture that is managed by the National Aeronautics and Space Administration (NASA) and the Japan Aerospace Exploration Agency (JAXA). The program's initial satellite was launched in February 2014. The satellite is low-orbit and only passes through the same area twice daily. Precipitation radar on the satellite is known as DPR. It has two bands: the Ku-band at 13.6 GHz and the Ka-band at 35.5 GHz. The Ku-band is more resistant to rain attenuation than the Ka-band and has a higher accuracy rate for detecting heavy rain than the Ka-band. The data's scanning width is 245 km, and the detection height is 22 km from the surface to the sky. The horizontal resolution is approximately 5 km, and the vertical resolution is about 250 m.

During the "0821" rainstorm, only one scan track passed through the rainstorm area. The time was at about 19:40 UTC on August 21, 2016, corresponding to the dissipating stage of the convective system. Fig. 12 shows the retrieved attenuation-corrected radar reflectivity factor, storm top height, precipitation types, and organizational structure of MCS based on radar reflectivity. It can be seen that the radar reflectivity factors above 40 dBZ exhibited nonlinear characteristics, with the highest storm top reaching about 13 km on the southwest part of the convective system. Moreover, both the retrieved precipitation types and the MCS organizational structure obtained based on attenuation-corrected radar reflectivity factors showed that the stratiform precipitation region of this convective system was mainly located in advance of the convective precipitation area. The organization structure of the convective system at the corresponding time obtained by the algorithm in this paper is displayed in Fig. 13. The main difference between the two is that the stratiform precipitation area of the convective system identified based on the optical imager onboard Himawari appeared larger than the results obtained by GPM precipitation radar. The part that exceeded was recognized by GPM as a mixed-type precipitation area. This result is because the algorithm in this paper is mainly based on the structural characteristics but the phase of the cloud tops, so it cannot further distinguish mixed precipitation areas. Except for that, both GPM-based and Himawari-based results suggest that the convective system had non-linear structural characteristics, with the convective cores mainly concentrated at the southwest end of the convective system and the stratiform area located in advance of the convective cores in the convective system. The above results indicate that the algorithm proposed in this article accurately represented the characteristics of the convective organizational structure during the "0821" process.

Figure 12.  The GPM/DPR Ku-band retrieved (a) attenuation-corrected radar reflectivity factor (shadings, units: dBZ), (b) storm top height (shadings, units: km), (c) MCS organizational structure based on attenuation-corrected radar reflectivity factors, and (d) precipitation type classification (the black solid lines in figures indicate the scan track edges) at 19:40 UTC on August 21, 2016.

Figure 13.  Himawari/AHI 11.2-μm brightness temperature imagery (units: K, T_b≤240K) along with the convective and intense convective areas (black thick dashed contours pointed by the black arrows) on the mesoscale convective system at 19:40 UTC on August 21, 2016. The black solid lines in the figure indicate the GPM scan track edges.

According to the prototypes summarized by G08, the MCS discussed in this paper exhibited isolated cells (IC), squall lines with parallel stratiform rain (PS), and nonlinear system (NL) structural morphology in its life cycle (Gallus et al ^[8]). Fig. 14 shows the life cycle scenarios of the organizational structure variations in the MCS. The MCS's evolution was initiated by an isolated cell mode. After time went by, during the development stage of the rainfall, it gradually converted into a typical parallel stratiform structure, with a group of convective cells in the middle of the cloud arranged in a line. However, the PS structure was not quite stable. It would turn to an NL structure for a short time when the cores forming on the east end of the line weakened. A continuous process of alternating PS and NL structures in the rapid development stage resembling order-disorder transformation suggests an energy dispersion process within the system. Once such a transformation disappears, it may indicate an equilibrium in energy budgets. The weakening stage of the system occurs when the coverage of the non-convective region expands while convective cores are dispersed and irregularly arranged.

Figure 14.  The evolution pattern of the organizational morphology of the extreme-rain-producing mesoscale convective systems occurred in Northwest China on August 21, 2016.

PJ03 pinpoints LS and PS as the two most typical MCS structures that lead to incredibly heavy rain during spring and summer in North America. G08, however, indicates that PS and TS are the dominant structures that can augment heavy rainfall and induce flooding. Despite the differences between PJ03 and G08 in the LS and TS as the substantial flood producers, both agree on the PS as a crucial structure in predicting flooding potential. In this case, the extreme precipitation associated with the PS structure was in accordance with the results of these previous studies. Thus, our study confirms that linear systems, especially those with line-parallel stratiform rain, play an important role in causing flooding threats.

This study aimed to improve the understanding of the organizational structure of mesoscale convective systems in regions with limited radar observation, like the northwest region of China. For this purpose, a method for making an objective MCS classification and segmentation was proposed based on geostationary satellite observations. In this study, the algorithm was developed using multiple infrared and water vapor channel observations from Himawari/AHI with a spatial and temporal resolution of 2 km and 10 minutes.

The products were qualitatively and quantitatively validated using radar composite reflectivity data from July 1 to 10, 2020. The MCS structures obtained using the algorithm proposed in this paper were more proximate to radar classification than those obtained using traditional methods based solely on satellite infrared brightness temperature, according to validation. Compared with the convective areas identified by radar, the POD of the satellite-identified convective was 65%, FAR was 39%, and ACC was 94%. A quantitative evaluation of the similarity between 50 given MCS structures derived from satellite and radar observations shows that the similarity between the two was as high as 60%.

Based on the above, this paper further applied the algorithm to recognize the organizational structure characteristics of an MCS that triggered the extremely heavy rainfall in Northwest China on August 21, 2016. The analysis revealed that IC, PS, and NL were the three types of organizational morphology that occurred in this process. Of the three types, IC corresponded to the formation of the MCS. In the development stage of the MCS, the system was characterized by PS and NL structures. Compared with precipitation intensity, the PS mode might played a key role in causing this flooding. These findings suggest that the PS-organized MCS may be an essential structural feature of storms that trigger extreme heavy rainfall in the semi-arid region of Northwest China.

The structure of MCSs is quite complex. Due to the limitations of payloads onboard current generation geostationary satellites in detecting the variations inside clouds, it is challenging to directly understand the structure of this complex weather system. In the algorithm of this study, the variations inside clouds were revealed by calculating the brightness temperature change at the top of clouds. Areas that have been continuously cooling for the past 30 minutes were classified as intense convective and convective zones. However, due to the Himwari-8/AHI observations used in this study having a time interval of 10 minutes, there may be a certain number of the convective cores with a lifespan of less than 10 minutes being missed. Therefore, in future research, the algorithm can be improved from two aspects: (1) adding satellite observations from payloads having vertical detection capability to obtain more information inside clouds; (2) supplying observations with higher spatial-temporal resolutions, such as the rapid scan onboard FengYun-4B, which can provide continuous observations of clouds at the spatial-temporal resolution of 250 m and 1 minute. With these data, satellites can hopefully recognize the structure of mesoscale convective systems as detailed as radar.