A simplified method for the detection of convection using high resolution imagery from GOES-16

. The ability to detect convective regions and assimilating the proper heating in these regions is the most important skill in forecasting severe weather systems. Since radars are most directly related to precipitation and are available in high temporal resolution, their data are often used for both detecting convection and estimating latent heating. However, radar data are limited to land areas, largely in developed nations, and early convection is not detectable from radars until drops become 10 large enough to produce significant echoes. Visible and Infrared sensors on a geostationary satellite can provide data that are less sensitive to drop size, but they also have shortcomings: their information is almost exclusively from the cloud top. Relatively new geostationary satellites, GOES-16 and GOES-17, along with Himawari-8, can make up for some of this lack of vertical information through the use of very high spatial and temporal resolutions. This study develops two algorithms to detect convection at different life stages using 1-minute GOES-16 ABI data. Two case studies are used to explain the two 15 methods, followed by results applied to one month of data over the contiguous United States. Vertically growing clouds in early stages were detected using decreases in brightness temperatures over ten minutes. Of the detected clouds, the method correctly identifies 71.0% to be convective. For mature convective clouds which no longer show decreases in brightness temperature, the lumpy texture, and rapid temporal evolution can be observed using 1-minute high spatial resolution reflectance data. The algorithm that uses texture and evolution for mature convection detects with an accuracy of 85.8%. 20 54.7% of clouds that are identified as convective by the ground-based radars are missed by the satellite. These convective clouds are largely under optically thick cloud shields. scheme using VIS channels uses visible texture caused by strong updrafts. The use of VIS and IR sensors in detecting convection can benefit significantly with the launch of National Oceanic and 85 Atmospheric Administration’s (NOAA’s) Geostationary Operational Environmental Satellites (GOES) – R Series which has high resolution, rapidly updating (i.e. 1 minute) imagery. This study makes use of this new data, namely the 1 minute data available from GOES-16 and GOES-17 in “mesoscale sectors” to update methods for detecting convection in different stages. One is developed for CI using T b from an IR channel in GOES-R. As in previous papers by Mecikalski and Bedka (2006) or Mecikalski et al. (2010), temporal trends of the data were used but, since GOES-R has high temporal resolution, ten 90 consecutive data with 1-minute interval were used. This procedure eliminates errors from cloud movements that needed to be dealt with in previous studies. Another one is developed for mature convection using both reflectances from a VIS channel and T b . For this algorithm, lumpy and rapidly changing surface and high cloud top height from mature convective clouds were used to detect clouds both with and without OTs. These two methods were then combined to provide detection of convection in all stages. The above methods are not intended to replace ground-based radars where these are available. 95 Instead, the focus here is complementing ground-based networks, either off-shore or other regions lacking coverage.

The datasets that were used to detect convection and validate the results are described in Sect. 2, while the methods used to identify initial and established convection are explained in Sect. 3. Sect. 4 highlights the results of each method. Two case studies were examined followed by a one-month statistical study to quantify the operational accuracy of the methods. 100 2 Data

The Geostationary Operational Environmental Satellite R series (GOES-R)
Earth-pointing instruments of GOES-R consist of the Advanced Baseline Imager (ABI) with 16 channels, and the Geostationary Lightning Mapper (GOES-R Series Data Book, 2019). GOES-16 is the first of the two GOES-R series satellites to provide data for severe weather forecast over the United States and surrounding oceans (Smith et al., 2017). Both 105 T b and reflectance data from the ABI were used to detect convective regions. Spatial and temporal resolutions of the GOES-16 data are summarized in table 1. Mesoscale data with one minute temporal resolution were used to fully exploit its high temporal resolution of the new instrument.
Reflectance at 0.64µm (Channel 2) and T b at 6.2µm (Channel 8), 7.3µm (Channel 10), and 11.2µm (Channel 14) were used 110 in the study. Channel 2 is a "red" band with the finest spatial resolution of 0.5km. This fine spatial resolution is useful to resolve lumpy, or bubbling surfaces of clouds in their mature stage. Channel 2 reflectance data were normalized by solar zenith angle so that a single threshold can be used throughout the method regardless of locations of the sun. Channel 14 is an IR longwave window band, which is a good indicator of the cloud top temperature for cumulonimbus clouds (Müller et al., 2018). High reflectance and texture of the cloud top seen in channel 2 and cloud top height inferred from channel 14 are 115 combined to determine locations of mature convective clouds.
Channel 8 and 10 are ABI water vapor channels with 2km spatial resolution. Because Channel 8 sees WV at somewhat higher altitudes than Channel 10, they can observe WV associated with updrafts as they progress upwards, and were therefore used to detect early convection. 120

NEXRAD and MRMS
Multi-Radar/Multi-Sensor (MRMS) data developed at NOAA's National Severe Storms Laboratory were used for validation purposes. MRMS integrates the radar mosaic from the Next Generation Weather Radar (NEXRAD) with atmospheric environmental data, satellite data, lightning, and rain gauge observations to produce three dimensional fields of precipitation (Zhang et al., 2016). These quantitative precipitation estimation (QPE) products have a spatial resolution of 1km and 125 temporal resolution of 2 minutes. https://doi.org/10.5194/amt-2020-38 Preprint. Discussion started: 25 June 2020 c Author(s) 2020. CC BY 4.0 License.
A "PrecipFlag" variable contained in the standard MRMS product classifies precipitating pixels into seven categories: 1) warm stratiform rain, 2) cool stratiform rain, 3) convective rain, 4) tropical-stratiform rain mix, 5) tropical-convective rain mix, 6) hail, and 7) snow. Details of the classification can be found in Zhang et al. (2016). A reduced set of these classes 130 were used to validate the convective classification from GOES ABI data. In this study, warm stratiform rain, cool stratiform rain, and tropical-stratiform rain mix are all assigned a stratiform rain type while grids with convective rain, tropicalconvective rain mix, and hail are assigned a convective rain type. Along with the classification product, MRMS provides a variable called "Radar QPE quality index (RQI)". This product is associated with quality of the radar data, which is a combination of errors coming from beam blockages and the beam spreading/ascending with range (Zhang et al., 2016). This 135 flag is used to mask out regions with low radar data quality. Only data with RQI greater than 0.5 are used in this study.

Methodology
This study examines methods to detect convective clouds at each life stage. Convective clouds can be divided into actively growing clouds and mature clouds. Actively growing clouds are usually clouds at the initial stage that grow nearly vertically while mature clouds are capped, but continue to bubble due to the release of latent heat. They often move horizontally after 140 they reach the tropopause. The proposed method to detect actively growing cloud is similar to the GOES-R CI algorithm in the sense that the method uses temporal trends of T b . The availability of higher temporal resolution data, however, allows us to simplify the method somewhat to use only two channels in the water vapor absorption band. The high temporal resolution data further simplifies the method because the use of derived wind motion in tracking clouds is no longer necessary. One minute is short enough that cloud motion, at most, is to the adjacent grids, and clouds can be easily tracked by focusing on 145 overlapped scenes.
The method to detect mature convective clouds is similar to previous studies by Bedka et al. 2016 andBedka et al. 2019 in terms of using the texture of the cloud top surfaces to infer strong updrafts. Cloud top surfaces of mature convective clouds are much bumpier than any other clouds, and their bumpiness is most evident in VIS images with the finest resolution. The 150 following method uses horizontal gradients of reflectance to represent the bumpiness of cloud tops, and the magnitude of the gradients are used to distinguish convective cores from their anvil clouds. Cloud top temperatures from channel 14 are used to eliminate low cumulus clouds that might appear bubbling.

Detection of actively growing clouds with brightness temperature data
In the early stage of convection, updrafts of water vapor eventually lead to condensation, the release of latent heat, and 155 convective processes. Operational weather radars cannot observe this, but GOES-ABI, with water vapor absorption bands, can. During the early convective stages, T b s that are sensitive to water vapor will decrease due to the increasing height (and cooler temperature) of water vapor and cloud droplets. Two ABI channels around the water vapor absorption bands, channel were used to find small regions consistent with developing clouds. If a cloud develops continuously for ten minutes and 160 shows a large decrease in T b over ten minutes in either channel, the cloud is determined to be convective.
To compute the T b decrease in clouds, a window has to be defined as it is usually difficult to precisely define the boundary of clouds, especially at the early stages of convection. Since most of the early convective clouds are smaller than 10km in diameter, the window was defined as a 10km´10km box which is essentially a 5´5 matrix of satellite pixels consisting of 25 165 T b s with 2km resolution. Considering the fact that a convective core usually has the lowest T b within its neighborhood, the T b matrix was formed around a pixel only if that pixel had the lowest T b in the 5´5 matrix. However, this criterion alone could not distinguish convective cores from stratiform clouds and cloud edges which can also exhibit a local minimum. In addition to the lowest T b , the shape of convective clouds is therefore also considered. As shown in the Fig. 1a, convective clouds not only have the lowest T b in their cores in all directions, but also have increasing T b s away from the core, making their T b 170 distributions look like an inverted two-dimensional (2D) Gaussian distribution. To select T b matrices that have this inverted Gaussian shape, an inverted 5´5 Gaussian matrix that has mean and standard deviation of the T b matrix was created and compared with the T b matrices. To focus the comparisons on the shape of the T b distribution (Fig. 1b), the maximum T b found in the 5´5 matrix was subtracted from all values, and T b values were divided by the difference between maximum and minimum T b to normalize the T b matrix itself. If the T b matrix has a shape of a developing cloud (i.e. 2D inverse Gaussian), 175 the absolute value of the difference between two matrices will be small. A threshold of 10 for this absolute value of the difference between T b shape and inverse Gaussian shape (sum of residuals between normalized T b and inverse Gaussian) was empirically determined to exclude non-convective scenes. T b matrices with values greater than 10 are removed from the scene. This is done for all ten consecutive T b images that are one minute apart. Continuous overlaps of T b matrices for ten minutes imply that the cloud maintained a convective shape for ten minutes, and therefore, changes in T b are calculated to 180 assess if the cloud in the T b matrices was growing.
The minimum T b s of the T b matrices at each time step were linearly regressed against time to measure a decreasing trend.
Since one-minute data can be noisy, the decreasing trend was considered instead of an actual difference in T b during 10minutes. If the fitted line had a slope smaller than -1K/min for channel 10 or -0.5K/min for channel 8, the grid with the 185 lowest T b at each time step for ten minutes as well as the neighboring 8 grids in the window were classified as convective.
The threshold of -1K/min and -0.5K/min were determined empirically, although it might not be the perfect threshold. Some convective clouds in the early stage show smaller decreasing trend, but using a smaller value for the threshold can introduce clouds that do not grow into deep convective clouds in the end. Clouds that develop into deep convective clouds are eventually captured by these thresholds in later times even if they had small decrease in the beginning. Actively growing 190 clouds are usually detected by channel 10 first and then by channel 8. This makes sense because channel 10 sees water vapor https://doi.org/10.5194/amt-2020-38 Preprint. Discussion started: 25 June 2020 c Author(s) 2020. CC BY 4.0 License.
in lower parts of the atmosphere while channel 8 sees upper level water vapor. Using two channels help find the same clouds in different levels.

Detection of mature convective clouds with reflectance data
Mature convective clouds consist of convective cores and stratiform or cirrus regions where clouds have detrained from the 195 core. The lack of discrete boundaries between different types of clouds make it difficult to separate convective grids from surrounding stratiform regions. Overshooting tops and enhanced-V pattern are well-known features in mature convective clouds, but these do not appear until their strongest stage and not in all convective clouds. Using such features associated with the deepest convective cores will create a detection gap between early and mature stages of convection. The method described here tries to minimize the gap, while still accurately detecting convective clouds. 200 A distinct feature that appears in convective clouds, even in their early stages, is a bubbling cloud top. The lumpiness of cloud tops can be numerically represented by calculating horizontal gradients in the reflectance field with the Sobel-Feldman (Sobel) operator which is commonly used in edge detection. The horizontal gradient is calculated at each pixel. The Sobel operator convolves the target pixel and its surrounding eight grids with two kernels given in Eq. (1) to produce gradients in 205 the horizontal and vertical direction.
By using Eq. (2), gradients in each direction are combined to provide the absolute magnitude of the gradient at each point. 210 Magnitude of gradient = G " 7 + G * 7 (2) Flat surfaces will have low gradients while cloud edges or lumpy surfaces will have high gradients. This lumpy feature is most evident in a VIS channel with the finest spatial resolution of 0.5km. IR fields are not very useful as the brightness temperature variations in these lumpy surfaces tend to be quite small relative to the IR's 2km resolution, and only cloud 215 edges stand out.
Before evaluating the texturs, only the grids that are potentially parts of deep convection are selected using simple threshold values of VIS (ABI channel 2; 0.65µm) and IR (ABI channel 14; 11.2µm) channels. Channel 2 reflectance is highly correlated with the cloud optical depth (Minnis and Heck, 2012) while Channel 14 brightness temperature is related to cloud 220 top temperature (Müller et al., 2018). These channels are used in GOES-R baseline product retrieval of cloud optical depth https://doi.org/10.5194/amt-2020-38 Preprint. Discussion started: 25 June 2020 c Author(s) 2020. CC BY 4.0 License. and cloud top properties, respectively. Any grids with reflectance less than 0.8 or T b greater than 250K during ten time steps (10 minutes) are removed since they generally represent thin or low clouds such as cirrus or growing clouds that can be identified by the CI method described earlier. These thresholds are chosen rather generously to include some convective clouds that have not grown into deep convection yet, while still avoiding the misclassification of low cumulus clouds and thin 225 anvil clouds as convective. The threshold of 250K is much warmer than typical values used in detecting deep convective features such as overshooting tops (Bedka et al., 2010) or enhanced-V (Brunner et al., 2007. It is intentionally chosen so that the method considers warmer convective clouds without those features in the next step when evaluating lumpiness of cloud top. The choice of these thresholds is discussed in more detail in section 4.3.

230
Once cold, highly reflective scenes are identified, the horizontal gradients of reflectance are calculated using the Sobel operator. The average of the horizontal gradients over the ten 1-minute time steps is calculated for each grid, and grids are removed if the average was less than 0.4 or greater than 0.9. Values below 0.4 or above 0.9 generally implies either stratiform region with a flat surface or cloud edges with very high gradients. The remaining grids were then interpolated into 1km maps to be consistent with the spatial resolution of MRMS dataset. Neighboring grids were grouped to form clusters, and only the 235 clusters with more than 5 grid points were assigned as a mature convective cloud to remove noise.

Results
We begin the result section with two case studies that illustrate the technique as well as some of its limitations.

June 28 th , 2017
Supercell thunderstorms developed in Iowa and produced several tornado touchdowns. In Fig. 2a, deep convection had 240 already developed over central Iowa at 19:30UTC, and two convective cells in the red box started to develop in southeast Iowa, although they do not stand out from surrounding low clouds in the VIS image. These two convective clouds became parts of major storm system that formed around 21:30UTC, producing the tornadoes (Fig. 2b) in the area. The two cells appeared in the Omaha (KOAX), Des Moines (KDMX), and La Crosse (KARX) NEXRAD radars at 19:30UTC (Fig. 2c), but reflectivity was very weak (≤ 30dBZ). In addition, the MRMS PrecipFlag product is shown in Fig. 2d. Convection is colored 245 in pink and stratiform in green. Although deep convections over central and northeast part of Iowa were assigned as convective in MRMS at 19:30UTC, the two growing clouds in the red box in Fig. 2a were not assigned convective flag until 19:48UTC. 19:27UTC and 19:36UTC. However, not all of the matrices in these figures showed the evolution of the developing cells (decreasing minimum T b over 10K) between the two time steps. The only two matrices in this scene that satisfied both 255 criteria of maintaining the shape of developing cells and growing vertically over ten time steps were the two in blue circles.
These two matrices contain early convective clouds that grow into deep convection shown in Fig. 2b, and they are correctly captured by this method. Since the same method is used in each time step, the same window can be captured throughout an overlapping time period despite the starting time. Therefore, this method can be used continuously in time.

260
Results for the detection of mature convective clouds are shown in a step by step fashion in Fig. 4. Figure 4a is the same as in Fig. 2a, but is mapped using a different color table for better comparisons between steps. Figure 4b shows the pixels retained after eliminating all the grids that did not meet the reflectance and T b thresholds (minimum reflectance over ten time steps greater than 0.8 and maximum T b over ten time steps less than 250K). Figure 4c shows the horizontal gradient values after applying the Sobel operator. The colorbar is set to be within the range of 0.4 and 0.9 to display potential convective regions 265 that passed these thresholds in colors. White regions are either regions that have average gradients greater than 0.9 such as cloud edges or thin cirrus clouds, or regions that have average gradients less than 0.4 such as clear sky or stratiform regions.
Eventually, only the regions that meet both the criteria in Fig. 4b and 4c are assigned to convection, and shown as white shade in Fig. 4d. When the locations of convective regions in Fig. 4c and 4d were parallax corrected with a constant cloud top height of 10km and compared to retrieved products from radars in Fig. 2c and 2d, most of convective regions align well 270 with high reflectivity regions in Fig. 2c and convective regions in Fig. 2d. However, a straight line around 44N at the right edge of Fig. 4d is definitely not a convective region, and it is due to unrealistically high reflectance in the raw satellite dataset. These kinds of artifacts were removed later in section 4.3 when the method was applied to a full month of data.
However, multiple lines are difficult to remove at this stage in the processing and will result in false alarm. As quality control procedures on ABI are improved, this may no longer be a source of significant errors. 275

June 18 th , 2018
Another case was examined to evaluate the methods under different conditions. Severe storms developed over the Great Plains in June 18 th , 2018, producing hail on the ground. At 22:30UTC, sporadic storms across Kansas and Oklahoma were observed by GOES-16. This scene contains both growing and mature convective clouds, and MRMS PrecipFlag for the scene is shown in Fig. 5a and 5b. Green color represents stratiform and pink color represents convective clouds. Figure 5c and 5d 280 are brightness temperature maps of the same scene at 22:30UTC and 22:40UTC, respectively. Growing clouds are shown in purple, blue, yellow, and green boxes. Convection in the purple box is detected using ten consecutive T b data starting at 22:19UTC. Considering the last data used at 22:28UTC, it was detected six minute earlier than MRMS detection which was at 22:34UTC. The growing cloud in the blue box was detected 7 minute after MRMS detection at 22:38UTC. This cloud did not grow rapidly enough and did not meet the T b threshold for channel 10 at the onset of convection. However, it was 285 detected by channel 8 as it grew higher altitudes. Similarly, a cloud in the green box was detected by channel 8 starting at https://doi.org/10.5194/amt-2020-38 Preprint. Discussion started: 25 June 2020 c Author(s) 2020. CC BY 4.0 License. 22:27UTC. Although it precipitated, this cloud did not grow into a severe storm. The growing cloud in the yellow box was detected by GOES using data from 22:29UTC and detected by MRMS at 22:38UTC. These results show that even though the thresholds for the T b method can be strict for some growing clouds, the thresholds were adequate for detecting convective storms in their earliest stages. 290 Black regions superimposed on the brightness temperature map in Fig. 5c represent convective regions identified by the mature convection method. There are slight misalignments of detected convective clouds between MRMS PrecipFlag products and GOES results possibly due to sheared vertical structures of the storms, but overall, they seem to match well.

Statistical results with one-month data 295
Thresholds used in the two case studies were chosen based on experiments applied to one month of data during June of 2017.
The T b method for detecting early convection only requires one threshold of T b decrease. However, since water vapor channels have different sensitivity to water vapor, different values for the threshold are required for each channel (channel 8 and 10). Since clouds do not grow at the same speed, and growth rate can vary at different evolution stages, it is important to find an appropriate threshold that best represents growth rate for clouds in their early stages. In order to investigate proper 300 values, the 5´5 T b windows that maintained the developing shape and had a decreasing trend of T b during ten minutes are collected over the one month period. A total of 27971 and 73204 (for channel 8 and 10, respectively) 5´5 windows were collected, and precipitation types from MRMS were assigned for each window. Future MRMS convective flags up to 30 minutes were included in the analysis because some time delays were observed in MRMS product when assigning convective flags, especially for early convection. When comparing GOES products to future MRMS products, future locations of GOES 305 products were calculated assuming convection moves at a constant speed. Tables 2 and 3 show results applying different thresholds ranging from -0.1K/min to -2.0K/min. Numbers in the table represent the number of 5´5 windows that MRMS precipitation flags were assigned to either non-convective or convective at the corresponding 10-minute time window, as well as pixels that were flagged as convective by MRMS in the next 20 minutes to account for the fact that GOES can detect convection before the radar sees precipitation. For channel 8, using a threshold less than -1.4K/min has 100% accuracy, but 310 in return, it misses much of the convection and loses an ability to detect convection earlier than radar because most of early convection does not have such a strong updraft. Thresholds are chosen so that it achieves at least 85% accuracy and detects reasonable amounts of convections. Therefore, -0.5K/min and -1.0K/min are chosen for channel 8 and 10, respectively. It is interesting to note that some clouds did not produce precipitation even with rapid growth over -2.0K/min (for channel 10).
This could be because evaporation occurred while it fell, leading to virga, or it did not have enough time to grow bigger due 315 to such a strong updraft.
The reflectance method used to identify mature convection uses three thresholds; reflectance at channel 2 and T b at channel clouds with flat cloud top surfaces. Cloud top texture (horizontal gradients of reflectance) is the key feature to detect 320 convection in this method, but since low cumulus clouds can also have lumpy cloud top surfaces, reflectance and T b thresholds are used additionally to distinguish these low or shallow cumulus clouds from convective clouds.
Results using different combinations of the three thresholds are compared. First, the lower bound of the texture thresholds are varied, keeping the upper threshold and the T b and reflectance thresholds constant. Resulting FAR and POD are shown in Fig.  325 6. The upper threshold does not change results much, and cloud edges are effectively removed by the threshold of 0.9. Using 0.5 (yellow) misses significant amounts of convective regions while using lower values (blue and red) substantially misclassifies stratiform regions with flat cloud tops as convective, although their PODs are much higher. A value of 0.4 (green diamond in Fig. 6) was chosen as a reasonable compromise between POD and FAR.

330
POD and FAR using different combinations of T b and reflectance thresholds are plotted in Fig. 7, and this time texture thresholds are kept constant with 0.4 and 0.9. The T b threshold is varied from 230K to 250K, and the reflectance threshold is varied from 0.7 to 0.9. There is a trade-off between detecting more mature convective clouds in the earlier stage and incorrectly assigning cumulus clouds as convective clouds. Having lower value for the T b threshold or higher value for the reflectance threshold leads to small FAR, but also leads to small POD. To make this method effective and not too harmful for 335 the data assimilation, 250K for the T b and 0.8 for the reflectance threshold (black diamond in Fig. 7a) are chosen. 240K and 0.7 (orange) also showed similar results, but 250K and 0.8 were chosen due to lower FAR. Figure 7b shows results including MRMS data 10 minutes after the detection period to show its ability to detect convection earlier than MRMS. Most of the results showed improvements in both FAR and POD (lower FAR and higher POD) when later data are included.

340
Despite FAR being relatively small, the method misses significant amounts of convective areas observed by MRMS.
Therefore, regions that were missed are evaluated further to investigate which threshold was most effective in preventing the method from detecting convective regions. Figure 8 shows histograms of T b , reflectance, and texture in the convective regions that were missed by the above method. It is clear from the figure that the largest number of misses were due to low texture values. There are many reasons why convective regions appear to have flat cloud top surfaces. Anvil or thick cirrus 345 clouds above convective regions can smooth out or cover bubbling cloud tops, and there is simply no way to avoid this problem. Another reason may be the nature of the classification method. Since classification by MRMS is determined by rain rate, even if convective clouds are in a decaying mode and do not bubble anymore, clouds can still continue to precipitate considerable amounts, which would lead to convective category in the MRMS product. It is also possible that it is due to a misclassification of trailing stratiform regions using radars. Previous studies (Qi et al. 2013;Shusse et al. 2011) have indeed 350 tried to improve the radar classification schemes.
As shown from these results, there are no perfect thresholds that can separate convective and stratiform clouds. Nevertheless, threshold values were chosen based on our purpose, which are to avoid FAR as much as possible and have decent POD comparable to radar products. Results applying both methods to one month data with the chosen thresholds are summarized 355 in table 4, and based on the table, FAR is 14.4% and POD is 45.3%. Most of the detection was from mature convective clouds, and its accuracy of 85.8% was much higher than that of the detection of growing clouds, which was 71.0% at a pixel level. When MRMS data up to 30 minutes were included in the analysis, considering methods' ability to detect convection earlier than radar, the accuracy of both methods increased, and overall FAR was reduced to 11.2%. Although FAR is slightly over 10%, 96.4% of these pixels were at least raining. Since the main objective of data assimilation is to have good prediction 360 of precipitation, applying these methods during data assimilation can still be beneficial in case the forecast model did not produce precipitation.

Conclusion and summary
This study explores two methods to detect convective clouds using GOES-R ABI data with one minute interval. Using such high temporal resolution data facilitates cloud tracking and helps the accuracy of the detection method when calculating 365 decreases in T b of the same cloud. Convective clouds in the early stage were detected using T b s of ABI channels 8 and 10.
These channels were used to find cloud scenes with the developing shape of convective clouds. They were then used again to calculate the T b decrease for those which maintained the developing shape for ten minutes. A cloud scene that had a consistent developing shape and a large decrease in T b over ten minutes was classified as convective by this method. Mature convective clouds were detected by masking out regions with high T b in ABI channel 14 and low reflectance in ABI channel 370 2 and finding regions with high horizontal gradients of reflectance over the course of ten minutes. Results from this reflectance method were mostly consistent with the radar-derived products, although this method is limited to daytime use only. Nevertheless, it detects a wide range of convective area, not just regions with overshooting tops.
These methods work well for well-structured convective clouds, but there are limitations to this method as with most 375 algorithms using IR and VIS sensors have. Cirrus cloud shields are the biggest problem as they block T b decreases underneath and smooth out lumpy reflectance surfaces. However, these methods can still be extremely useful for defining convection for assimilation into models where radar data is not available. Because regions identified as convective are most likely convective (~90% accuracy), this can easily be assimilated while setting cloudy regions to "missing" since the accuracy of detecting convection under large cirrus shields is poor. 380

Author contributions
All three authors designed the experiments. YL processed and analysed the data. CK and MZ gave feedbacks with their insights at every step of the data analysis. The manuscript was written jointly by YL, CK, and MZ. https://doi.org/10.5194/amt-2020-38 Preprint. Discussion started: 25 June 2020 c Author(s) 2020. CC BY 4.0 License.