the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 05 Feb 2025
A machine learning-based perspective on deep convective clouds and their organisation in 3D. Part II: Spatial-temporal patterns of convective organisation
Abstract. This sequence of papers examines spatio-temporal patterns of convective cloud activity and organisation. In response to the limitations of current remote sensing sensors, our analysis employs a machine learning (ML)-based contiguous 3D extrapolation of 2D satellite data. In Part 2, we investigate spatio-temporal patterns of convective organisation over West Africa and assess their connection to 3D convective cloud and core properties. We employ three organisation indices (COP, SCAI, ROME) to statistically quantify convective organisation. Our results show that convective organisation increases for long-lasting mesoscale cloud systems with numerous deep convective cores. It is likely connected to the Inter-Tropical Convergence Zone (ITCZ) and its northward shift in summer. In spring (March–May), strong convective organisation appears around the Gulf of Guinea and the remote Atlantic Ocean between 15–30° S. The seasonality of convective cloud development induces an increase of the indices between 5–20 % in summer. For instance, the landmass distribution and the influence of extra-tropical dynamics may cause a considerably higher variability in the southern hemisphere. Over the ocean, the organisation indices (COP, SCAI) are about 5–10 % higher than over land. However, derived statistics may be affected by overlapping effects of isolated and clustered convection occurring in the same region. To disentangle their impact calls for an adaptive index. To summarise, combining the information of multiple remote sensing instruments may deliver more profound insights into convective organisation. For future research, we emphasise a further need for a robust quantification of convective organisation.
- Preprint
(4383 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 24 Mar 2025)
-
RC1: 'Comment on egusphere-2025-376', Anonymous Referee #1, 25 Feb 2025
reply
General comments
This study presents the spatial and temporal patterns of convective organization based on a three-dimensional cloud field constructed using a machine learning technique that combines data from horizontal distribution from geostationary satellite MeteoSat-11 and vertical profiles from Cloud-Sat Cloud Radar. The area of interest (AOI) spans from 30°S to 30°N and 30°W to 30°E, where parameters such as cloud area, cloud top height, cloud lifetime, the number of deep convective clouds, core size, and the area ratio between the core and the anvil are analyzed.
Detailed results illustrate seasonal and geographic variations in organization indices and related quantities. The newly constructed dataset provides unique insights into convective organization within the AOI, making it a valuable resource for understanding these characteristics. Therefore, the paper could be published only after clarifying the following results.
The reviewer’s main request is to clarify the description of the results. For example, the term “convective activity” is used without clearly specifying its characteristics, leading to confusion rather than clarification. Although the manuscript presents various findings on geographic distributions and seasonal evolution, it would benefit to present a summarizing analysis that highlights the key conclusions. In particular, the differences in convective organization between sea and land, as well as between the Northern and Southern Hemispheres (Figs. 14 and A1), remain unclear and should be clarified.
Moreover, many of the key results are not exclusively derived from the unique three-dimensional dataset with continuous object tracking. Ideally, the analysis should emphasize aspects uniquely obtainable from this dataset. It was disappointing that the authors chose indices of convective organization that could be derived from 2D imagery alone. The authors should clearly summarize the advantages of their unique dataset and highlight how it advances our understanding of convective organization.
The characteristics of the dataset, especially the 3D objects obtained through the machine learning method, should be described more clearly in this paper, even if detailed explanations are provided in Part I. In particular, it remains unclear whether the identified 3D objects are smoothly connected over time.
Specific comments
L26, “convective organisation (or aggregation)”: The authors should distinguish “convective organization” and “convective aggregation” by giving their definitions.
L30, “The spatial distribution of convective clouds is not arbitrary.”: This sentence is unclear.
L37, “several regions”: It is unclear. What types of “regions” are meant in this context?
L44, “So far, the models show convective organisation increases with a warming climate”: Wing et al. (2020) also showed a change in convective organization with warming in RCEMIP.
L58-60, “At the same time, food security and a high climate risk exposeWest Africa to multiple threats (Berthou et al., 2019). Changing atmospheric conditions
could intensify those hazards.”: We know that these points are important but not specifically related to the current research. These sentences should be moved to the final section, instead of the introduction, or removed.
L104-105, “It is characterised by lower temperatures and a strong vertical ascent, which we identify by an extensive vertically contiguous 105 layer and a high radar reflectivity (e.g., Igel et al. (2014); Takahashi et al. (2017)).”: Vertical ascent is not directly analyzed by the proposed method, neither in Igel et al. (2014) nor in Takahashi et al. (2017). This sentence should be modified.
L173-174: Please describe the methodology of the moving windows in more detail. What types of iterations are applied to calculate the indices?
L179, “the frequency distribution shows an overlap of lower index values for all indices”: This sentence is unclear. What does this mean by an overlap?
L230-232: What is the meaning of “a diverging convective activity”? In Fig. 7a, we cannot see where the maximum cloud area is. Where is “a lower cloud lifetime” in Fig. 7c?
L244, 245: What does “convective activity” mean in the sentences, and which figure shows this? Which figure and location show “a higher convective activity comes with a lower area ratio, a higher number of DCCs, a larger cloud and core area”?
L254, “the convective organisation is overall weaker around the equator”: In 4.1, “convective organization” is not clearly defined and is not specifically described. The definition of organization must be clarified.
L255, “Their impact on large-scale patterns of organisation is limited
compared to MCSs”: The meaning of this sentence is unclear.
L263, “(Figure 9, a-b,g-h,i-j)”: Fig.9-j does not exist.
L264: “Figure 7” should be “Figure 6”. Figure 7 does not show seasonal distributions.
L274: What does “convective activity” indicate here?
L274-279: In Figure 7, which index is used to define a most or least organized group?
L303-304, “Overall, organised clouds (P90) come along a larger cloud anvil area, a
longer cloud lifetime, a lower CTH, a lower area ratio, and more and larger DCCs”: It is unclear from Figure 13 to see these relationships. The relationships between the indices and these properties should be directly compared. A part of the comparison between the indices and the number of DCC is shown in Fig. 4g-i.
L320-336: The authors describe notable characteristics which can be seen from Fig. 14. However, some points are not convincing from the figure. Please check whether the description is consistent with the figures. For example, in L328-330, I cannot see a noticeable narrowing in summer (JJA) for ROME. The differences in the effect size are significant according to the numbers in the Figure, but not clearly visible.
L350, L359-360, “the microphysical cloud properties”: Cloud anvil area is a cloud macrophysical characteristic rather than a microphysic property. It is true that cloud microphysics affect cloud anvil area through the balance between sedimentation and the outflow, the present study does not examine microphysical cloud properties specifically.
L361-362, “For continental cloud clusters in the northern hemisphere, we find more distinct results regarding the relationship between DCCs and the degree of organization”: This is one of the noticeable results discovered in this paper. However, this conclusion is indirectly shown by the figures in 4.1.2. The authors are suggested to show a more direct analysis showing the conclusion.
L468: Spell out “JAS”.
-
RC2: 'Comment on egusphere-2025-376', Anonymous Referee #2, 06 Mar 2025
reply
General comments
This paper assesses the distribution of convective organisation and deep convective cloud properties over tropical Africa, using machine-learning derived satellite radiance fields. This work extends existing literature by tracking convective clouds and deriving cloud properties over the lifecycle of clouds rather than snapshots alone. They investigate spatial, diurnal and seasonal variations and show that, in their dataset, clouds in organised regimes tend to live longer, have large maximum cloud areas and have more convective cores.
This work offers a few key advances to existing literature, and, with a number of major corrections, offers a useful addition to the scientific body and should be accepted for publishing.
The main recommendation is to add new analyses to address the scientific objective more directly. The aim of the study is to get at the relationship between organisation and cloud properties, but very little of the analysis progresses this aim despite the opportunity to do so with the dataset developed. The focus on seasonal, land/ocean and hemispheric differences are relatively arbitrary, and not well developed or discussed. Further, with only 6 months of data, from one year, the seasonal results in the paper are of limited value as they do not capture the entire seasonal cycle or account for inter-annual variability. Instead, we suggest more focus on direct comparisons between key cloud properties and the level of organisation measured using convective indices and the tracking of convective cores.
The hemispheric differences emphasised in the text may mainly result from the large oceanic anomaly (-15 S, 0 E; Figures 6f, 7acdef, 8abcd). But this tends to be a region of large-scale descent covered mainly by low level stratocumulus, rather than deep convection. Is it possible that systems here are being misidentified as DCCs (eg., mid-latitude cyclones or atmospheric rivers)? Given the small sample size (Figures 5,6), average properties here may be unreliable. The authors should address this, and focus on distinct convective regimes rather than hemispheric land/ocean divisions.
The quality of written expression needs improvement. In particular, the authors frequently make claims that are not evidenced by the results. Comparable spatial patterns alone do not justify inferred relationships. In addition, there are issues with referencing throughout. Citations are frequently used as evidence for a methodological choice or finding that are merely examples of comparable works. Please clarify why a statement needs referencing, or merely state the result.
The dataset was not adequately detailed, and relevant limitations and validation were not addressed. Please justify whether the derived-radiances and cloud-tracking methods are suitable, and clarify uncertainties across the AOI, seasons, diurnal cycle, cloud regimes or lifecycle. Can mid-level cloud and cumuli arise that may instead determine the cloud area extent? How were the machine learning derived reflectivities evaluated? A summary figure of the tracked clouds would help introduce the data, and help to interpret the robustness of the organisation indices in different areas/seasons/time of day. Additionally, normalised density maps alone do not convey absolute occurrences of the data shown (Figures 5,6,9,10 and 11); this information is important for interpreting the results and should be provided, perhaps present frequency maps instead of densities.
It should be clarified what advances are made from using the 3D fields and whether these justify uncertainties associated with additional data processing, as comparable cloud tracking and properties can be derived from 2D data.
“Anvil cloud” and “anvil area” are discussed, but at no point is a cloud anvil defined. If no work was done to segment the anvil from the cloud objects tracked or ensure the cloud area extent always results from an anvil feature, all references to cloud anvils should be specified as “cloud” or “cloud area”.
The calculation of cloud properties and spatial density distribution was not described.
Assessing the area / core ratio does not seem to add much information to this work, additional justification of why this was included and what it shows would be welcome.
Specific comments
L36: “contiguous convective regions” unclear whether this refers to a convective cloud, multiple connected convective clouds, or an MCS. There are issues throughout on the use of “MCS” and “organised states” (also L260, L276). In addition, the following line regarding MCS organisation is a little unclear.
L45-46 [discussion on convective aggregation]: Contradicts statement at L53-54. Aggregation is different to organisation, and is generally a phenomena seen in RCE models, please clarify.
L58: MCSs contribute not only the majority of extreme rainfall, but all rainfall, and the stratiform component of their precipitation is important for this. Highlighting could provide more justification for the tracking of the entire anvil, not just the convective cores
L71: “DCC” used as an acronym for deep convective cores, while it is more commonly used as an acronym for deep convective clouds. Would recommend referring to the convective cores simply as “cores” as used in some figure headings
L83: This talks about “cloud development”, but the lifecycle of tracked convective clouds is not assessed in this manuscript. Is this meant to refer to the distribution of observed cloud properties?
L86, L350, L359, L406: No microphysical properties of clouds (i.e. Effective radius, droplet number density etc.) are discussed in this paper. Instead, the authors may be referring to the bulk or macrophysical properties of clouds
L93: Do you include the 3.9 micron channel (channel 4) from SEVIRI? This channel includes a contribution from reflected solar IR, so has a different response between day time and nighttime. Has the consistency of the input data been verified across the diurnal cycle?
L101: The resolution is 3km at nadir, however this increases further away from the sub-satellite point. Should the comment on the vertical resolution say that it is that of Cloudsat CPR, not SEVIRI?
L102 onwards (discussion of cloud tracking): The detection starts with a very low radar reflectivity threshold of -15dBz, which will detect all clouds, not just convective cloud features. Is any restriction on cloud height applied, or will this also detect low level liquid clouds? How are scenes with multiple layers clouds handled?
L108: only one threshold was used to define the cloud boundary, you apply some filtering to elongated clouds, but does this mean that in other cases some organised systems are counted as one cloud object?
L110: “convective updraft”, data for an “updraft” (i.e. vertical velocity) is not used. Is this instead based on a radar reflectivity threshold?
L111: please clarify how the local extrema are used to detect core objects. Is there any contiguity of these required in time, and does using the local extrema to define cores result in there always being >= 1 core at each time? If a cloud has many cores over its lifetime but they don’t co-occur in time, is it still reported as having multiple cores?
L118: In the introduction (L48) you state that observational studies have been limited as there is a “low frequency of events most relevant for aggregation”, is your 6 months of data sufficient?
Section 3.1: It would be useful to have a final paragraph comparing the pros and cons of each organisation metrics and in which situations they are more or less reliable, rather than isolated comments or their capabilities which are difficult to compare
L131: Please state why these three indices were chosen over others
L139-140: Please clarify how shifts in time and space occur and why they matter. Does the calculation of convective organisation indices take into account multiple time steps, or is the calculation independent per time step?
L152: “continuous convective regions”, a “convective region” has not been defined and this is unclear
L177: I suggest you keep consistent the ordering of the indices, which you introduced as (SCAI, COP, ROME) but here and in the figures list as (COP, SCAI, ROME). I would also suggest that, for readability, SCAI is not placed in the middle of the two that have the same direction for increasing organisation.
L178: The phrasing of the statement about values of COP and ROME vs SCAI is confusing. I would recommend rephrasing along the lines of “By design, COP and ROME produce larger values for more organised domains, while SCAI instead results in smaller values”
L184: Can the metrics be compared to each directly in this manner?
L187-188: By “the diurnal cycle is opposed” are you referring to lower SCAI values meaning more organised as opposed to higher values for COP/ROME?
L189 (on diurnal cycle of organisation metrics): Yes, but isolated convection tends to only exist for a short period around the diurnal maximum, whereas more organised convection lasts for longer throughout the diurnal cycle, so it is not contradictory that observed convection is less organised during the diurnal maximum.
L190: Most studies have found convective maximum over the ocean to occur during early morning hours, but the diurnal differences are much smaller than over land
L193-195 and Figure 4 d,e,f: I interpret this as only small latitudinal differences in COP and ROME, excluding a peak in the south around -20 degrees, yet a large weakening in organisation reported by SCAI in the equatorial region. I don’t see that the variability differences between land and ocean are significant. I would like the commentary to instead explain the strong weakening reported by SCAI, and whether this relates to data sensitivity due to the high number of clouds found in this region.
Fig 3: It would be clearer to plot these distributions as proportions, rather than frequencies, to allow easier comparison between land and sea. Also, it would be good to increase the number of bins to show more detail in the distributions
L198: Can the convective core detection be used here to resolve this issue rather than calculating organisation purely from the MCS cloud shields?
L203: How are the spatial densities calculated? Is it using a gaussian kernel estimation?
L204-205 and 210: “isolated convective cells”, “highly clustered systems” and “clustered systems”, the link between the number of cores and the “clustering” is not defined or justified. The assumption is that one detected core means that the cloud is isolated, but this may not be the case. Further a system with multiple cores may be isolated. Could the authors provide additional detail in the text justifying the categories used here?
Fig 4: The layout of the figures could be rotated to make comparisons clearer, e.g. change COP, SCAI and ROME values to be along one row each, and make all diurnal cycle plots one column etc. (i.e. current positions a,b,c would move to a,d,g and so forth)
Fig 4e: Why does SCAI reduce so much towards the northern and southern edges of the domain?
L214: Should this read less than 1000km2? The description of how systems are binned by area is clearer in the caption of figure 5 than in the text.
L214-215: why define moderate cloud area as (mean, 10 x mean)? The skewed distribution of anvil cloud area results in a mean that is greater than the median, so the majority of clouds will end up in the smallest category.
L222-223: “we generally find highly clustered systems to be accompanied by a larger cloud anvil size and vice versa”, this claim not supported. These plots mainly show that most of the tracked clouds occur in the equatorial region. Also the number of cores a cloud has has not been directly compared with the cloud area, which would be interesting to include in the discussion.
Figure 6. I interpret this figure differently. I find the differences between the distribution of small and medium cloud areas to be small, and not very meaningful. I think the main result here is that the largest clouds predominantly occurred over land, and not in the equatorial belt. This may also relate to the number of samples in this partition, which should be stated.
L224: “convective activity”, not defined or assessed. Could the estimated radar reflectivity be used to provide a measure of convective intensity?
L225-226: I disagree partially with this description. A “reduced cloud lifetime, “enhanced area ratio” and “fewer DCCs” is not characteristic of the equator region, but rather characteristic of most of the domain when compared against the large anomalous region over the Atlantic (15S 0E)
L227: I don’t think this reference is really appropriate here, as only SCAI reported less organisation over the equator in your dataset?
L229-230: The reference here is unnecessary without a link between your results and the previous study.
L250: Why is the number of systems included in the 10th and 90th percentile bins unequal?
L256: The relationship between number of convective cores and convective organisation could be shown much more clearly by showing a scatter plot of the two to show how correlated they are, rather than comparing spatial distributions
L258: Figure 8e shows core size to decrease near the equator in a large region
L264: The patterns in figures 9 and 10 appear much more complex than discussed here. In particular, ROME appears very different during JJA, and appears to show the opposite locations for more organised convection than COP. Why is this?
L269: The description of how the convective indices are aggregated is unclear, particularly given the differences between them shown in fig. 10
L270: “averaged of the AOI”, I believe you mean “averaged zonally”
L272: It would be clearer to show this directly. An additional plot, plotting the average of different cloud properties against organisation would show whether this is the case. In addition, this statement refers to vertical cloud core properties, but only CTH has been assessed. It would be interesting to see how e.g. convective core height and anvil height vary as well
L273-274: The claim here is not well supported, it would be clearer to show this directly. An additional plot, plotting the average of different cloud properties against organisation would show whether this is the case
L276: is “MCS” here referring to overall organisation (which is inconsistent with earlier usage)? If not, statement is not supported as no assessment of the cloud types in this regime has been performed
L297: I wouldn’t describe this as a linear decrease as it appears more like a rapid drop from May to June
L299: I don’t think the trends are clear enough to say that there is a relative increase, just that organisation tends to be higher over the ocean
Figure 13: Unclear that the percentiles refer to organisation metrics. Are the percentiles calculated overall or individually for each month? If the latter, then this could cause differences not because the DCC properties are changing for a given level of organisation, but because the average organisation is changing and hence the percentiles are at different organisation values
L303: Number of DCCs doesn’t appear to trend over time, while core size increases except for SH-P90.
L315: While Welsh’s t-test can better handle distributions with different sizes and variances, it does not account for skewness in the distributions.
Figure 14: It’s difficult to interpret how meaningful these results are. As they only represent the 90th percentile, it is difficult to see how much they differ from the general population of DCCs. It would be helpful to plot the distributions of all observed systems, as well as the 90th percentiles, to show how the properties of the most organised systems differ
L352: There was not much discussion of the core/area ratio.
L361: The claim here is that grouping by organisation allows significant differences to be shown between different regions and different seasons, but are the differences statistically significant for all levels of organisation?
L366: Agreed, and I believe with your dataset you have the opportunity to progress this, which would greatly improve the value of this and future works. For instance, by focusing on more considered comparison between cloud properties and the organised sate (as in Figure 13), or by considering key synoptic or flow regimes separately rather than the somewhat arbitrary partitioning along hemisphere or latitude, or by making better use of 4D radiances with vertical cloud and core properties and development information. What additional properties would be useful for future studies?
L384-385: To achieve this, I recommend also considering the above in your analysis work. There were a few missed opportunities in this, mainly descriptive, work. Using a better measure of organisation alone will not elucidate the relationships you seek. Further, assessment of the tracked cloud distribution and sample size in space and time (diurnal and seasonal) will help to clarify which of your results are most robust, and which are subject to most uncertainty in the organisation metrics.
Technical corrections:
L84: both “Section” and “Sect.” are used in this passage. Recommend “sec.” as an acronym for section instead
L143: COP accounts for the areas of both objects i and j as per eq. 2
L186: “zonal changes” should be “zonal mean changes” or “latitudinal variations” or similar
L210: typo “custered” -> clustered
Citation: https://doi.org/10.5194/egusphere-2025-376-RC2
Data sets
Convective organisation indices based on 3D radar reflectivities Sarah Brüning https://doi.org/10.5281/zenodo.14724869
Model code and software
Analysing spatial-temporal patterns of convective organisation from 3D data Sarah Brüning https://doi.org/10.5281/zenodo.14710228
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 109 | 34 | 7 | 150 | 6 | 5 |
- HTML: 109
- PDF: 34
- XML: 7
- Total: 150
- BibTeX: 6
- EndNote: 5
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1