the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 23 Sep 2021
CycloneDetector (v1.0) – Algorithm for detecting cyclone and anticyclone centers from mean sea level pressure layer
Abstract. Automatic methods for identifying and tracking cyclones were firstly constructed in 1990's and since then there was a big increase in a precision and probability of detection. These methods have been traditionally focused on cyclones (and particularly on tropical cyclones), but the question of anticyclone centers detection remained unsolved since they are usually not a source of turbulent weather, precipitation etc. However, this issue can be important in the era of the climate change. In this paper, an algorithm for an automatic detection of both, cyclones and anticyclones based on mean sea level pressure field, is presented. The algorithm uses two-dimensional raster data as an input and returns a list of detected pressure systems. The main advantages of our solution are easy implementation since it is based on the standard image processing algorithm, sufficient performance of the algorithm, and especially the possibility of high-pressure systems detection. Moreover, the presented solution does not need a direct terrain filtering needed for some algorithms to be done. To validate the quality of detection algorithm results, a comparison against manually prepared data by Met Office was used. It follows from the comparison that the presented algorithm produces results similar to those by Met Office. The most significant differences can be found in the detection of cyclones at the beginning or the end of the lifespan stage. Met Office detects more cyclones in these stages than the presented solution.
This preprint has been withdrawn.
-
Withdrawal notice
This preprint has been withdrawn.
-
Preprint
(2549 KB)
Interactive discussion
Status: closed
-
RC1: 'Comment on gmd-2021-266', Anonymous Referee #1, 19 Oct 2021
General comments:
In this manuscript the authors present an algorithm for detecting both cyclone and anticyclone centres on a Sea Level Pressure chart. This constitutes an interesting subject but unfortunately, as a climatologist, I cannot say that I would trust the method presented, at least in its current version. The method is based on:
- Computing the isobars.
- I assume (not completely clear from the text) that in a second step, “closed” isobars are searched for and selected as “system candidates”.
- What (accordingly to the text) are “clearly not supposed to be pressure system centres” (the technical meaning of such a claim should be clarified) are eliminated based on the area enclosed by the contour, which is approximated by the axis-aligned bounding box (AABB). The authors do not offer any convincing clue on the precision of this approximation or the influence of the latitude or projection on the relative error. Could this approximation introduce a latitude-dependent bias in the algorithm?
- The remaining “systems” are classified as local maxima or minima based on the application of a mask of variable radius. The size of the radius can be set manually or automatically estimated. The authors recommend a value between 5 and 50 “pixels” (do the authors mean grid-points?) (line 156) but the only clue they offer to support such affirmation is that it is based on “their experiments”. These “experiments” should be detailed, and specific sensibility test included, otherwise it is not clear to the user the effect of this variable on the results.
As a test of the performance, the authors show a global map for a specific day (Figure 5) which seems reasonable (but with several caveats, see below) and then a comparison of their results with two Met Office analysis for the North Atlantic. Again, with reasonable results. The three cases shown are at least suggestive. But the practical utility of an algorithm aimed to track pressure systems relies on the confidence a researcher can put on it, and at the current state, the algorithm presented by Prantl et al leaves many questions open as:
- It is difficult to believe that the performance of the algorithm is not affected by the projection used (as it is claimed by the authors). At least, the relation between the area of a centre and the area of its corresponding AABB could be projection-dependent and the latter area is part of the rejection procedure. The authors should provide a sensitivity test on this issue. In my opinion, the algorithm should be optimized for a fixed projection and resolution and then convert any input array to meet the optimal algorithm requirements. Otherwise for the same meteorological situation, the user could obtain different outputs depending on the projection or the resolution.
- Apparently, the results depend on the isobar spacing (in fact it is one of the user selectable parameters). In my opinion, an automated method for detecting pressure systems should not be dependent upon the step at which the isolines are drawn. The area “T” enclosed by a “valid” system is again user-selectable. I assume this area also depends on the isobar step. The authors recommend using between 10.000 and 20.000 km2, but this seems rather arbitrary, and it is difficult to believe (in absence of a specific test) that this area can be considered the same for low- and high-pressure centres. Figure 4 clearly shows that the number of systems detected heavily depends on the area threshold “T” and the isobar step chosen (see their Figure 4).
- It is not completely clear (the contours are not labelled nor coloured) but from Figure 5, apparently some important centres are not detected (as one in the middle of the equatorial Pacific or a very clear intense cyclone next to the Kamchatka Peninsula) while other, probably less meteorologically relevant (along Central America, for example) are classified as pressure centres by the algorithm. The list given here is not exhaustive.
- Apart of the three case-studies of Figures 5, 6 and 7, the authors evaluate the performance of their results by comparing them with a sample of 385 Met Office Analysis. They obtain a reasonable match (92% of the detected systems are also found in Met Office maps and conversely, 83% of the Met Office systems are found by the algorithm). Although these figures are reasonable, in my opinion the analysis is somewhat simplistic. First because the performance test is only done over the relatively organised North Atlantic (there are areas of the world much more complex as Africa or sub-tropical latitudes over land) and second, because the authors only use the number of detected systems to check the method.
Summarising, the authors’ idea is interesting, but the algorithm given here is too simple and the performance tests are not convincing. The problem of localising by means of automatic algorithms meteorologically meaningful low- or high-pressure centres is, believe it or not, one of the most difficult in meteorological analysis. This is the reason because previous research on this topic usually considers not only the pressure field, but vorticity, local gradients, several levels, etc… because one needs more than a “large enough” local maxima or minima in a gridded SLP dataset to deal with the complexities of the synoptic scale meteorology.
Specific comments:
The quality of Figures 1 to 3 is not suitable for publication in an EGU journal. Figure 4 is clearly “over engineered”.
Citation: https://doi.org/10.5194/gmd-2021-266-RC1 -
RC2: 'Comment on gmd-2021-266', Anonymous Referee #2, 19 Nov 2021
The authors present in this manuscript a cyclone (anticyclone) tracking algorithm based on SLP analysis. Although the variable has been widely used in other studies of cyclones, the method could be interesting. According to the manuscript, the method can be summarized as follows: search the closed isobars and in order to remove unwanted areas, a filtering based on the area of the contour area is used with some considerations about the projection of the areas (AABB). Perhaps we could say that the method has somehow, similarities with a clustering approach using isobars as centroids (Kmeans). Such a method might not be entirely accurate in the case of multiple cyclones in small areas, particularly when they may have different lifetimes; it seems to underestimate the complexity of these events and their dynamics.
I find, according to what is expressed in the manuscript, that the algorithm needs decisions in several steps that are somewhat arbitrary and / or depend on the user. For example, the area “T” enclosed is user-selectable, the threshold value can be adjusted by the user, the isobar spacing etc. I don't think that an algorithm for cyclone detection should strongly depend on users. This could lead to very dissimilar results depending on the options chosen, and therefore not valid for comparisons. Precisely, the idea of ââdeveloping powerful algorithms to track and study these events should contribute to producing increasingly detailed statistical and dynamic analyzes that are comparable with others already performed (whether based on SLP, vorticity, etc.).
It is also important, as the referee # 1 mentioned, that the dependence (no dependence) of the results on the projections used (if no evidence can be show to support it), should be at least tested with a wide number of events. I believe that more powerful and well-founded arguments are needed to support this idea.
I must also agree with the previous referee that the analysis of the results is simple and at least insufficient. The idea of ââvalidating the algorithm should cover broader areas of study (there are all over the world), and not just be based on number of events.
There are several methods used to track cyclones and some recently published. In these methods, the algorithm is automated and also provides substantial information regarding not only the position of these events but also lifetimes and trajectories.
I believe that the method proposed here, even though the idea could be interesting, should be improved and completed in order to contribute to the methods already in use, that in addition to detecting low and high pressure centers, can provide trajectories, and lifetimes (very important characteristics for a cyclone analysis)
-
EC1: 'Comment on gmd-2021-266', Juan Antonio Añel, 22 Nov 2021
Dear authors,
After receiving the comments from the reviewers, I do not support the submission of a reviewed version of your manuscript. My decision is to reject it for publication in Geosc. Mod. Dev.
I hope that you understand the rationale for the decision in this case. We are happy of continuing to receive your research works for publication in our journal in the future.
Yours sincerely,
Juan A. Añel
Geosc. Mod. Dev. Executive Editor
Citation: https://doi.org/10.5194/gmd-2021-266-EC1
Interactive discussion
Status: closed
-
RC1: 'Comment on gmd-2021-266', Anonymous Referee #1, 19 Oct 2021
General comments:
In this manuscript the authors present an algorithm for detecting both cyclone and anticyclone centres on a Sea Level Pressure chart. This constitutes an interesting subject but unfortunately, as a climatologist, I cannot say that I would trust the method presented, at least in its current version. The method is based on:
- Computing the isobars.
- I assume (not completely clear from the text) that in a second step, “closed” isobars are searched for and selected as “system candidates”.
- What (accordingly to the text) are “clearly not supposed to be pressure system centres” (the technical meaning of such a claim should be clarified) are eliminated based on the area enclosed by the contour, which is approximated by the axis-aligned bounding box (AABB). The authors do not offer any convincing clue on the precision of this approximation or the influence of the latitude or projection on the relative error. Could this approximation introduce a latitude-dependent bias in the algorithm?
- The remaining “systems” are classified as local maxima or minima based on the application of a mask of variable radius. The size of the radius can be set manually or automatically estimated. The authors recommend a value between 5 and 50 “pixels” (do the authors mean grid-points?) (line 156) but the only clue they offer to support such affirmation is that it is based on “their experiments”. These “experiments” should be detailed, and specific sensibility test included, otherwise it is not clear to the user the effect of this variable on the results.
As a test of the performance, the authors show a global map for a specific day (Figure 5) which seems reasonable (but with several caveats, see below) and then a comparison of their results with two Met Office analysis for the North Atlantic. Again, with reasonable results. The three cases shown are at least suggestive. But the practical utility of an algorithm aimed to track pressure systems relies on the confidence a researcher can put on it, and at the current state, the algorithm presented by Prantl et al leaves many questions open as:
- It is difficult to believe that the performance of the algorithm is not affected by the projection used (as it is claimed by the authors). At least, the relation between the area of a centre and the area of its corresponding AABB could be projection-dependent and the latter area is part of the rejection procedure. The authors should provide a sensitivity test on this issue. In my opinion, the algorithm should be optimized for a fixed projection and resolution and then convert any input array to meet the optimal algorithm requirements. Otherwise for the same meteorological situation, the user could obtain different outputs depending on the projection or the resolution.
- Apparently, the results depend on the isobar spacing (in fact it is one of the user selectable parameters). In my opinion, an automated method for detecting pressure systems should not be dependent upon the step at which the isolines are drawn. The area “T” enclosed by a “valid” system is again user-selectable. I assume this area also depends on the isobar step. The authors recommend using between 10.000 and 20.000 km2, but this seems rather arbitrary, and it is difficult to believe (in absence of a specific test) that this area can be considered the same for low- and high-pressure centres. Figure 4 clearly shows that the number of systems detected heavily depends on the area threshold “T” and the isobar step chosen (see their Figure 4).
- It is not completely clear (the contours are not labelled nor coloured) but from Figure 5, apparently some important centres are not detected (as one in the middle of the equatorial Pacific or a very clear intense cyclone next to the Kamchatka Peninsula) while other, probably less meteorologically relevant (along Central America, for example) are classified as pressure centres by the algorithm. The list given here is not exhaustive.
- Apart of the three case-studies of Figures 5, 6 and 7, the authors evaluate the performance of their results by comparing them with a sample of 385 Met Office Analysis. They obtain a reasonable match (92% of the detected systems are also found in Met Office maps and conversely, 83% of the Met Office systems are found by the algorithm). Although these figures are reasonable, in my opinion the analysis is somewhat simplistic. First because the performance test is only done over the relatively organised North Atlantic (there are areas of the world much more complex as Africa or sub-tropical latitudes over land) and second, because the authors only use the number of detected systems to check the method.
Summarising, the authors’ idea is interesting, but the algorithm given here is too simple and the performance tests are not convincing. The problem of localising by means of automatic algorithms meteorologically meaningful low- or high-pressure centres is, believe it or not, one of the most difficult in meteorological analysis. This is the reason because previous research on this topic usually considers not only the pressure field, but vorticity, local gradients, several levels, etc… because one needs more than a “large enough” local maxima or minima in a gridded SLP dataset to deal with the complexities of the synoptic scale meteorology.
Specific comments:
The quality of Figures 1 to 3 is not suitable for publication in an EGU journal. Figure 4 is clearly “over engineered”.
Citation: https://doi.org/10.5194/gmd-2021-266-RC1 -
RC2: 'Comment on gmd-2021-266', Anonymous Referee #2, 19 Nov 2021
The authors present in this manuscript a cyclone (anticyclone) tracking algorithm based on SLP analysis. Although the variable has been widely used in other studies of cyclones, the method could be interesting. According to the manuscript, the method can be summarized as follows: search the closed isobars and in order to remove unwanted areas, a filtering based on the area of the contour area is used with some considerations about the projection of the areas (AABB). Perhaps we could say that the method has somehow, similarities with a clustering approach using isobars as centroids (Kmeans). Such a method might not be entirely accurate in the case of multiple cyclones in small areas, particularly when they may have different lifetimes; it seems to underestimate the complexity of these events and their dynamics.
I find, according to what is expressed in the manuscript, that the algorithm needs decisions in several steps that are somewhat arbitrary and / or depend on the user. For example, the area “T” enclosed is user-selectable, the threshold value can be adjusted by the user, the isobar spacing etc. I don't think that an algorithm for cyclone detection should strongly depend on users. This could lead to very dissimilar results depending on the options chosen, and therefore not valid for comparisons. Precisely, the idea of ââdeveloping powerful algorithms to track and study these events should contribute to producing increasingly detailed statistical and dynamic analyzes that are comparable with others already performed (whether based on SLP, vorticity, etc.).
It is also important, as the referee # 1 mentioned, that the dependence (no dependence) of the results on the projections used (if no evidence can be show to support it), should be at least tested with a wide number of events. I believe that more powerful and well-founded arguments are needed to support this idea.
I must also agree with the previous referee that the analysis of the results is simple and at least insufficient. The idea of ââvalidating the algorithm should cover broader areas of study (there are all over the world), and not just be based on number of events.
There are several methods used to track cyclones and some recently published. In these methods, the algorithm is automated and also provides substantial information regarding not only the position of these events but also lifetimes and trajectories.
I believe that the method proposed here, even though the idea could be interesting, should be improved and completed in order to contribute to the methods already in use, that in addition to detecting low and high pressure centers, can provide trajectories, and lifetimes (very important characteristics for a cyclone analysis)
-
EC1: 'Comment on gmd-2021-266', Juan Antonio Añel, 22 Nov 2021
Dear authors,
After receiving the comments from the reviewers, I do not support the submission of a reviewed version of your manuscript. My decision is to reject it for publication in Geosc. Mod. Dev.
I hope that you understand the rationale for the decision in this case. We are happy of continuing to receive your research works for publication in our journal in the future.
Yours sincerely,
Juan A. Añel
Geosc. Mod. Dev. Executive Editor
Citation: https://doi.org/10.5194/gmd-2021-266-EC1
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 1,307 | 510 | 58 | 1,875 | 36 | 45 |
- HTML: 1,307
- PDF: 510
- XML: 58
- Total: 1,875
- BibTeX: 36
- EndNote: 45
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
