Classification of flying insects in polarimetric weather radar using machine learning and aphid trap data

Kwakye, Samuel; Kalesse-Los, Heike; Maahn, Maximilian; Seifert, Patric; van Klink, Roel; Wirth, Christian; Quaas, Johannes

doi:https://doi.org/10.5194/amt-2023-69

Preprints

https://doi.org/10.5194/amt-2023-69

Preprints

11 Apr 2023

| 11 Apr 2023

Status: this preprint was under review for the journal AMT. A revision for further review has not been submitted.

Classification of flying insects in polarimetric weather radar using machine learning and aphid trap data

Samuel Kwakye, Heike Kalesse-Los, Maximilian Maahn, Patric Seifert, Roel van Klink, Christian Wirth, and Johannes Quaas

Abstract. Over the past decades, studies have observed strong declines in biomass and the abundance of flying insects. However, there are many locations where no surveys of insect biomass or abundance are available. Weather radars are known to provide quantitative estimates of flying insect biomass and abundance, and can therefore be used to fill knowledge gaps in space and time. In this study, we go beyond previous studies by combining a machine-learning approach with ground- truth observations from an aphid trap network. In this study, radar echoes from Level-II (Base) data of three Next Generation Weather Radar (NEXRAD) stations in the U.S. are classified using machine learning approaches. Weekly aphid counts from suction traps at Manhattan (Kansas), Morris (Illinois), and Rosemount (Minnesota) are used as validation data. Variability and distribution of the radar signals of four scatterer classes (insects, light rain, heavy rain, and plankton) are assessed. Probability density functions (PDF) of reflectivities of insects and plankton were found to be distinct from those of light- and heavy rain. Furthermore, the PDF of radar variables of the insect scatter class was also characterized by a broad distribution of spectrum width, cross-correlation ratio, and a broad range of differential reflectivity values. Decision trees, random forests, and support vector machine models were generated to distinguish three combinations of scatterers. A random forest classifier is found to be the best-performing model.

Received: 05 Apr 2023 – Discussion started: 11 Apr 2023

Download & links

Samuel Kwakye et al.

Status: closed (peer review stopped)

RC1:
'Comment on amt-2023-69', Anonymous Referee #1, 16 May 2023

Results of classification of radar echoes are presented in the manuscript. The classification categories are rain, atmospheric biota, and plankton. The authors have applied three machine learning algorithms to distinguish these radar returns. Data from three American Nexrad radars have been used in the study.
I am confused with the authors’ approach. On one hand, it follows from the manuscript’s title and main results (i.e., Fig. 5) that this is a study of classification of radar echoes using the machine learning algorithms. This implies that the input data are some variables from the initial base radar products, i.e., reflectivity, ZDR, Doppler velocity, spectrum width, differential phase, and correlation coefficient. On the other hand, results of echo classification from the Nexrad system are used first by the authors as an input to the authors’ algorithms (lines 77-78). The Nexrad radars have the Hydrometeor Classification Algorithm (HCA) which produces numerous classes of radar returns including biological and rain echoes. Have the authors used the HCA output to preselect types of radar echoes as it follows from ll. 77-78? The HCA is based on a fuzzy logic approach. It would be of interest to compare classification results from the HCA and the authors’ approach if the latter uses the base radar data but not the HCA outputs. If my understanding is not correct, i.e., the HCA output is not used to preselect types of echoes, the manuscript should contain comparisons of the classes obtained from the authors’ algorithms and HCA. At least, the manuscript should contain comparisons of insect identifications using the HCA and authors’ algorithms. If it turns out that the HCA exhibits a better performance than the authors’ algorithms, then the latter cannot be considered as a replacement for HCA.
Classification categories considered in the manuscript are produced by the HCA in more detail. For instance, the Nexrad radars measure the rain rate, which is more valuable data than just light and heavy rain used by the authors. Also, I have never seen the term ‘plankton’ used by the authors in their classification of Nexrad echoes. The category ‘plankton’ is absent in the HCA. The authors define the plankton as clear air returns at a wind speed > 11.3 m/s (line 94). If this is clear air (is this Bragg scatter?), why the term plankton is introduced? This term implies that clear air is filled with some substances. Since the radar data from warm seasons were used in the study, such echoes are typically produced by insects as well, but not tree leaves as the authors explain.
A question can be asked about comparisons of aphids in the ground traps and insects/birds causing radar returns. The main radar return comes from heights about 1 km above the ground at the ranges indicated in the manuscript. Is there any evidence that insects near the ground and those at a height of 1 km are the same species? Nexrad data frequently shows a height dependence of reflectivity from insects, which could be an indication of different taxa. Also, could birds be present aloft and contribute to radar returns?
The manuscript contains numerous not correct radar parameters and terms, but there is no need to discuss them now because the main authors’ approach is not clear.

Citation: https://doi.org/10.5194/amt-2023-69-RC1
- AC1:
  'Reply on RC1', Samuel Kwakye, 01 Jun 2023
  
  Thank you for your comment. The idea of comparing NEXRAD's HCA output with the output of the machine learning algorithm is profound. Shedding more light on the scatterers, especially plankton; in various literature, plankton has been considered as minute flying atmospheric biota but here plankton was used as a term for lifted leaves and debris as a result of high wind speed conditions not just clear air Bragg scatterers.
  The performance of machine learning algorithms is sensitive and conforms to the input variables/data. To address this; specific insect occurrences were identified from aphids suction trap data and the associated radar scans were further processed using the NEXRAD HCA. These two validation steps were seen to ensure that the radar moments of the echoes were insects. The same procedure was used for the weather echoes and plankton. The HCA does not classify plankton but classifies specifically ground clutter, biota, and various weather phenomena; the unclassified echoes were presumed and used as the plankton classification. Two base radar moments (reflectivity and spectrum width) and three polarimetric radar moments(cross-correlation coefficient, differential reflectivity, and specific differential phase) were used.
  The question of the height of insects radar returns was considered in the selection of the sites as the height of a radar beam is a function of the range. Suction traps of the Midwest suction trap network as used in the study were within the range of 50km of the nearest NEXRAD radar.
  
  Citation: https://doi.org/10.5194/amt-2023-69-AC1
  - RC3: 'Reply on AC1', Anonymous Referee #1, 28 Jun 2023
    
    I do not see a compelling reason on the use of NEXRAD’s HCA output to process it further with the machine learning algorithms (MLA). The authors found one of the MLAs with the best performance. What new information can we get from this finding? The HCA has already classified a case as “biological” echo. Jatau et al. JTECH, 2021 proposed a MLA to classify such echoes as caused by insects or birds. Their MLA is based on the base radar products and can be considered as a development of the NEXRAD’s products. It is not clear how the manuscript’s findings can be used to get new information because they are based on the NEXRAD’s HCA.
    The authors claim that aphids in the ground traps prove their presence aloft. No evidence is presented in the manuscript on that and this remains the authors’ guess. It is known that different species can fly at various heights in the atmosphere (i.e., the monograph by V. Drake and D. Reynolds). The authors’ argument that the traps are located within the range of 50 km from radars seems weak. The center of radar beam at 50 km is at a height of 440 m. The species at this height can differ from species at the ground.
    Another authors’ guess is the “plankton” radar echo. The authors claim that this echo is caused by lifted leaves at the with speed > 11.3 m/s. No evidence is presented on the source of this echo. There is evidence that strong winds leads to more turbulent atmospheric layers, which can create Bragg scatter (i.e., Davison et al., JAS, 2013, p. 3047; Richardson et al., JTECH, 2017, p.479). The authors rule out the presence of Bragg scatter and insects at that wind speed, but do not present any evidence that the echo is caused by lifted leaves.
    
    Citation: https://doi.org/10.5194/amt-2023-69-RC3
    
    AC3: 'Reply on RC3', Samuel Kwakye, 28 Nov 2023
    
    Thank you for your questions and comments. The question on the significance of using the HC algorithm to process the radar data before machine learning classification was a recommendation suggested by Jatau et al., 2021 in the classification of radar echoes from different scatterers. Aphids are extensively known as crop pest in the Midwest of United States croplands where the climatology and weather conditions are seen to favor aphids’ assemblage and abundance in addition to the cropland and vegetation cover shown in figure 1 which has led to the establishment of a suction trap network Lagos-Kutz, Doris, et al. 2020. The phenology and migration characteristics of aphids was not the focus of the study and as stated the aphid occurrence as used in the study were selected from the best estimates for the days of maximum aphid counts from the suction traps, and optimal weather conditions for aphids’ assemblage and abundance. These factors proves their presence aloft in the vicinity of the weather radars. The atmospheric conditions used to describe uplifting of leaves debris i.e. plankton is fundamentally recognized by the Beaufort scale at that wind speed.
    
    Citation: https://doi.org/10.5194/amt-2023-69-AC3
RC2:
'Comment on amt-2023-69', Anonymous Referee #2, 31 May 2023

GENERAL COMMENTS
The manuscript studies machine learning methods in classification of insects in polarimetric weather radar echoes. Weekly trap catches of aphids are used in selection of periods for the weather radar data, but are these really "validation data" for this study? Much talk about insects' abundance changes, possibly related to climate change is included as an introduction of using weather radar data for this purpose, but I see no critical view of the method. For instance the size of some aphid species may vary a lot.
SPECIFIC COMMENTS
1. The role of trap data does not seem to be very important, and figure 1 showing the land use around the traps has no meaning at all while there is no selection of periods accroding to wond direction or any other criteria considered. Only purpose for the figure seems to be to show that the radar close by has similar kind of natural environment, and this could be more precisely be commented in the text, while the population characteristics of even one aphid species are in any case quite obscure for the reader based only on the maps.
2. The scatterer classes were selected usinng, perhaps best estimates for the days of maximum aphid abundance, and local weather data. There may be a problem in aphid data, because of the relatively short period winged aphids are developed and that migrate in the radar monitored air space, and on the other hand because of possible long range migrations that occur unrelated to local weather. In other scatterer classes the local meteorological observations seem to give the "ground truth", but there is no exact description on how these observations were geographically co-located with radars.
3. The "plankton" class, this may be more precisely described in some of your references, but if 22 knots is a wind speed limit, is there any idea of how abundant this kind of debris could be in the radar monitored air space compared to the actual reflectivities?
4. You comment a reference stating much larger values of differential reflectivity of insects than what you find, and "small size and plump shape of aphids" is your explanation. However, aphids may have very elongated bodies as welll, in the trap data you may find more precise information on this, the species and its shape. However, there are other things involved as well, as in insect migrations common orientation may be quite common, and this means differences depensing on the azimuth the radar data is gathered from.

Citation: https://doi.org/10.5194/amt-2023-69-RC2
- AC2: 'Reply on RC2', Samuel Kwakye, 01 Jun 2023
  
  Thank you for your question and suggestion. The general question about aphids count being used as reference data for insect occurrence can be much more piercing if it pointed to the temporal differences with the radar data. Specifically the aphids count data gives a definite fact that there are insects in the aerosphere within the period of trapping and counting. Aphids are widely known as crop pest and the plots of land cover in the vicinity and coverage of the suction trap and radar at the three sites as seen in figure 1 are cropland and vegetation cover which are favorable for aphids.
  The phenology and migration characteristics of aphids was not the focus of the study and as stated the insect occurrence were selected from the best estimates for the days of maximum aphid counts, and local weather data. The weather data although not described in the manuscript were obtained from the nearest weather station to the radars with their geographical coordinates.
  To understand your question on plankton; plankton as described as lifted debris and leaves litter are significant to echo reflectivity power on the scale of the insect and weather scatterers. The radar cross section of plankton to obtain quantitative estimates of the inhomogeneous nature of the plankton scatterer can be explored.
  As compared to other insects the symmetry of aphids shape in the horizontal and vertical is much higher than other insects hence the low values of differential reflectivity.
  
  Citation: https://doi.org/10.5194/amt-2023-69-AC2

Status: closed (peer review stopped)

RC1:
'Comment on amt-2023-69', Anonymous Referee #1, 16 May 2023

Results of classification of radar echoes are presented in the manuscript. The classification categories are rain, atmospheric biota, and plankton. The authors have applied three machine learning algorithms to distinguish these radar returns. Data from three American Nexrad radars have been used in the study.
I am confused with the authors’ approach. On one hand, it follows from the manuscript’s title and main results (i.e., Fig. 5) that this is a study of classification of radar echoes using the machine learning algorithms. This implies that the input data are some variables from the initial base radar products, i.e., reflectivity, ZDR, Doppler velocity, spectrum width, differential phase, and correlation coefficient. On the other hand, results of echo classification from the Nexrad system are used first by the authors as an input to the authors’ algorithms (lines 77-78). The Nexrad radars have the Hydrometeor Classification Algorithm (HCA) which produces numerous classes of radar returns including biological and rain echoes. Have the authors used the HCA output to preselect types of radar echoes as it follows from ll. 77-78? The HCA is based on a fuzzy logic approach. It would be of interest to compare classification results from the HCA and the authors’ approach if the latter uses the base radar data but not the HCA outputs. If my understanding is not correct, i.e., the HCA output is not used to preselect types of echoes, the manuscript should contain comparisons of the classes obtained from the authors’ algorithms and HCA. At least, the manuscript should contain comparisons of insect identifications using the HCA and authors’ algorithms. If it turns out that the HCA exhibits a better performance than the authors’ algorithms, then the latter cannot be considered as a replacement for HCA.
Classification categories considered in the manuscript are produced by the HCA in more detail. For instance, the Nexrad radars measure the rain rate, which is more valuable data than just light and heavy rain used by the authors. Also, I have never seen the term ‘plankton’ used by the authors in their classification of Nexrad echoes. The category ‘plankton’ is absent in the HCA. The authors define the plankton as clear air returns at a wind speed > 11.3 m/s (line 94). If this is clear air (is this Bragg scatter?), why the term plankton is introduced? This term implies that clear air is filled with some substances. Since the radar data from warm seasons were used in the study, such echoes are typically produced by insects as well, but not tree leaves as the authors explain.
A question can be asked about comparisons of aphids in the ground traps and insects/birds causing radar returns. The main radar return comes from heights about 1 km above the ground at the ranges indicated in the manuscript. Is there any evidence that insects near the ground and those at a height of 1 km are the same species? Nexrad data frequently shows a height dependence of reflectivity from insects, which could be an indication of different taxa. Also, could birds be present aloft and contribute to radar returns?
The manuscript contains numerous not correct radar parameters and terms, but there is no need to discuss them now because the main authors’ approach is not clear.

Citation: https://doi.org/10.5194/amt-2023-69-RC1
- AC1:
  'Reply on RC1', Samuel Kwakye, 01 Jun 2023
  
  Thank you for your comment. The idea of comparing NEXRAD's HCA output with the output of the machine learning algorithm is profound. Shedding more light on the scatterers, especially plankton; in various literature, plankton has been considered as minute flying atmospheric biota but here plankton was used as a term for lifted leaves and debris as a result of high wind speed conditions not just clear air Bragg scatterers.
  The performance of machine learning algorithms is sensitive and conforms to the input variables/data. To address this; specific insect occurrences were identified from aphids suction trap data and the associated radar scans were further processed using the NEXRAD HCA. These two validation steps were seen to ensure that the radar moments of the echoes were insects. The same procedure was used for the weather echoes and plankton. The HCA does not classify plankton but classifies specifically ground clutter, biota, and various weather phenomena; the unclassified echoes were presumed and used as the plankton classification. Two base radar moments (reflectivity and spectrum width) and three polarimetric radar moments(cross-correlation coefficient, differential reflectivity, and specific differential phase) were used.
  The question of the height of insects radar returns was considered in the selection of the sites as the height of a radar beam is a function of the range. Suction traps of the Midwest suction trap network as used in the study were within the range of 50km of the nearest NEXRAD radar.
  
  Citation: https://doi.org/10.5194/amt-2023-69-AC1
  - RC3: 'Reply on AC1', Anonymous Referee #1, 28 Jun 2023
    
    I do not see a compelling reason on the use of NEXRAD’s HCA output to process it further with the machine learning algorithms (MLA). The authors found one of the MLAs with the best performance. What new information can we get from this finding? The HCA has already classified a case as “biological” echo. Jatau et al. JTECH, 2021 proposed a MLA to classify such echoes as caused by insects or birds. Their MLA is based on the base radar products and can be considered as a development of the NEXRAD’s products. It is not clear how the manuscript’s findings can be used to get new information because they are based on the NEXRAD’s HCA.
    The authors claim that aphids in the ground traps prove their presence aloft. No evidence is presented in the manuscript on that and this remains the authors’ guess. It is known that different species can fly at various heights in the atmosphere (i.e., the monograph by V. Drake and D. Reynolds). The authors’ argument that the traps are located within the range of 50 km from radars seems weak. The center of radar beam at 50 km is at a height of 440 m. The species at this height can differ from species at the ground.
    Another authors’ guess is the “plankton” radar echo. The authors claim that this echo is caused by lifted leaves at the with speed > 11.3 m/s. No evidence is presented on the source of this echo. There is evidence that strong winds leads to more turbulent atmospheric layers, which can create Bragg scatter (i.e., Davison et al., JAS, 2013, p. 3047; Richardson et al., JTECH, 2017, p.479). The authors rule out the presence of Bragg scatter and insects at that wind speed, but do not present any evidence that the echo is caused by lifted leaves.
    
    Citation: https://doi.org/10.5194/amt-2023-69-RC3
    
    AC3: 'Reply on RC3', Samuel Kwakye, 28 Nov 2023
    
    Thank you for your questions and comments. The question on the significance of using the HC algorithm to process the radar data before machine learning classification was a recommendation suggested by Jatau et al., 2021 in the classification of radar echoes from different scatterers. Aphids are extensively known as crop pest in the Midwest of United States croplands where the climatology and weather conditions are seen to favor aphids’ assemblage and abundance in addition to the cropland and vegetation cover shown in figure 1 which has led to the establishment of a suction trap network Lagos-Kutz, Doris, et al. 2020. The phenology and migration characteristics of aphids was not the focus of the study and as stated the aphid occurrence as used in the study were selected from the best estimates for the days of maximum aphid counts from the suction traps, and optimal weather conditions for aphids’ assemblage and abundance. These factors proves their presence aloft in the vicinity of the weather radars. The atmospheric conditions used to describe uplifting of leaves debris i.e. plankton is fundamentally recognized by the Beaufort scale at that wind speed.
    
    Citation: https://doi.org/10.5194/amt-2023-69-AC3
RC2:
'Comment on amt-2023-69', Anonymous Referee #2, 31 May 2023

GENERAL COMMENTS
The manuscript studies machine learning methods in classification of insects in polarimetric weather radar echoes. Weekly trap catches of aphids are used in selection of periods for the weather radar data, but are these really "validation data" for this study? Much talk about insects' abundance changes, possibly related to climate change is included as an introduction of using weather radar data for this purpose, but I see no critical view of the method. For instance the size of some aphid species may vary a lot.
SPECIFIC COMMENTS
1. The role of trap data does not seem to be very important, and figure 1 showing the land use around the traps has no meaning at all while there is no selection of periods accroding to wond direction or any other criteria considered. Only purpose for the figure seems to be to show that the radar close by has similar kind of natural environment, and this could be more precisely be commented in the text, while the population characteristics of even one aphid species are in any case quite obscure for the reader based only on the maps.
2. The scatterer classes were selected usinng, perhaps best estimates for the days of maximum aphid abundance, and local weather data. There may be a problem in aphid data, because of the relatively short period winged aphids are developed and that migrate in the radar monitored air space, and on the other hand because of possible long range migrations that occur unrelated to local weather. In other scatterer classes the local meteorological observations seem to give the "ground truth", but there is no exact description on how these observations were geographically co-located with radars.
3. The "plankton" class, this may be more precisely described in some of your references, but if 22 knots is a wind speed limit, is there any idea of how abundant this kind of debris could be in the radar monitored air space compared to the actual reflectivities?
4. You comment a reference stating much larger values of differential reflectivity of insects than what you find, and "small size and plump shape of aphids" is your explanation. However, aphids may have very elongated bodies as welll, in the trap data you may find more precise information on this, the species and its shape. However, there are other things involved as well, as in insect migrations common orientation may be quite common, and this means differences depensing on the azimuth the radar data is gathered from.

Citation: https://doi.org/10.5194/amt-2023-69-RC2
- AC2: 'Reply on RC2', Samuel Kwakye, 01 Jun 2023
  
  Thank you for your question and suggestion. The general question about aphids count being used as reference data for insect occurrence can be much more piercing if it pointed to the temporal differences with the radar data. Specifically the aphids count data gives a definite fact that there are insects in the aerosphere within the period of trapping and counting. Aphids are widely known as crop pest and the plots of land cover in the vicinity and coverage of the suction trap and radar at the three sites as seen in figure 1 are cropland and vegetation cover which are favorable for aphids.
  The phenology and migration characteristics of aphids was not the focus of the study and as stated the insect occurrence were selected from the best estimates for the days of maximum aphid counts, and local weather data. The weather data although not described in the manuscript were obtained from the nearest weather station to the radars with their geographical coordinates.
  To understand your question on plankton; plankton as described as lifted debris and leaves litter are significant to echo reflectivity power on the scale of the insect and weather scatterers. The radar cross section of plankton to obtain quantitative estimates of the inhomogeneous nature of the plankton scatterer can be explored.
  As compared to other insects the symmetry of aphids shape in the horizontal and vertical is much higher than other insects hence the low values of differential reflectivity.
  
  Citation: https://doi.org/10.5194/amt-2023-69-AC2

Samuel Kwakye et al.

Viewed

Total article views: 729 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
521	181	27	729	18	17

HTML: 521
PDF: 181
XML: 27
Total: 729
BibTeX: 18
EndNote: 17

Views and downloads (calculated since 11 Apr 2023)

Month	HTML	PDF	XML	Total
Apr 2023	231	52	4	287
May 2023	93	23	2	118
Jun 2023	54	25	10	89
Jul 2023	34	14	2	50
Aug 2023	25	15	1	41
Sep 2023	30	17	5	52
Oct 2023	20	15	1	36
Nov 2023	10	8	1	19
Dec 2023	19	12	1	32
Jan 2024	5	0	5

Cumulative views and downloads (calculated since 11 Apr 2023)

Month	HTML	PDF	XML	Total
Apr 2023	231	52	4	287
May 2023	93	23	2	118
Jun 2023	54	25	10	89
Jul 2023	34	14	2	50
Aug 2023	25	15	1	41
Sep 2023	30	17	5	52
Oct 2023	20	15	1	36
Nov 2023	10	8	1	19
Dec 2023	19	12	1	32
Jan 2024	5	0	5

Viewed (geographical distribution)

Total article views: 729 (including HTML, PDF, and XML) Thereof 729 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 05 Jan 2024

Short summary

Insect numbers in the atmosphere can be calculated using polarimetric weather radar but they have to be identified and separated from other echoes, especially weather phenomena. Here, the separation is demonstrated using three machine-learning algorithms and insect count data from suction traps and the nature of radar measurements of different radar echoes is revealed. Random forest is the best separating algorithm and insect echoes radar measurements are distinct.


Total:	0
HTML:	0
PDF:	0
XML:	0