Open Access

Application of insect songs in monitoring population density level of Locusta migratoria migratoria (Orthoptera: Acrididae)

Zoological Studies201453:55

DOI: 10.1186/s40555-014-0055-x

Received: 4 May 2014

Accepted: 13 August 2014

Published: 22 August 2014

Abstract

Background

The locusts Locusta migratoria migratoria (Orthoptera: Acrididae) is the most destructive agricultural pests worldwide, the population and distribution of L. migratoria migratoria growing rapidly in recent years. It is crucial to find a green, economical way to monitor this insect's population for effective control tactics. In this study, acoustic samples were recorded and analyzed under three different density levels of Asian migratory locust L. migratoria migratoria.

Results

The results showed that the songs of L. migratoria migratoria had a very stable acoustic feature in time domains; then, we used duration of pulse as a tool for identifying and counting the numbers of pulse to classify the population size. After removing the background noises, an automatic density classification and monitoring system was established based on the backpropagation (BP) neural network. The field sample test showed that the accuracy of the density level recognition reached 96.67%.

Conclusions

The results indicated that the calling songs of insects could be an effective character to distinguish population density level of locust plagues, and it could be potentially used as a green and environmental protection solution in monitoring the dynamics of locust plagues and other acoustic agriculture pests.

Keywords

Insect song Asian migratory locust Monitoring system

Background

The various subspecies of migratory locusts such as Locusta migratoria migratoria (Orthoptera: Acrididae) is the most destructive agricultural pests worldwide. In recent years, the areas damaged by L. migratoria migratoria have enlarged geographically in China (Ma et al. [2005]; Zhang et al. [2009]), and it has caused billions of dollars in property damage. Thus, knowing the dynamics of pest populations is crucial in determining effective control tactics, when to initiate the tactics, and the tactics, once implemented, are successful. Knowing how a population changes could also give us a better understanding of occurrence regularity of the species studied. The current monitoring methods for such species mainly depend on field observations and remote sensing data (Drake et al. [2002]; Ceccato et al. [2006]; Shi et al. [2003]). However, due to the high cost of radar and satellite sensor-based monitoring system, it is meaningful to find an effective and economical solution. The application of acoustic signals in monitoring pest populations has inherent advantages: It only needs some acoustic sensors set in the field with connection to a computer which is already programmed, and it is more economic than the satellite sensor-based monitoring system. Moreover, we are willing to establish a system that can automatically calculate the classification results without manual observations. Finally, the application of acoustic signals is also a pollution-free and environment-friendly method compared with some chemical methods such as pheromone traps.

Sound productions have great significance for intraspecific communication, such as interaction and courtship (Ragge and Reynolds [1998]), disturbance (García et al. [2003]), and performance in a chorus (Ewing [1989]). Most insects of Cicadidae and Orthoptera can make sounds, but the methods used to emit sounds are very diverse (Uvarov [1996]). The femoro-elytral method is the most widespread in suborder Caelifera, producing stridulation by rubbing the posterior femora against the tegmina (Schmidt and Stelzer [2005]). Other reports claimed, however, that some apterous grasshoppers developed their own methods to produce sounds (López et al. [2007]). The calling songs have a good specificity and steadiness that can be used as an effective indicator in species identification and evolutionary studies (Ngo and Ngo [2013]; Wang et al. [2011]; Hemp and Kehl [2010]; Montealegre-Z and Morris [2004]; Pace et al. [2010]). For more than several decades, researchers had used sounds produced by insects to detect their presence (Riede [1998]; Parkman et al. [1996]); in contrast, few attempts have been made to estimate population size and density. There is only one report about monitoring locust density based on some acoustic characters such as diurnal song activity, the song quality, and the audible distance, but this study is mainly focus on some ecological characters (Fischer et al. [1997]). Until now, there are no reports of precasting insect densities base on their intraspecies acoustic characters, especially in automatic identifications on a particular species.

In order to test whether the intraspecies acoustic characters could be used to estimate the population density of a particular species, the calling samples of L. migratoria migratoria in three different density levels were recorded and analyzed, and then an automatic density classification and monitoring system was established based on the backpropagation (BP) neural network. The results showed that insect songs could be potentially used as an effective and green method in monitoring locust plagues.

Methods

Insects

We used fourth instar larvae of L. migratoria migratoria as specimen, which were collected in Haituo (45° 23′ N, 129° 91′ E, 133 m in altitude), Da'an, Jilin province, Northeast China. The grassland in Haituo is covered with reeds over 500 ha, and nearly 60% was attacked by the Asian migratory locust. Gregarious nymphs were raised to adults in the lab, fed with fresh reed leaves, and kept in transparent containers (55 cm × 53 cm × 50 cm) at 30 ± 2°C with a 12:12-h light–dark cycle. The male/female ratio is nearly 1:1 per cage. The life cycle of L. migratoria migratoria in this place was also studied by field observation, and the records of the local government were checked. For that, only sexual maturity locust could make songs, and it was important to decide the proper time to monitor the density using this system.

Song recording and analysis

The experiment was divided into three groups, which were defined with actual locust infestation during the past several years according to the local government's records: slight locust plagues, 5 individuals per box (0.5 m3); moderate locust plagues, 25 individuals per box (0.5 m3); and heavy locust plagues, 50 individuals per box (0.5 m3). Within 2 months of collection, we recorded the calling songs of both males and females in our lab as sound files using a digital voice recorder (PCM-D50 Digital Recorder, Sony Corporation, Tokyo, Japan). The sampling rate was 96 kHz, and the recorder was placed in the center of the box. It was reported that the acoustic behaviors and the traits of songs changed with temperature (von Helversen [1972]), so the temperature was taken under natural sunlight during recording. Each recording sample took 1 h, which was divided into six 10-min samples; each density got 30 samples, and thus, a total of 90 samples were obtained. Thereinto, 45 samples were used as training samples to train the BP neural network, and the remaining 45 samples were used as testing samples to test the system we built. The width of the pulses, the intervals between pulses and pulses groups, the normalized amplitude of pulses, and the amplitude ratio between pulses were analyzed using Cool Edit (Cool Edit pro V2.1, Adobe Systems) and Matlab (Matlab 7.0, Mathworks) with at least 16 different individuals.

BP neural network

For the purpose of automatically classifying and monitoring the population density of L. migratoria migratoria pest, we used the backpropagation neural network. Standard backpropagation is a gradient descent algorithm, as is the Widrow-Hoff learning rule, in which the network weights are moved along the negative of the gradient of the performance function. The term ‘backpropagation’ refers to the manner in which the gradient is computed for nonlinear multilayer networks. There are a number of variations on the basic algorithm that are based on other standard optimization techniques, such as conjugate gradient and Newton methods.

We used three independent layers: input, hidden, and output layers. As for our purpose, after removing the background noise, the number of efficient pulses was used as input layers, and three different density levels were used as output layers. The number of input neurons depends on the dimension of input vector; in this text, the dimension of input vector was three-dimensional, so we used three neurons in input layers. The number of neurons in output layers was three; it was based on the three density levels of locust pest we mentioned above. The number of neurons in hidden layers was determined by the formula n = n i + n o + a , where n was the number of neurons in hidden layers, n i was the number of neurons in input layers, n o was the number of neurons in output layers, and a was a constant between 1 and 10 (we took 5 here, which was an intermediate value). Finally, we got n = 7.4, so we used seven neurons in hidden layers. The structure of the BP neural network that we used is shown in Figure 1.
Figure 1

The structure of the BP neural network. Thereinto, input layer n i = 3, output layer n o = 3, constant a = 7, and then we get hidden layer n = 7.4.

Data preprocessing

Background noise can be divided into stationary noise and impulse noise. The former can be conveniently removed by setting a proper threshold and clamping it. We set the normalized threshold as 0.1 (this threshold is based on the signal to noise (S/N) ratio); the latter cannot be removed by clamping because its amplitude of shortwave is similar to that of locust songs. Therefore, we converted the migratory locust songs and impulse noise into Bohr square wave after removing the stationary noise (Figure 2a,b,c). No overlapping pulses were mainly distributed between threshold 0.1 and 0.7. As will be mentioned in the ‘Results’ section, the calling songs of L. migratoria migratoria have an obvious character of double pulses, and the average width values of pulse A and pulse B were 15.45 and 17.74 ms, respectively (Table 1). In actual situations, most of the impulse noise differs greatly from locust songs in the width of pulse. The impulse noise would not be counted by cutting off the bad data. The impulse noises, which had the same width as locust songs in minor cases, should exert no effect on the statistical result.
Figure 2

Locusta migratoria migratoria calling songs with impulse noise. (a) The time domain plot of Locusta migratoria migratoria songs and impulse noise. (b) The corresponding short-time energy diagram. (c) The corresponding Bohr square waveform.

Table 1

Time domain song traits of L. migratoria migratoria

 

Number of samples

Average

Standard error

Relative error (%)

Utest value

Width of pulse A (ms)

187

15.45

0.25

1.7

5.329

Width of pulse B (ms)

187

17.74

0.36

2.1

Intervals between A and B (ms)

186

7.49

0.36

4.9

 

Intervals between double pulses (ms)

186

73.1

1.5

2.1

Amplitude of pulse A (normalized)

187

0.293

0.019

6.3

11.68

Amplitude of pulse B (normalized)

187

0.0736

0.0036

4.9

Amplitude ratio A/B

187

4.12

0.25

6.0

 

When the population density is very high, there maybe some overlapping pulses; when it occurs, it will not only increase the amplitude but also will prolong the durations of the pulses. To identify this situation, there are two symbols: firstly, the amplitude was over the normalized threshold 0.7; secondly, the duration of pulse A is between 15.45 and 30.90 ms, and pulse B is between 17.74 and 35.48 ms (more than one pulse time but less than two pulses time). For three or more pulses that overlap, it is really rare and will not take effect on the statistical results, so we did not use them when counting. All the preprocessing methods were programmed with Matlab (Matlab 7.0, Mathworks), and after processing, the counting number of the three thresholds was used as the input layers to the BP neural network.

Field sample test

After the system was established, another 90 field samples were collected in the same location (45° 23′ N, 129° 91′ E, 133 m in altitude) to examine the system. When recording, the recorder was placed on a 0.3 m high tripod with a 0.5 m2 ranged marker in the grassland, with a wind shelter on the recorder's microphone; each sample took an hour and divided into six 10-min samples. After recording, the locust density were counted by eye to define the level; after that, all the samples were subjected to data preprocessing and input to the system, and the output results were compared with man-made counting, which meant that the man-made results were completely blindly obtained before they were input to the system. We just counted the final number of the locust within 0.5 m2 which does not include the ‘in and out’ insects during the recording time.

Results

Song analysis

The calling songs of Asian migratory locust have an obvious character of double pulses (Figure 3). For convenience, the first pulse was called pulse A and the other one was called pulse B, the former having higher amplitude and narrower width than the latter. Moreover, the intervals between pulse A and pulse B were stable, and so are the intervals between double pulses. The S/N ratio (the ratio of signal/noise) is about 10. By measuring 187 pulse groups of 16 individuals, we got a statistical result of song traits (Table 1). As listed in Table 1, the U test value of pulse amplitude and pulse width between pulse A and pulse B are very large, suggesting a significant difference. The relative errors of all traits were 1.7% to 6.3%, reflecting a good stability. Therefore, the duration traits of pulses were used as parameters in our system.
Figure 3

Time domain plot of L. migratoria migratoria calling songs. L. migratoria migratoria's songs showed obviously double pulses, with the stable duration time in pulse A and pulse B, the intervals between A and B, and also the intervals between pulse groups.

Data preprocessing results

The counted results of the number of pulses in three different thresholds have been listed in Table 2, and each density resulted from 100 sound samples. The result indicated that the pulse number in the same threshold was significantly different among the three densities. Taking the data as network input and the density levels of locust pest as network output, we established the classification system.
Table 2

Count results of the number of pulse in three different thresholds after removing the bad data

Density (individuals/0.5 m2)

Pulse number (×103)

Threshold 0.1

Threshold 0.4

Threshold 0.7

5

1.060 ± 0.048

0.226 ± 0.011

0.0469 ± 0.0028

25

5.11 ± 0.11

1.380 ± 0.023

0.218 ± 0.068

50

10.26 ± 0.29

2.171 ± 0.062

0.416 ± 0.014

U test value

Between 5 and 25

337.0

357.6

25.1

Between 25 and 50

166.1

156.3

28.6

Between 5 and 50

312.9

309.0

258.7

The effective number of pulses in three different thresholds is listed in Table 2. After removing noise, U test showed that there is greatly significant differences in three densities, and when the threshold was over 0.7, the number of pulses in three densities were 0.0469 ± 0.0028, 0.218 ± 0.068, and 0.416 ± 0.014, respectively, which is much less than the value of the threshold between 0.4 and 0.7. It is suggested that when the density is high, the overlapping pulses are still very rare and take no big effect on the counting results, but, indeed, it was also increasing with the locust density. Half of the data were input to the BP neural network for training, and the rest were used as testing samples.

Training results of neural networks

For training neural networks, each density level randomly selected 15 samples from 30 samples, and the rest were the testing samples. Thus, the total number of training samples and testing samples was 45. We trained neural networks 600 times, and the training target was 0.01. Figure 4 shows the error curve of the network training. Table 3 shows the testing results of this BP neural network. As listed in Table 3, the correct rate of density level recognition reached 100%, which showed that the system we established worked correctly in the laboratory environment.
Figure 4

The error curve of neural network training. We trained neural networks 600 times; the training target was 0.01, and the correct rate of density level recognition was 100%.

Table 3

Testing results of BP neural network

 

Testing samples

Outputs

Density 1

Density 2

Density 3

Inputs

Density 1

15

15

0

0

Density 2

15

0

15

0

Density 3

15

0

0

15

Correct number of output

15

15

15

Field sample test

After the system was established, field samples were used to test whether this system could be used in actual situations. After field recording, finally, we got 90 recording samples; thereinto, 34 samples were identified as low density, 23 were medium and 11 were high. The correct recognition rate of our system in field working environment is listed in Table 4. It is shown that the system also could be used in the field even if the rate is a little bit fewer than the BP neural network testing results.
Table 4

Field sample testing results of monitoring system

 

Testing samples

Outputs

Low

Medium

High

Inputs

 Low

34

34

0

0

 Medium

23

0

21

2

 High

11

1

0

10

Correct rate of output

100 %

91.30 %

90.91 %

Total identified rate

96.67 %

Discussion

The calling songs of L. migratoria migratoria have an obvious character of double pulses which is not a rare phenomenon in Orthoptera insects (Greenfield [1990]). The detailed study had been done in katydid Neoconocephalus affinis (Orthoptera: Tettigoniidae). The Fourier analysis of the stimulus envelopes revealed that females respond only when both the first and second harmonics of the AM spectrum are of similar amplitude. The second harmonic is generated by the amplitude difference between the two pulses making up a pulse pair. Females respond to double pulses that have been merged into a single pulse only if this amplitude modulation is preserved (Bush et al. [2009]).

Insect songs had a good stability that is already used as a powerful tool in taxonomy, especially in some close-related species (Schul [1998]). Our result also suggested a stable acoustic trait in L. migratoria migratoria, which means that these methods could not only be used in L. migratoria migratoria but also in other acoustic insects such as katydids, crickets, and cicadas. Some studies showed that there was a differentiation of songs between geographically isolated populations (Heady and Denno [1991]; Paillette et al. [1997]). It is still not very clear for us whether there was a big differentiation of songs in L. migratoria migratoria; if so, we need to adjust the system to adapt to its changes in other locations, but it is convenient to just change some parameters.

There are several reasons why we choose the BP neural network as a statistical method: firstly, the calling songs of L. migratoria migratoria showed double pulses. Even if the calling time varied between different insects, the number of insects may not have a linear relationship with the number of pulses. Secondly, the number in three thresholds took different effects on the classification results. To avoid this, we just input the data in different thresholds and let the BP neural network train the system. The results showed that the BP neural network can provide accurate and effectual information in population density recognition.

There are some points worthy of notice for practical application. Firstly, the calling songs of L. migratoria migratoria were the symbol of sexual maturity, i.e., only adults could make songs by flapping the forewings. This system can accurately recognize the density level of locust plague, but not for the early monitoring. For several years of field observation and combined with the previous record of the local government, we obtained adult emergence time in the field of L. migratoria migratoria. August and September are the most proper time for our monitoring system because most of the adults appear in these two months in Northeast China. It is crucial to do so because the age of inspection is the major factor for us to determine when to use our system; moreover, a common application for density estimation is to determine whether a given population has increased or decreased significantly. Often density estimates between two time periods might be statistically significantly different; in our case, a single classification results cannot be used as a final conclusion. Comprehensive analysis was needed, that is to say, when monitoring a large area of the field, and the number of acoustic sensors and the time to record need to be considered.

Conclusions

Our results showed that the intraspecies acoustic characters could be used to identify the L. migratoria migratoria density based on the BP neural network and could be potentially used in other acoustic agriculture pest. In this paper, we just provided a theoretical method for the monitoring system which is a green, effective, and economical solution to automatically monitor the locust plagues and could be a potential guide to the manufacture of commercial monitoring system which is used in the field.

Declarations

Acknowledgements

This work is supported by the Special Fund for Agro-scientific Research on Public Interest (No. 200903021), Natural Science Foundation of China (No. 31172133), and the Fundamental Research Funds for the Central Universities (No. 11SSXT153). The system had been authorized for national patent of China (No. ZL201010518492.5). We are extremely grateful to John Cryan for the language correction (University College Cork, Cork, Ireland) and the members of our lab for the material collection. The materials used in this work are supported by the Central Lab, School of Life Sciences, Northeast Normal University, Changchun, China.

Authors’ Affiliations

(1)
Jilin Key Laboratory of Animal Resource Conservation and Utilization, School of Life Sciences, Northeast Normal University
(2)
School of Physics, Northeast Normal University

References

  1. Bush SL, Beckers OM, Schul J: A complex mechanism of call recognition in the katydid Neoconocephalus affinis (Orthoptera: Tettigoniidae). J Exp Biol 2009, 212: 648–655. 10.1242/jeb.024786View ArticlePubMedGoogle Scholar
  2. Ceccato P, Bell MA, Blumenthal MB, Connor SJ, Dinku T, Grover-Kopec EK, Ropelewski CF, Thomson MC: Use of remote sensing for monitoring climate variability for integrated early warning systems: applications for human diseases and desert locust management. Paper presented at the geoscience and remote sensing symposium. Columbia University, NewYork; 2006.Google Scholar
  3. Drake VA, Wang HK, Harman IT: Insect monitoring radar: remote and network operation. Comput Electron Agr 2002, 35: 77–94. 10.1016/S0168-1699(02)00024-8View ArticleGoogle Scholar
  4. Ewing AW: Arthropod bioacoustic. In Neurobiology behaviour. Comstock Publishing Associates, New York; 1989.Google Scholar
  5. Fischer FP, Schulz U, Schubert H, Knapp P, Schmöger M: Quantitative assessment of grassland quality: acoustic determination of population sizes of orthopteran indicator species. Eco Soc America 1997, 7: 909–920.Google Scholar
  6. García MD, Lorier E, Clemente E, Presa JJ: Sound production in Parapellopedon instabilis (Rehn, 1906) (Orthoptera: Gomphocerinae). Annales de la Socie′te entomologique de France (NS) 2003, 39: 335–342. 10.1080/00379271.2003.10697391View ArticleGoogle Scholar
  7. Greenfield MD: Evolution of acoustic communication in the genus Neoconocephalus : discontinuous songs, synchrony, and heterospecific interactions. In The Tettigoniidae: biology, systematics and evolution. Edited by: Bailey WJ, Rentz DCF. Springer, Heidelberg; 1990:71–97. 10.1007/978-3-662-02592-5_5View ArticleGoogle Scholar
  8. Heady SE, Denno RF: Reproductive isolation in Prokelisia planthoppers (Homoptera: Delphacidae): Acoustic differentiation and hybridization failure. J Insect Conserv 1991, 4: 367–390.Google Scholar
  9. Hemp C, Kehl S: Taxonomic changes and new species of the flightless genus Parepistaurus Karsch, 1896 (Orthoptera: Acrididae, Coptacridinae) from mountainous East Africa. J Orthop Res 2010, 19: 31–39. 10.1665/034.019.0106View ArticleGoogle Scholar
  10. López H, García MD, Clemente E, Presa JJ, Oromí P: Sound production mechanism in pamphagid grasshoppers (Orthoptera). J Zool 2007, 275: 1–8. 10.1111/j.1469-7998.2007.00394.xView ArticleGoogle Scholar
  11. Ma J, Han X, Hasibagan A, Wang C, Zhang Y, Tang J, Xie Z, Deveson T: Monitoring East Asian migratory locust plagues using remote sensing data and field investigations. Int J Remote Sens 2005, 26: 629–634. 10.1080/01431160310001595019View ArticleGoogle Scholar
  12. Montealegre-Z F, Morris GK: The spiny devil katydids, Panacanthus Walker (Orthoptera: Tettigoniidae): an evolutionary study of acoustic behaviour and morphological traits. Syst Entomol 2004, 29: 21–57. 10.1111/j.1365-3113.2004.00223.xView ArticleGoogle Scholar
  13. Ngo BV, Ngo CD: Reproductive activity and advertisement calls of the Asian common toad Duttaphrynus melanostictus (Amphibia, Anura, Bufonidae) from Bach Ma National Park, Vietnam. Zoological Studies 2013, 52: 12. 10.1186/1810-522X-52-12View ArticleGoogle Scholar
  14. Pace F, Benard F, Glotin H, Adam O, White P: Subunit definition and analysis for humpback whale call classification. Appl Acoust 2010, 71: 1107–1112. 10.1016/j.apacoust.2010.05.016View ArticleGoogle Scholar
  15. Paillette M, Bizat N, Joly D: Differentiation of dialects and courtship strategies in allopatric populations of Drosophila teissieri . J Insect Physiol 1997, 43: 809–814. 10.1016/S0022-1910(97)00030-9View ArticlePubMedGoogle Scholar
  16. Parkman JP, Frank JH, Walker TJ, Schuster DJ: Classical biological control of Scapteriscus spp. (Orthoptera: Gryllotalpidae) in Florida. Environ Entomol 1996, 25: 1415–1420.View ArticleGoogle Scholar
  17. Ragge DR, Reynolds WJ: The songs of the grasshoppers and crickets of Western Europe. Harley Books, Colchester; 1998.Google Scholar
  18. Riede K: Acoustic monitoring of Orthoptera and its potential for conservation. J Insect Conserv 1998, 2: 217–223. 10.1023/A:1009695813606View ArticleGoogle Scholar
  19. Schmidt GH, Stelzer R: Characterization of male structures, and the stridulatory organs of Pantecphylus cerambycinus (Ensifera: Tettigonioidea: Pseudophyllidae). Entomol Gen 2005, 27: 143–154.Google Scholar
  20. Schul J: Song recognition by temporal cues in a group of closely related bushcricket species (genus Tettigonia ). J Comp Physiol A 1998, 183: 401–410. 10.1007/s003590050266View ArticleGoogle Scholar
  21. Shi RX, Liu C, Li D-M, Xie B-Y: Application of MODIS_NDVI to monitoring locust pest in Baiyang Shallow Lake. J Nutr Dis 2003, 12: 155–160.Google Scholar
  22. Uvarov B: Grasshoppers and locusts: a handbook of general acridology, vol 1. Cambridge University Press, London; 1996.Google Scholar
  23. von Helversen D: Gesang des MSnnchens und Lautschema des Weibchens bei der Feldheuschrecke Chorthippus biguttulus L. J Comp Physiol A 1972, 81: 381–422. 10.1007/BF00697757View ArticleGoogle Scholar
  24. Wang Y, Zhang J, Li X, Ren B: Acoustic and molecular differentiation between macropters and brachypters of Eobiana engelhardti engelhardti (Orthoptera: Tettigonioidea). Zoological Studies 2011,50(5):636–644.Google Scholar
  25. Zhang DX, Yan L-N, Ji Y-J, Hewitt GM, Huang Z-S: Unexpected relationships of substructured populations in Chinese Locusta migratoria . BMC Evol Biol 2009, 9: 144. 10.1186/1471-2148-9-144PubMed CentralView ArticlePubMedGoogle Scholar

Copyright

© Wang et al.; licensee Springer. 2014

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited.