Detection and classification of flying insects by using radar doppler signatures of wing movement

Abstract

Disclosed is a method for detecting and classifying a flying insect. The method includes providing a radar sensor that is configured to transmit a radio wave of a predefined carrier frequency onto the flying insect, and recording a signal reflected from the flying insect, recording harmonic spectrum related to insect's wingbeat in the reflected signal, extracting a plurality of harmonic, spectral and temporal features from the harmonic spectrum resulting from the flapping of wings of the insect, selecting a plurality of relevant features from the extracted features, and analysing the extracted features for determining characteristics of the insect that allow its classification, by using signal processing techniques and machine learning model that is trained on features extracted from radar signatures of one or more flying insect species.

Description

[0001] Title

[0002] Detection and classification of flying insects by using Doppler signatures of wing movement

[0003] Field

[0004] The present disclosure relates to detection and classification of flying insects, and more particularly to classification of flying insects by using Doppler signatures of wing movement, as obtained from any radar system.

[0005] Biodiversity refers to the variety of living beings found within the global ecosystem. The existence of insects plays a vital role in human activities, ensuring the stability of our ecosystem and global food production, as well as their own well-being. Important pollinator species, such as the bee, for instance, are facing unprecedented decline due to factors like climate change, pesticide overuse, and habitat loss. This decline raises significant concerns due to its potential impact on food security, underscoring the urgent need for effective monitoring and conservation strategies. By embracing real-time monitoring technologies, it is possible to transition from re-active to proactive measures, providing early warning systems of potential threats. The traditional approaches to insect monitoring, such as human visual surveillance, insect trapping, and manual identification by specialists, are known to suffer from multiple inherent limitations. For example, these methods often require sacrificing the insects themselves and involve labour-intensive data collection. Additionally, many insects remain unseen and uncounted due to their minute size, natural camouflage, and limitations posed by human observers, which create blind spots and further compromise the accuracy of traditional approaches.

[0006] More recently, the advent of new solutions that allow for the automatic identification of insects aims to combat many of the limitations posed by traditional methods by employing sophisticated sensors and algorithms. Among these approaches, audio and vibration monitoring have become some of the most popular since they produce distinct vibroacoustic signals in response to various triggers. In most of the associated literature, acoustic sensors have been used due to their low-cost, portability, and non- invasive nature. These sensors can cover a wide range of spatio temporal scales However, they also come with notable limitations, including challenges posed due to ambient background noise which can severely degrade the accuracy of the solution, changes in the proximity of the monitored insects relative to the location of the microphone, and an overlap of extraneous sounds produced by other nearby animal species within the vicinity of the microphone, which has previously been a main cause for their restricted use by researchers.

[0007] Also, the state-of-the-art still overly rely on human factor, i.e. highly trained ecologists manually analysing the data. However, the vast amounts of data to be collected to achieve sufficient spatial and temporal resolutions quickly saturates human cognitive capabilities. Most works propose the machine learning classification of pollinators based on visual image processing. However, the classification of insects based on visual image processing is subject to the image quality, light conditions, weather and even the background of the image.

[0008] Some recent research has delved into radar cross-section (RCS) processing techniques that can provide more efficient detection, alongside prior findings of correlation between insect body mass and RCS. However, challenges associated with the overall size and expensiveness of complex systems based on pulsed radar or frequency-modulated continuous wave (FMCW) hinder their widespread implementation. There has also been progress in the direct identification of aerial targets, such as windborne migration insects or large insects at night. Those identifications are carried out by planes, kites, or searchlight trappings. Some works focused on extracting morphological parameters from the reflected signal, such as wingbeat, body length, width, and mass. Based on these parameters, they attempted to identify the targets. Even though manually extracted mass, wingbeat, body length width ration using stroboscope and steel rule shown possibility of species classification with 80% accuracy, the precision of retrieving these morphological parameters from even from the laboratory radar measurements from the insect is not yet good for species level classification.

[0009] WO 2024 / 050617 discloses insect detection and classification using a "centimeterwave radar" for laboratory data collection. Said radar primarily relies on spectrogram analysis without pre-selecting or filtering segments based on biological periodicity. W02024050617A1 describes the use of basic statistical features that provide only coarse statistical representations of the signal. CN 115565041 discloses a deep learning-based solution for "wing flapping pattern recognition on only three coarse “wing flapping patterns” (continuous, beat, irregular) and does not perform specieslevel or genus-level classification of insects. The document titled as “Entomological Radar Overview: System and Signal Processing” serves as general background art, confirming the known field of entomological radar. US 2024 / 0103130 A1 discloses a generic radar target classification system designed to identify classes of objects based on radar reflections, with no specific focus on biological targets, such as insects. The document titled “On pollinator monitoring using mmWave micro-Doppler Radar” discuss the extraction of important characteristic wingbeat information and noting that wing fluttering generates high-frequency harmonics.

[0010] SUMMARY

[0011] In an aspect of the present invention, as set out in the appended claims, there is provided a method for detecting and classifying a flying insect. The method includes providing a radar sensor that is configured to transmit a radio wave of a predefined carrier frequency onto the flying insect, and recording a signal reflected from the flying insect, recording harmonic content related to insect’s wingbeat in the reflected signal, extracting a plurality of harmonic, spectral and temporal features from the harmonic content resulting from the flapping of wings of the insect, selecting a plurality of relevant features from the extracted features; and analysing the extracted features for determining characteristics of the insect that allow its classification, by using signal processing techniques and machine learning model that is trained on features extracted from the radar signatures of one or more flying insect species.

[0012] In an embodiment of the present invention, the carrier frequency is at least 20 GHz.

[0013] In an embodiment of the present invention, the radar sensor operates at mm-wave frequencies.

[0014] In an embodiment of the present invention, the extracting the plurality of harmonic, spectral and temporal features from the reflected signal, comprises applying amplitude normalization to the recorded signal to reduce sensitivity to signal amplitude variations; and high-pass filtering the signal with a cut-off frequency to remove slow moving background noise, wherein the filter frequency is based on reported wing-beat frequency of one or more flying insects.

[0015] In an embodiment of the present invention, the method further includes removal of signal portions without harmonic content by calculating the harmonic ratio from the peak value of the normalized autocorrelation function within a specified lag range. The frames that possess a harmonic ratio exceeding a predefined harmonic ratio threshold are selected as the segments that include a Doppler signal from the insect.

[0016] In an embodiment of the present invention, the harmonic features include fundamental wingbeat frequency, harmonic spacing, harmonic-to-noise ratio, and relative harmonic amplitudes, extracted using Short-time Fourier transform (STFT), wavelet transform, cepstral analysis, normalized correlation function, and pitch estimation filters. The harmonic features relate to periodic structure, and harmonicity associated with wingbeat motion.

[0017] In an embodiment of the present invention, the spectral features relate to statistical and descriptive properties of signal’s spectral content, and include at least one of: spectral centroid, spectral flatness, spectral bandwidth, spectral roll-off (e.g. 95 % energy threshold), spectral flux, band energy ratios, skewness, kurtosis, or Mel- frequency cepstral coefficients (MFCC), and wherein the spectral features are computed from frequency-domain representations, and characterize energy distribution and variation across frequencies.

[0018] In an embodiment of the present invention, the temporal features wingbeat interval regularity, envelope modulation of the Doppler signal, time between successive wingbeat peaks, zero-crossing rate, or total signal energy.

[0019] In an embodiment of the present invention, the supervised methods include at least one of: mutual information and recursive feature elimination, and unsupervised methods include at least of: variance thresholding, clustering-based methods, and SHAP values.

[0020] In an embodiment of the present invention, the method further includes modelling the received radar signal as a combination of components including: a wingbeat modulation signal swingbeat(t), expressed as a harmonic series: swingbeat(t)= n=1Anexp(2mnfwwt + cf>n) , a body movement component, and an additive noise component; applying signal separation techniques, including bandpass filtering or harmonic extraction, to isolate the wingbeat component; and estimating the wingbeat frequency fw.

[0021] In an embodiment of the present invention, the recorded signal reflected from the flying insect is a complex Doppler signal that contains a first signal containing reflections from the background and noise, a second signal containing doppler signal from the movement of the insect, and a third signal containing micro-doppler periodic reflection due to wing movement, wherein the third signal is a sum of harmonic signals with lowest frequency corresponding to the frequency of the movement of the insect wings.

[0022] In another aspect of the present invention, there is provided a system for detecting and classifying a flying insect. The system includes a radar transmitter to transmit a radio wave of a predefined carrier frequency towards the flying insect; a radar receiver to record a radar signal reflected from the flying insect; a memory to store the one or more instructions; a processor to execute the one or more instructions to extract signal segments associated with wingbeat motion by computing a harmonic ratio, defined as ratio of energy in harmonic portion of signal to the total signal energy, and selecting Doppler segments exceeding a predefined harmonic ratio threshold; extract a plurality of features from the extracted signal segments resulting from the flapping of wings of the insect, the features comprising at least one of: harmonic, spectral and temporal features; automatically learn features obtained through a feature extractor implemented via a deep learning model; optionally select a subset of features based on supervised or unsupervised feature selection techniques that evaluate each feature’s discriminative importance towards classification; and classify the insect in one or more classes, using a machine learning trained on selected features, ora deep learning model configured to jointly learn discriminative features and perform classification from raw or transformed signal inputs, wherein the classification yields a taxonomic-level classification of the flying insect, including classes selected from species, genus, or wingbeat behaviour.

[0023] In yet another aspect of the present invention, there is provided a non-transitory computer readable medium having stored thereon computer-executable instructions which, when executed by a processor, cause the processor to record a radar signal reflected from the flying insect; extract signal segments associated with wingbeat motion by computing a harmonic ratio, defined as ratio of energy in harmonic portion of signal to the total signal energy, and select Doppler segments exceeding a predefined harmonic ratio threshold; extract a plurality of features from the extracted signal segments resulting from the flapping of wings of the insect, the features comprising at least one of: harmonic, spectral and temporal features; and automatically learn features obtained through a feature extractor implemented via a deep learning model; optionally select a subset of features based on supervised or unsupervised feature selection techniques that evaluate each feature’s discriminative importance towards classification; and classify the insect in one or more classes, using a machine learning trained on selected features, or a deep learning model configured to jointly learn discriminative features and perform classification from raw or transformed signal inputs, wherein the classification yields a taxonomic-level classification of the flying insect, including classes selected from species, genus, or wingbeat behaviour.

[0024] In yet another aspect of the present disclosure, there is provided a method for detecting and classifying a flying insect. The method includes transmitting by a radar sensor, a radio wave of a predefined carrier frequency towards the flying insect, recording a radar signal reflected from the flying insect, extracting signal segments associated with wingbeat motion by computing a harmonic ratio, defined as ratio of energy in harmonic portion of signal to the total signal energy, and selecting Doppler segments exceeding a predefined harmonic ratio threshold, extracting a plurality of features from the extracted signal segments resulting from the flapping of wings of the insect, the features comprising at least one of: harmonic, spectral and temporal features; automatically learning features obtained through a feature extractor implemented via a deep learning model; optionally selecting a subset of features based on supervised or unsupervised feature selection techniques that evaluate each feature’s discriminative importance towards classification; and classifying the insect in one or more classes, using a machine learning trained on selected features, ora deep learning model configured to jointly learn discriminative features and perform classification from raw or transformed signal inputs, wherein the classification yields a taxonomic-level classification of the flying insect, including classes selected from species, genus, or wingbeat behaviour.

[0025] Various embodiments of the present invention provide a multi-domain approach that includes harmonic, spectral, temporal features specifically derived from wingbeat- induced signal segments. These features are tailored to detect and classify taxonomic- level variations in wingbeat patterns. Thus, the present invention facilitates analyzing back reflected Doppler signal spectrum, i.e. the fundamental frequency and its harmonics, which originate from the movement of the wing of an insect. This component is typically buried within much larger reflected signal originating from other movements and stationary objects within the antenna range. Furthermore, the use of the whole harmonic spectrum allows not only the extraction of the frequency of the wing beat, but also its species-specific characteristic. In the proposed invention, this is considered a crucial morphological parameter for target identification, and this study incorporates all relevant features that depict the Doppler signatures resulting from wing flapping. The signal is analyzed, looking for the Doppler frequency spectrum, i.e. the main Doppler frequency and its harmonics, that includes information characteristic for each insect species. From the harmonic spectrum, several discrete features of the signal are extracted that were found to be the most significant distinguishing factors between various species. Those features are analyzed with a trained Al / machine learning algorithm, which is trained on similar features extracted from various flying insect species. Thus, by analysing in this way the Doppler signatures from individual flying insects, the proposed system enables an autonomous insect classification technique for flying insects using any radar operating at mm-Wave frequencies, or alternatively a backscattered signal used for telecommunication at mm-Wave frequencies, i.e. of at least 20 GHz. This method holds promise for establishing a foundation for sustainable and cost-effective insect monitoring, enabling farmers to identify pollinators without requiring expertise in insect taxonomy.

[0026] This work focuses on the design and implementation of a low-cost, portable, and power-efficient radar system which leverages the benefits of the Doppler effect induced by the wingbeats of bees for their identification. The portable device eliminates the need for human intervention and facilitates in-field monitoring of pollinators, enabling timely warnings about any changes in their biodiversity and the resulting ecological impacts they face. By conducting various simulated and real-life experiments, the feasibility of the approach is demonstrated for providing low-cost and effective portable in-field monitoring of pollinators for biodiversity conservation purposes.

[0027] In an embodiment of the present invention, the entire signal of the wing-flapping duration is analyzed, and it is assumed that the Doppler signature depends on each species' wingbeat, size, permittivity, mass, and other factors. Therefore, all signal properties are used, mainly harmonicity-related parameters and energy spread, for classifying targets without solely relying on morphological parameter extraction. Further, the harmonic spectrum is analyzed which allows the distinction between wingbeat patterns of different species. Consequently, several features from the harmonic Doppler signal are extracted, which - in combination with Al techniques - allow to classify the species of the insect from which the signal originates.

[0028] BRIEF DESCRIPTION OF DRAWINGS

[0029] The invention will be more clearly understood from the following description of an embodiment thereof, given by way of example only, with reference to the accompanying drawings, in which: -

[0030] FIG.1 illustrates a system for detecting and classifying of flying animals, in accordance with an embodiment of the present invention;

[0031] FIG.2 illustrates prior-art standard demodulator used in the Doppler radar system;

[0032] FIG.3 illustrates signal processing of output signals generated by the radar sensor as a result of detecting the insects, in accordance with an embodiment of the present invention;

[0033] FIG.4 illustrates an example of micro-Doppler segment extraction;

[0034] FIG.5A shows a set of insects. FIG.5B shows performance metrics for the set of insect classes, and FIG.5C shows exemplary reflected micro-Doppler spectrograms, recorded using 30 GHz continuous wave signal of the radar using a trained machine learning model;

[0035] FIG.6 lists the species and the number of subjects that in the experimental data demonstrated wing flapping for a minimum of 2 seconds; and

[0036] FIG.7 is a flowchart illustrating a method for detecting and classifying of flying animals, in accordance with an embodiment of the present invention. DETAILED DESCRIPTION OF THE DRAWINGS

[0037] FIG.1 illustrates a system 100 for detecting and classifying of flying animals, in accordance with an embodiment of the present invention. The system includes a radar sensor 102 that may be used to detect the Doppler signatures produced by pollinator insects, such as bees. The sensor 102 includes a K-Band transceiver. Two different embodiments of the invention were realised, one using transceiver from Vector Network Alayzer device with various continuous-wave signals of 20, 25 and 30 GHz, the second embodiment involves a continuous-wave generated by off-the-shelf mm- Wave Doppler radar. The Millimetre-wave frequencies offer distinct advantages for micro-Doppler analysis of small targets like insects due to shorter wavelengths providing higher resolution and sensitivity to small movements, which is particularly beneficial for extracting detailed harmonic and micro-Doppler signatures for accurate classification.

[0038] The sensor 100 may operate as a single-frequency transceiver, and the signal may be modulated, at any frequency above 20 GHz. An optional amplification circuit 104 may be integrated into the system. Example of the amplification circuit 104 includes an amplification circuit of 53.44 dB gain using the Texas Instruments rail-to-rail low-noise TL974 operational amplifier integrated circuit (IC).

[0039] FIG.2 illustrates a conventional receiver of radar sensor 200 in detail, in accordance with an embodiment of the present invention. The radar sensor 200 is a Continuous wave (CW) radar that generates a continuous-wave unmodulated signal at 30 GHz, and transmits it via an antenna. The reflected signal is received by the same antenna and is down-converted by mixing with the initially generated signal, thus extracting Doppler frequency. This is done for two reasons, to allow collection of generalized data, with as little hardware-specific aspects as practicable for a radar system and keep the cost of the device as low as possible, to support large-scale deployment. It is to be noted that the radar sensor 200 does not allow to measure the exact locations of the target, however for the investigated use-case such a location is irrelevant, as long as it is in the relative proximity (i.e. several centimetres) of the antenna - a functionality ensured by using low power to limit range. However, the location of the target can be easily added by person trained in engineering via the use of pulsed or frequency modulate radar, without affecting the functionality described in this invention. For pollinator biodiversity, the abundance of species is of the utmost significance, therefore the sensor 200 is intended to detect and recognise to which species the insect belongs.

[0040] For the generic case, where an insect is flying in front of the antenna, the complex down-converted Doppler signal can be expressed as a combination of three components: where snoise(t) stands for noise, including multiplicative background noise from all non-insect reflective objects; sbodymovement (t) is the Doppler signal originating from the movements of the insect due to its flight pattern; and swingbeat( is the Doppler periodic reflection due to wing movement. Since the latter component is periodic, i.e. modulated by the repeated wing movement, it can be expressed as a sum of harmonic signals with the lowest frequency corresponding to the frequency fw at which the insect moves its wings: where Anand cp„ are respectively the effective amplitude and phase shift induced by the wing-beating at (n - l)t / lharmonic frequency. Different insect species exhibit different patterns of wing movements, which may be captured by the composition of the harmonic spectrum of swingbeat(t).

[0041] Thus, the received radar signal may be modelled as a combination of components including: a wingbeat modulation signal swingbeat(t), expressed as a harmonic series: swingbeat(t), a body movement component, and an additive noise component. The signal separation techniques, including bandpass filtering or harmonic extraction may be applied, to isolate the wingbeat component, and estimate the wingbeat frequency fw.

[0042] FIG.3 illustrates signal processing of output signals generated by the radar sensor 200 because of detecting the insects, in accordance with an embodiment of the present invention. The overall processing includes data normalisation 304, high pass filtering 306, micro doppler segmentation 308, and feature selection 314.

[0043] The data acquisition includes acquiring raw sensor data, using the conventional radar sensor 200. The raw sensor data is first pre-processed to remove slow-moving background noise and signal portions without Doppler patterns from wing flapping. The signal is normalized 304, and high-pass filtered 306, then divided into windowed frames. The removal of signal portions without Doppler patterns is automated by analyzing the harmonic content of the signal. The harmonic content of a signal segment due to insect wing flapping extraction is quantified by calculating the harmonic ratio from the peak value of the normalized autocorrelation function within a specified lag range. The frames that possess a harmonic ratio exceeding a predefined harmonic ratio threshold are selected as the segments that include a Doppler signal from the insect. The harmonic ratio approach enables isolation of biologically relevant motion patterns, reduction of irrelevant or noisy segments, and enhancement of downstream feature computation and classification accuracy. This biologically informed segmentation strategy is functionally and conceptually distinct from existing systems. Also, the segmentation focuses on isolating portions of the signal that exhibit periodicity and motion signatures uniquely characteristic of flapping insect wings. The targeted segmentation improves both biological relevance and computational efficiency.

[0044] An example of segment extraction of the recorded Doppler signal has been illustrated with reference to FIG.4, where the short-time Fourier transform (STFT) of the measurement signal is presented at the bottom, benchmarked against the same signal with a Doppler segment highlighted at the top. In the STFT spectrogram, the Doppler patterns appear as parallel horizontal bands from 6 to 10 seconds.

[0045] The micro-doppler segments are used to extract hand crafted features 310 and automated deep learned features 312.

[0046] The handcrafted features 310 include temporal, spectral, and harmonic features from the signal as well as the automated deep learning features extracted from the signal. These three distinct categories of extracted features are essential for robust and efficient classification across diverse insect species and facilitate a structured approach to feature engineering.

[0047] The harmonic features relate to periodic structure, and harmonicity associated with wingbeat motion, and wherein the harmonic features include at least one of: fundamental wingbeat frequency, harmonic spacing, harmonic-to-noise ratio, and relative harmonic amplitudes. The harmonic features are extracted using at least one of: Short-time Fourier transform (STFT), wavelet transform, cepstral analysis, normalized correlation function, and pitch estimation filters. The spectral features relate to statistical and descriptive properties of signal’s spectral content, and include at least one of: spectral centroid, spectral flatness, spectral bandwidth, spectral rolloff (e.g. 95 % energy threshold), spectral flux, band energy ratios, skewness, kurtosis, or Mel-frequency cepstral coefficients (MFCC). The spectral features are computed from frequency-domain representations and characterize energy distribution and variation across frequencies. The temporal features include at least one of: wingbeat interval regularity, envelope modulation of the Doppler signal, time between successive wingbeat peaks, zero-crossing rate, or total signal energy.

[0048] These features are inherently capable of capturing complex patterns and representations that might be missed by handcrafted features. The handcrafted temporal, spectral, and harmonic features 310 are meticulously combined with automated deep learning features 312. This extensive feature set ensures that a wide range of signal characteristics is captured, which is crucial for accurate classification. In the feature selection 314, a subset of features is optionally selected based on supervised or unsupervised feature selection techniques that evaluate each feature’s discriminative importance towards classification. The supervised methods include at least one of: mutual information and recursive feature elimination, and unsupervised methods include at least of: variance thresholding, clustering-based methods, and SHAP values. The contribution and importance of these features are determined by using feature selection methods such as SHAP and then the best features are selected. The features are selected specially linked to different aspects of the Doppler spectrum, which are linked to species behaviour characteristics. The specific combination of these features and the application of SHAP for feature selection in insect classification task has not been explored before. The selected features form a doppler dataset for the Al classification. The Al classification classifies the insect in one or more classes, using a machine learning trained on selected features, or a deep learning model configured to jointly learn discriminative features and perform classification from raw or transformed signal inputs, wherein the classification yields a taxonomic-level classification of the flying insect, including classes selected from species, genus, or wingbeat behaviour.

[0049] The Al classification includes a three-level hierarchical classification to categorize the signals taxonomically. The first-level model 416 is trained to predict the 'Family' of the insect, the second-level model 418 to predict the 'Genus,' and the third-level model 420 to predict the 'Species.' The classification architecture consists of 3 models that leverages the complementary capabilities of ensemble learning and probabilistic modelling, encapsulated in a structured process designed to enhance predictive accuracy and adaptability.

[0050] Referring to FIG.3, the Al classification includes model initialization and training. The idea of using Al for radar-based insect classification is novel. The non-obvious step is the realisation that different wing movement patterns (i.e. specific to different insect species) can generate different Doppler reflections including harmonics, and that those reflections can be generalised (by the means of feature extraction and Al) to identify the species. The algorithm begins with the training of three models, for example, two CatBoost classifiers and a Gaussian Naive Bayes classifier. For each model, 80% of the data is used as a training set, and 20% is used for testing. The separation of data for training and testing ensured that the same subject data do not appear in multiple sets, preventing any potential data leakage. The CatBoost models are selected for their resistance to overfitting, and their capacity for handling highdimensional datasets. Each classifier is trained on a separate subset of the training data, allowing for diverse learning paths and the capture of varied data patterns. The first model, i.e. catboost model is trained with data labelled as a wasp (Vespula vulgaris) or bees. The second model, i.e. catboost model is trained with data labelled honeybee (Apis mellifera) or bumble bees. Following this, a third model such as a Gaussian Naive Bayes classifier is trained on another distinct subset of the training data that is labelled with bumble bee species names (Bom bus lapidarius, Bom bus terrestris and Bombus muscorum). After the initial training phase, the CatBoost models undergo a feature importance analysis. This step involves evaluating the impact of each feature on the model’s predictive performance and identifying those that contribute most significantly to the outcome. The analysis leads to the selection of the top features for the CatBoost models, based on their importance scores. This process is rooted in the belief, that concentrating on highly predictive features can improve model effectiveness and efficiency.

[0051] With the key features identified, the models are retrained to focus exclusively on these attributes. This retraining phase is pivotal, as it refines the models’ focus on the most informative features, thereby enhancing their predictive accuracy. The Gaussian Naive Bayes model is similarly retrained, ensuring that all models are optimized for performance. The final stage involves the application of the trained models to the test dataset in a sequential manner. The predictions from the first model are initially generated, followed by the second and then third model. The model predicts the family of the insect between Apidae and Vespidae. If the first model equals 0, indicating the family Apidae, the algorithm then consults the prediction from the second model which predict the genus of insect. If the prediction from the second model indicates 0 means the prediction is “Apis”(Apis mellifera), otherwise it indicates the genus “bombus" and a further subdivision of the bombus to species, the prediction of third model is considered. The outcome at this stage can lead to multiple specific classifications (’B.lapidarius,’ ’B.muscorum,’ or ’B.teresstris’), depending on the value of model.

[0052] FIG.5A shows a set of insects. FIG.5B shows performance metrics for the set of insect classes, and FIG.5C shows exemplary reflected Doppler spectrograms, recorded using 30 GHz continuous wave signal of the radar sensor 200 using a trained machine learning model. While some spectrogram differences between the species can be observed visually, for other cases, a machine learning algorithm is trained to make the correct distinction. In the table shown in FIG.5B, mellifera and vulgaris have good precision and recall, indicating balanced performance for predictions with a strong F1- score of 0.94 and 0.85. The lapidaries exhibit a lower precision of 0.61 , indicating a notable rate of false positives. However, lapidaries achieve a perfect sensitivity score of 0.98, denoting the model’s ability to capture all instances of this class with a good F1 score of 0.75. The muscorum and terrestris have good precision and lower sensitivity with good F1 -score of 0.75 and 0.73 indicating that it is conservative in predicting these classes; it does not capture as many muscorum and terrestris instances as there are, but when it does predict, it is usually correct. The given Al model demonstrates an 83% overall accuracy in classifying data into these five insect classes.

[0053] FIG.7 is a flowchart illustrating a method 700 for detecting and classifying of flying animals, in accordance with an embodiment of the present invention.

[0054] At step 702, a radar sensor is provided that is configured to transmit a radio wave of a predefined carrier frequency onto the flying insect, and recording a signal reflected from the flying insect.

[0055] At step 704, harmonic spectrum related to insect’s wingbeat is recorded in the reflected signal. The harmonic spectrum is recorded by extracting signal segments associated with wingbeat motion by computing a harmonic ratio, defined as ratio of energy in harmonic portion of signal to the total signal energy, and selecting Doppler segments exceeding a predefined harmonic ratio threshold.

[0056] At step 706, a plurality of harmonic, spectral and temporal features is extracted from the harmonic spectrum resulting from the flapping of wings of the insect.

[0057] At step 708, a plurality of features is automatically learned through a feature extractor implemented via a deep learning model and then a subset of features is optionally selected based on supervised or unsupervised feature selection techniques that evaluate each feature’s discriminative importance towards classification.

[0058] At step 710, the insect is classified in one or more classes, using a machine learning trained on selected features, or a deep learning model configured to jointly learn discriminative features and perform classification from raw or transformed signal inputs, wherein the classification yields a taxonomic-level classification of the flying insect, including classes selected from species, genus, or wingbeat behaviour.

[0059] In the specification the terms "comprise, comprises, comprised and comprising" or any variation thereof and the terms include, includes, included and including" or any variation thereof are considered to be totally interchangeable and they should all be afforded the widest possible interpretation and vice versa. The invention is not limited to the embodiments hereinbefore described but may be varied in both construction and detail.

Claims

Claims1 . A method for detecting and classifying a flying insect, comprising: transmitting by a radar sensor, a radio wave of a predefined carrier frequency towards the flying insect; recording a radar signal reflected from the flying insect; extracting signal segments associated with wingbeat motion by computing a harmonic ratio, defined as ratio of energy in harmonic portion of signal to the total signal energy, and selecting Doppler segments exceeding a predefined harmonic ratio threshold; extracting a plurality of features from the extracted signal segments resulting from the flapping of wings of the insect, the features comprising at least one of: harmonic, spectral and temporal features; automatically learning features obtained through a feature extractor implemented via a deep learning model; optionally selecting a subset of features based on supervised or unsupervised feature selection techniques that evaluate each feature’s discriminative importance towards classification; and classifying the insect in one or more classes, using a machine learning trained on selected features, or a deep learning model configured to jointly learn discriminative features and perform classification from raw or transformed signal inputs, wherein the classification yields a taxonomic-level classification of the flying insect, including classes selected from species, genus, or wingbeat behaviour.

2. The method as claimed in claim 1 , wherein the carrier frequency is at least 20 GHz.

3. The method as claimed in claim 1 , wherein the radar sensor operates at mm- wave frequencies.

4. The method as claimed in any preceding claim, wherein the extracting the plurality of harmonic, spectral and temporal features from the reflected signal, comprises: applying amplitude normalization to the recorded signal to reduce sensitivity to signal amplitude variations; and high-pass filtering the signal with a cut-off frequency to remove slow moving background noise, wherein the filter frequency is based on reported wing-beat frequency of one or more flying insects.

5. The method as claimed in any preceding claim, wherein the harmonic features relate to periodic structure, and harmonicity associated with wingbeat motion, and wherein the harmonic features include at least one of: fundamental wingbeat frequency, harmonic spacing, harmonic-to-noise ratio, and relative harmonic amplitudes, extracted using at least one of: Short-time Fourier transform (STFT), wavelet transform, cepstral analysis, normalized correlation function, and pitch estimation filters.

6. The method as claimed in any preceding claim, wherein the spectral features relate to statistical and descriptive properties of signal’s spectral content, and include at least one of: spectral centroid, spectral flatness, spectral bandwidth, spectral roll-off (e.g. 95 % energy threshold), spectral flux, band energy ratios, skewness, kurtosis, or Mel-frequency cepstral coefficients (MFCC), and wherein the spectral features are computed from frequency-domain representations, and characterize energy distribution and variation across frequencies.

7. The method as claimed in any preceding claim, wherein the temporal features include at least one of: wingbeat interval regularity, envelope modulation of theDoppler signal, time between successive wingbeat peaks, zero-crossing rate, or total signal energy.

8. The method as claimed in any preceding claim, wherein the supervised methods include at least one of: mutual information and recursive feature elimination, and unsupervised methods include at least of: variance thresholding, clustering-based methods, and SHAP values.

9. The method as claimed in any preceding claim, further comprising: modelling the received radar signal as a combination of components including: a wingbeat modulation signal swingbeat(t), expressed as a harmonic series: swingbeat(t)= n=1Anexp(2mnfwwt -i- c|)n) , a body movement component, and an additive noise component; applying signal separation techniques, including bandpass filtering or harmonic extraction, to isolate the wingbeat component; and estimating the wingbeat frequency fw.

10. A system for detecting and classifying a flying insect, comprising: a radar transmitter to transmit a radio wave of a predefined carrier frequency towards the flying insect; a radar receiver to record a radar signal reflected from the flying insect; a memory to store the one or more instructions; and a processor to execute the one or more instructions to: extract signal segments associated with wingbeat motion by computing a harmonic ratio, defined as ratio of energy in harmonic portion of signal to the total signal energy, and selecting Doppler segments exceeding a predefined harmonic ratio threshold;21 extract a plurality of features from the extracted signal segments resulting from the flapping of wings of the insect, the features comprising at least one of: harmonic, spectral and temporal features; automatically learn features obtained through a feature extractor implemented via a deep learning model; optionally select a subset of features based on supervised or unsupervised feature selection techniques that evaluate each feature’s discriminative importance towards classification; and classify the insect in one or more classes, using a machine learning trained on selected features, or a deep learning model configured to jointly learn discriminative features and perform classification from raw or transformed signal inputs, wherein the classification yields a taxonomic-level classification of the flying insect, including classes selected from species, genus, or wingbeat behaviour.

11. A non-transitory computer readable medium having stored thereon computerexecutable instructions which, when executed by a processor, cause the processor to: record a radar signal reflected from the flying insect; extract signal segments associated with wingbeat motion by computing a harmonic ratio, defined as ratio of energy in harmonic portion of signal to the total signal energy, and select Doppler segments exceeding a predefined harmonic ratio threshold; extract a plurality of features from the extracted signal segments resulting from the flapping of wings of the insect, the features comprising at least one of: harmonic, spectral and temporal features; automatically learn features obtained through a feature extractor implemented via a deep learning model; optionally select a subset of features based on supervised or unsupervised feature selection techniques that evaluate each feature’s discriminative importance towards classification; and11 classify the insect in one or more classes, using a machine learning trained on selected features, or a deep learning model configured to jointly learn discriminative features and perform classification from raw or transformed signal inputs, wherein the classification yields a taxonomic-level classification of the flying insect, including classes selected from species, genus, or wingbeat behaviour.

Images