Over the past decade there has been a substantial increase in the number of micro-drones, i.e., Unmanned Aerial Vehicles (UAVs), available on both the commercial and consumer markets. This increase can be attributed to their continually improving capabilities, ease of use, and low cost, making them attractive to businesses and hobbyists alike. There have been countless positive applications of micro-drone technology, e.g., agricultural and environmental surveying, disaster response and search and rescue. However, micro-drones have become infamous for their misuse, in performing anti-social, and often criminal activities, such as illegal filming of private/sensitive locations or dangerous flying through restricted airspace creating a collision hazard to other aircraft. Of particular interest in this work is the ability for micro-drones to be utilized as vehicles for transporting payloads of dangerous substances e.g. explosives, chemical/biological weapons and illicit drugs. This capability is of concern to security agencies worldwide (Forbes, 2019), with calls being made to limit commercial drone payload capacities in order to limit the risk posed (BBC, 2016).

Current conventional radar based air surveillance systems often perform poorly when tasked with detecting and tracking micro-drones, due to their considerably smaller Radar Cross Section (RCS) when compared with standard aircraft, thus establishing the need for a dedicated system with adequate sensitivity for detection of these small RCS targets (Rahman and Robertson, 2018). Additionally, birds present an issue with drone detection systems due to their comparable RCS, altitude and speed of flight, all of which make discrimination between these two targets challenging using RCS information alone (Rahman and Robertson, 2018). In order to reduce the false alarm rate and thus mature radar based drone detection technology, there is a widening body of research into methods for discriminating between birds and micro-drones. The micro-Doppler signatures of drones have been identified as one method for discriminating micro-drones from other targets and general clutter (Tahmoush, 2014; Ritchie et al., 2016; Samiur and Robertson, 2019; Beasley et al., 2020). Micro-Doppler signatures are produced by periodic micro-motions of a target or object, such as flapping of a bird’s wings, the gait of a human walk or rotation of a micro-drone’s propellers (Chen, 2019). Micro-Doppler is the respective Doppler shift induced by these target micro-motions which appear as side bands around the target’s bulk Doppler. Further details on the mechanisms that produce micro-Doppler have been extensively documented for a range of radar targets in Miceli (2014) and Chen (2019). There are several publications detailing micro-Doppler signatures of drones, where comparisons are made with bird micro-Doppler e.g., Tahmoush (2014), Ritchie et al. (2016), Rahman and Robertson (2018), and Beasley et al. (2020). This work has been taken further in Molchanov et al. (2014), Kim et al. (2017), Samiur and Robertson (2019), and Rahman and Robertson (2020), where drone micro-Doppler signatures are fed into classifiers in order to discriminate between both different classes of drones and between drones and birds. For a more comprehensive review of the current literature on drone detection and classification using machine learning, please refer to a recent paper published on the topic by Taha and Shoufan (2019).

There have been a number of works published on multi-frequency micro-Doppler based classification of targets, with gesture recognition accounting for a large portion. Such works look to compare the suitability of particular frequency bands in the role of micro-Doppler signature generation for classification. A recent paper by Gurbuz et al. (2021), investigated the use of three frequencies S-bands, namely 77 GHz, 24 and 10 GHz, for use in sign language recognition. It was reported that, in nearly all cases, using the 77 GHz radar data resulted in the highest classification accuracy, where the lowest classification accuracy was reported to result from using the 10 GHz radar data.

A comparatively more limited body of literature is published on the topic of micro-Doppler based classification of unloaded and loaded drones. A single data-set collected by University College London’s (UCL) multistatic S-Band radar system, NETRAD (Doughty, 2008), has been used in a number of publications (Fioranelli et al., 2015; Ritchie et al., 2017; Patel et al., 2019; Pallotta et al., 2020), to test the accuracy of a variety of payload classification algorithms. In this data-set, a DJI Phantom 2 was loaded with a variety of payloads between 0 and 600 g and made to either hover or fly towards the central radar node. Fioranelli et al. (2015) and Ritchie et al. (2017) report high classification accuracies using a feature extraction based approach. In the hovering scenario, accuracies of 96–97% were achieved using centroid based micro-Doppler features. In contrast, for payload classification of the drone in flight, a maximum classification accuracy of 95% was achieved when using singular value decomposition (SVD) derived features. A similar feature based approach was used on the same data-set in a more recent publication by Pallotta et al. (2020), differing mainly through the use of spectral kurtosis (Dwyer, 1983) to extract features from the spectrograms. The algorithm produced by Pallotta et al. (2020) achieved classification accuracies for the hovering drone of 92.61%, marginally lower than the result reported by Ritchie et al. (2017). In the case of the moving drone, their classifier achieved an accuracy of 70.83%, considerably lower than the 95% accuracy reported in the work by Ritchie et al. (2017). The UCL data-set is again used by Patel et al. (2019); however, a different approach is used to tackle the classification problem. In Fioranelli et al. (2015), Ritchie et al. (2017), and Pallotta et al. (2020) work, a single spectrogram time-frequency representation was used, whereas in the work by Patel et al. (2019), two additional transforms are used, namely the cepstrogram and the cadence velocity diagram (CVD). Instead of selecting individual features to feed a classifier, a Convolutional Neural Network (CNN), AlexNet, is provided with the three time-frequency representation images. In both the hovering and moving drone scenarios, the AlexNet based image classifier had an overall accuracy of around 95–96%, comparable to the results reported by Fioranelli et al. (2015) and Ritchie et al. (2017).

Rahman et al. (2020) from The University of St Andrews have since released a substantial piece of work in which a 24 GHz Frequency Modulated Continuous Wave (FMCW) radar and 96 GHz Continuous Wave (CW) radar were simultaneously used to sample the micro-Doppler signatures of two large drones (DJI S900 hexacopter and a Joyance JT5L-404 crop spraying quadcopter) with a variety of payloads up to a maximum of 5 kg. Rahman et al. (2020) concluded through visual inspection of the spectrograms derived from the FMCW radar data that there were no consistent unique signatures that could solely be attributed to a drone carrying a payload. However, more success was had in the case of the CW radar data, though a conclusion was still reached that observed features could not be used as the sole basis of a classification algorithm. Nevertheless, an attempt was made to classify between unloaded and loaded drones using two machine learning approaches, namely feature extraction and a CNN. However, both approaches were found to have limited success.

The work presented in this paper looks to expand on previous work by comparing the effectiveness of radar data captured at three different frequency bands to be used for determining the payload weight which different drones are carrying. Radar systems operating at C-band, S-band and W-band (mm-wave) are used to measure two different types of small quadcopter type drones carrying three different payload weights and performing two different movement styles each. A series of investigations with increasing complexity are then carried out using the micro-Doppler signatures obtained from the radar measurements to be able to classify the payload type. The rest of this paper is organized as follows. Section 2 provides an overview of the experimental setup and the equipment used, including information about the radar systems and drones. Section 3 presents the theory and the methodology of the signal processing and the classification procedures used. Section 4 reports the results obtained and provides some analysis from these. Discussion on the observations are then summarized in Section 5, before conclusion are finally presented in Section 6.

In this section, the two radar platforms used in this research to capture radar data, the C-band and S-band ARESTOR RF sensor system and the W-band Texas Instruments radar system, are described. Details of the drone models flown as targets and the experimental setup employed to gather the radar data are then provided.

In this section, the signal processing methodology used to facilitate the generation of Doppler-time spectrograms is presented. The extracted features and machine learning model used for classification are described, and details of the classification investigations for which results will be reported are given.

This section first presents examples of Doppler-time spectrograms obtained from captures using two different payload weights at each of the three frequency bands before presenting the results from the different classification tests in the order introduced in the preceding section.

A series of example Doppler-time spectrograms corresponding to 1 min of radar data are shown in Figure 7. The spectrograms show the differences between data collected at the three different frequency bands being used to generate a Doppler signature corresponding to the same capture. Spectrograms are also shown for two different payloads: No payload and heavy payload (500 g). All the example spectrograms are generated from data collected using the big drone and the “Moving” motion type.

The initial classification investigation aimed to classify the weight of the carried payload by the drone given that the training set data was formed only from samples corresponding to the same drone and motion type as test samples being used, along with samples from a static target. Figure 8 shows the results from this classification investigation for both drone and motion types at each of the three frequency bands.

It can be seen from Figure 8 that when the motion type is “Hovering,” higher payload classification accuracies were achieved using S-band and C-band when the small drone was used, whereas a higher classification accuracy was achieved using W-band when the big drone was used. When the big drone is used, the performance ranking of the three frequency bands, from best to worst is: W-band, C-band, S-band. When the small drone is used, the ranking is reversed: S-band, C-band, W-band. This shows that despite the smaller payload weight to body weight ratio, when the smaller drone is used, the payload classification problem is still easier for the small drone when the motion type is “Hovering” and either S-band or C-band are being used.

When the motion type is “Moving,” it is seen that the performance ranking of the three frequency bands, from best to worst is: W-band, S-band, C-band. This is the same ranking order for both drone types; however, higher classification accuracies are observed for the big drone compared to small across all frequency bands. The classification accuracies for samples from the big drone were higher than that from samples from the small drone by the following amounts: S-band = +28.90%, C-band = +20.60%, W-band = +6.50%, showing that the smallest performance spread due to different drone types was seen when using W-band. When the motion type is “Moving,” payload classification accuracy for both drones at all frequencies is lower than for the “Hovering” motion type, showing that classification of payload weight using signatures from the drone hovering is more accurate than classification using signatures from moving drones.

By taking into account the graphs for both motion types in Figure 8, it can be seen that W-band data leads to the best classification accuracy in three out of the four cases and S-band data has the best payload classification accuracy for the remaining case. C-band ranks second best when the motion type is “Hovering,” for both drone types. S-band ranks second best for both drones when the motion type is “Moving.”

Out of the four cases for which data is shown in Figure 8, the largest performance spread between any two frequency bands occurs in the case of small drone and a motion type of “Moving,” between W-band and C-band, where the classification accuracy difference is 47.65%. Across the four cases included in the first investigation, the best overall payload classification accuracy is achieved at W-band when using the big drone in a “Hovering” motion, while the worst overall payload classification accuracy is achieved at C-band when using the small drone in a “Moving” motion.

Figure 9 shows the results from the second classification investigation, in which training sets are constructed using samples from a specific drone type but for both moving and hovering motions, as well as static non-drone target samples.

From Figure 9 it can be seen that the ranking order of the frequency bands from best to worst payload classification accuracy is the same for both drone types: W-band, S-band, C-band. The payload classification accuracy is seen to be higher for the small drone compared to the big drone for S-band and C-band, though this difference is very small (<2%). The opposite is seen for results obtained using the W-band data. The classification accuracy spread across the three frequency bands for the big drone is 23.85%, while for the small drone this spread is 19.50%, meaning less variation in performance across frequency bands is seen when using the smaller drone compared to the bigger drone.

Figure 10 shows the results from the third classification investigation where payload classification accuracies are found using mixed drone data but known movements (the left hand set of bars correspond to the hovering movement and the central set of bars correspond to the moving drone movement), as well as for the fourth classification investigation where mixed drone data and mixed movement data is used (shown in the set of bars on the right side of the graph).

From the first two sets of bars (bars for “Hovering” and “Moving”) in Figure 10, it can be determined that when the drone type is unknown, the highest payload classification accuracy is achieved for motion type “Hovering” for all three frequency bands. The smallest variation in performance across S-bands also occurs for hovering data, where all three frequency bands performed within 4.35% of each other. While the performance differences are small for hovering data, the frequency band ranking from best to worst is seen to be: W-band, C-band, S-band. In the case where the data type is from the drones performing the “Moving” motion, the frequency band performance ranking is: W-band, S-band, C-band. The performance differences seen between the “Hovering” and “Moving” signatures for each band are: S-band = −17.55%, C-band = −38.25% and W-band = −6.90%, showing the smallest variation in performance across different motion types to be achieved by W-band.

When the signatures used in an investigation are of unknown motion and unknown drone (corresponding to the set of bars on the right hand side of Figure 10), the payload classification accuracy of the frequency bands from best to worst is: W-band, S-band, C-band. It is also seen that S-band and C-band perform better in this investigation than when moving only data is used, but worse than when hovering only data (i.e., the results from investigation three), while W-band shows a minor performance reduction compared to the case where the motion type is known to be “Moving” (<1%).

Figure 11 includes three confusion matrices pertaining to each of the three frequency bands. Each confusion matrix is constructed by recording the total number of classification attempt outcomes for each investigation carried out within the research presented. The values in the confusion matrices shown the proportional breakdown of outcome class predictions given a particular true payload class.

It can be seen from the confusion matrices that the highest proportion of correct classifications for each payload type occurred when using W-band data, followed by S-band data, and the worst performance is seen from C-band data. When using C-band data, the highest misclassification situation occurs when the correct payload is “Heavy” but the “Empty” class is predicted.

The second highest misclassification situation when using C-band data occurs jointly when the true payload is “Light” but the “Empty” class is predicted, and when the true payload is “Empty” but the “Heavy” class is predicted. Three misclassification scenarios occur with the same rate as these two variations when using S-band. For all three frequency bands, no instances were observed where a situation was classed as having no drone present. For all three frequency bands, the most likely true payload to be misclassified is the “Heavy” payload. Similarly, for all three frequency bands, the most likely true payload to be correctly classified is the ‘Empty’ payload, though the true class of “Light” payload has an equal misclassification rate as this in the case of S-band.

All classification tests reported in this section of were repeated using a set of eight conventional SVD features, previously used for micro-Doppler classification in Fioranelli et al. (2017) and Dhulashia et al. (2021). The four investigations reported include a total of 27 different reported classification accuracy results. A comparison of the classification accuracy achieved in each of the tests using the SVD features with that achieved using the features proposed in this work can be seen in Figure 12.

As shown in Figure 12, the feature set reported in this paper outperformed the SVD feature set in 25 out of the 27 classification tests, with classification accuracies being an average of 26.67% higher when using the feature set proposed in this paper.

The results shown within this paper provide evidence that it is possible to classify the payload of different quadcopter micro-drone platforms. The tests included used payloads varying from 0% up to 40% of the drone body weight and achieved a peak classification success of 98.55%. From the results shown in Figures 8, 10, it can be seen that higher payload classification accuracies were achieved when using signatures from the “Hovering” scenarios, compared to those from when the drone is performing the “Moving” motion. This may be due to a greater number of degrees of freedom in the drone movement for the “Moving” motion, causing greater variations within the signatures for this motion type. Examples of these include more deviations in: Flight path, drone altitude and drone tilt.

It can be determined that the best payload classification accuracy can be achieved using signatures obtained at W-band frequencies. However, the W-band radar system is significantly range limited compared to the C-band and S-band systems, and suffers from rapid SNR reduction as the range to the target increases. For the W-band data, the lack of observable Helicopter Rotation Modulation (HERM) lines within Doppler signatures means that the features extracted were dominated by the bulk Doppler contribution which appears to be sufficiently different in each payload example to allow high classification success rates despite the lack of HERM lines. This may be due to the changed reflectivity of the target from the added metallic weights creating the payload. As such, this conclusion is limited to the short range examples tested in this research and the W-band system is likely to have a significant drop off in classification success rate with increases in range, unless highly directive antennas coupled with significant power amplification is built into the radar system.

In the ideal case, a single system capable of simultaneous measurements of all three frequency bands would be used along with transmission and reception hardware to ensure no difference in gain values. Such a system would allow for a comparison of the classifiability of micro-Doppler signatures for the three frequency bands while ensuring no other operating variables have any bearing on the results. The measured SNR of the S-band and C-band signals collected in this work were seen to be similar, thus legitimizing a comparison of data obtained using different hardware architectures.

In comparing the C-band and S-band results, the S-band was found to generally outperform the C-band measurements, particularly for the “Moving” and combined movement cases. Therefore, the recommendation based on the findings of this work would be to utilize the S-band frequencies for the task of micro-drone payload classification, due to this frequency band’s long range capabilities and high classification success rates. While the superior classification accuracy of the W band is likely due to the greater magnitude of Doppler frequencies spanned by the bulk Doppler signature, the S-band and C-band signature classification results from a greater contribution of the HERM lines observed within the spectrograms. The bulk Doppler contribution will be lower for these bands since this component of the signature spans only a small proportion of the overall Doppler frequency span. The S-band is shown to result in a higher overall classification accuracy than the C-band. Two factors are thought to contribute to this. The first of these results from the differences in the distribution of the HERM lines within the signatures from the two bands. In the S-band spectrograms, it can be observed that the HERM lines are unambiguously captured, while in the C-band spectrograms HERM lines may extend beyond the Doppler frequency range and result in overlap. Varying the payload on the drone results in changes in the rate at which the motor rotates which induces changes in the distribution of the HERM lines. For a higher payload and a higher rate of motor rotation this would appear only as the spreading of HERM line spacing in the S-band spectrograms. Conversely, this change would result in some increase in HERM line spacing and others becoming closer in the C-band spectrograms, due to the overlap wrapping effect. The S-band and C-band performance differences are also partially likely due to the higher SNR of the HERM lines within the S-band spectrograms compared to the C-band spectrograms. Since the features used for classification are based on measuring the distribution of energy within the signature, a more sensitive distinction between signatures of different payload classes is achievable for higher variations between the energy distributions in the spectrograms from different payload classes. The stronger HERM lines observed in the S-band spectrograms would result in higher variations between the energy distributions within segments from different payload classes, thereby resulting in a higher overall classification accuracy compared to the use of C-band spectrograms, which exhibit less energy distribution variations across segments from different payloads due to the lower SNR HERM line contributions. The disadvantage of S-band is that a system operating at this frequency band requires a larger antenna due to the longer wavelengths involved. This may present additional logistical challenges if deployment is required on a lightweight or small platform. Another potential reason for the higher classification success rates achieved when using S-band could be that the features being extracted have a greater propensity towards characteristics more readily observed within the S-band spectrograms compared to the C-band spectrograms.

This paper has presented results of using micro-Doppler features extracted from radar data obtained at three frequency bands to classify between three different payloads carried by two types of quadcopter micro-drones performing two different motions. It has been shown that the features and classification methodology used in this work can successfully determine the payload being carried by a drone in problems chosen to have varying complexities and situational knowledge. The reported results show that W-band is the preferable choice of radar frequency band to use for micro-drone payload classification. Out of the three frequency bands investigated, S-band is the second most preferable and C-band is the least preferable. It has also been shown that micro-Doppler signatures collected from hovering drones lead to more accurate payload classification than signatures from moving drones.

Further work will aim at collecting additional data in different conditions to validate these preliminary results, including the use of additional models of micro-drones and selecting different locations in which to perform tests.

The results shown within this paper are all based on data captured using horizontal polarization (H-pol) geometries. Future work may look to investigate the benefit of a multiple polarized sensing system when looking to classify a drone payload. H-pol was selected to maximize the micro-Doppler signatures but vertical polarization (V-pol) may also give information on flight dynamics, particularly when the drone is moving at higher velocities and travelling at a significantly tilted orientation.

In all four of the classification investigations presented in this work, classification decisions were made solely on the data from a selected frequency band at any given decision instance. If a multi-frequency system were to be used with a voting system to make classification decisions, it is likely that the payload classification accuracy would be greatly increased. This could be achieved by using the data collected from all three frequency bands and applying separate classifiers, before voting on the outcomes to obtain a joint decision. It is expected that this fused decision would provide an improved overall success rate.