Ground penetrating radar used to detect drowning victims under ice

Abstract

 

Every year, people drown after falling through ice on rivers and lakes. In some cases, the body of the victim floats up to the underside of the ice, making detection and recovery difficult using traditional search methods with divers. A robust and contact-less sensing system is required to locate drowning victims that does not put rescue teams at risk of falling through the ice themselves. In this paper, we demonstrate the feasibility of a ground penetrating radar (GPR) for detecting deceased drowning victims that have floated up to the underside of the ice. We placed three euthanized pigs simulating drowning victims under ice ranging in thickness from 5 to 26 cm. We dragged a GPR at 500 MHz and 1 GHz across the ice to detect the simulated victims using an autocorrelation-based detection technique. Results showed that both frequencies were able to detect the rough shape of the simulated victims at ice thicknesses up to 42 cm, with the 1-GHz data showing slightly more resolution than the 500-MHz data. These results show promise and suggest future development of an autonomous drone-based GPR detection system.

Key points

• Floating bodies are successfully detected under both ice and snow using a commercial ground penetrating radar system with ice depths reaching up to 26 cm in a controlled environment.

• The differences between using radar systems operating at/around 500 MHz and 1 GHz were not pronounced from the point of view of detection.

• Future studies should investigate the capabilities for detecting bodies in more realistic settings.

Introduction

The past 20 years have seen the development, and increased use, of geophysical tools to search for “clandestine graves and physical evidence associated with criminal activity” [1]. Ground penetrating radar (GPR) has been the subject of numerous controlled studies simulating forensic settings. For example, pig carcasses have been used to determine effects of different burial times and types of soil on GPR scans [1, 2]. GPR has also been used in snow, ice and glacier environments for snow management and to detect ice thickness and volumes and even buried aircraft [3, 4]. However, to our knowledge, GPR has not been used to detect objects or human bodies under, or encased in, ice that covers bodies of water [5].

An urban winter drowning case has highlighted the potential usefulness of GPR for the novel purpose of detecting a drowning victim who was lost under the ice [6]. A 6-year-old boy went missing under the ice in Winnipeg, Canada, in early December 2010. The Winnipeg Police Service underwater search and recovery unit immediately started a search of the riverbed in zero visibility water. Volunteers from the Canadian Amphibious Search Team were asked to join the search, and each team took responsibility for parallel sections of the river using coordinated search patterns of the riverbed. The search was also augmented by side scan sonar technology (Figure 1). After one of the most extensive underwater searches in Canadian history, the search was discontinued, and the body was found floating 10 months later.

Figure 1.

Side scan sonar of potential drowning victim (at 10:30 position). The object was later determined to be a large piece of waterlogged wood. A white circle and ”Waterlogged Wood” text were added on the image to highlight the area of interest. ([6]; with permission)

An alternate hypothesis was developed that the victim may have been floating up under the ice. Normally, drowning victims sink to the bottom close to the point of submersion (unless there is a strong current) [7]. However, bodies usually refloat after some time, and may not sink at all in rare instances. Primary flotation results from bacteria-generated gas (putrification) within the digestive tract [8]. This increases the body’s buoyancy, and the body may float in as little as 12 h in warm water. Secondary flotation results from putrification within the circulatory system, muscles and viscera and takes longer to occur. Although these processes are slower in cold water, drowning victims have been found up under the ice after 64 days (Manuel Maendel, personal communication, 18 April 2022) (Figure 2).

Figure 2.

Drowning victim found floating under the ice, 64 days after going missing. The body had sunk and refloated after the lake had completely frozen over. A white chalk outline was added on the image to highlight the shape and parts of the body. ([6]; with permission)

In rare cases, drowning victims do not sink at all and are found floating on the surface within minutes to hours of drowning. For example, a review of body recoveries by several recovery divers from the Royal Canadian Mounted Police revealed that, of 225 bodies recovered, eight were found floating (seven face down, one undetermined) shortly after drowning in open water, and therefore not because of putrification (Jay White, personal communication, 23 March 2022). There have also been cases of drowning victims being found up under the ice, close to the point of entry ([9] and cases from Jamie Diemert, personal communication, 2 December 2022).

One explanation for these latter cases is that water in the larynx triggers laryngospasm resulting in asphyxia with little or no water entering the lungs. If the victim is face down the body may float at the surface even after laryngospasm relaxes after death since air may be trapped in the lungs much like in an inverted pop bottle prevents water from entering when it is submerged. Additionally, snowmobilers may not sink after drowning as snowmobile helmets have significant flotation (some have as much buoyancy as an approved personal flotation device). Finally, some air may be trapped in clothing thus increasing buoyancy.

If any of these scenarios were to occur in ice-covered water, the body would be floating up against the bottom of the ice, undetectable by search methods aimed at the riverbed. Based on this possibility, an extensive GPR search was performed 3 months after the 6-year-old drowned, and holes were augured at several sites indicated by anomalies in the GPR data. The body was not located during this process but was found a few miles downstream 10 months later floating due to the purification process [8].

Even though the GPR search was unsuccessful, this strategy should still be considered in searches where a body cannot be found by an extensive bottom search. Early location of the body would be advantageous for several reasons including forensic analysis before evidence is distorted by the purification process, and for timely return of the remains to the victim’s family.

Motivated by these advantages, we conducted a study to assess the ability to detect bodies below or, encased in, the ice surface with GPR. The following questions were addressed:

• a) Can bodies be detected under thin (5 cm) and thicker (26 cm) ice, as well as once they are completely encased by ice?

• b) What are the advantages and disadvantages, if any, of two different antenna frequencies (500 MHz and 1 GHz) available in a commercial GPR system?

• c) What is the effect of clothing and air in the lungs on the ability to detect carcasses?

Our goal in addressing these questions is to determine if GPR would be a valuable addition to the under-ice search and recovery tool kit.

Methods

Research site and ice burial simulation

The research site was located at the Sea-ice Experimental Research Facility (SERF) at the University of Manitoba in Winnipeg, Canada. The study was conducted from December 2021 to January 2022 during typically cold winter months. Trials were conducted in a standalone tank (Figure 3) filled with fresh water. Three euthanized pig carcasses (20–25 kg) were used as human analogues (animal carcasses were obtained from Maple Leaf Agra Farms, Landmark, Manitoba); this size was chosen to represent children who are more likely to drown under ice, and to provide the smallest realistic target for GPR detection.

Figure 3.

Submersion tank (). Nude carcass (open) and clothed carcass (shaded). Solid border: submersed under 5-cm-thick ice. Dashed border: submersed later under 26-cm-thick ice.

Initially when the ice was 5-cm-thick, a hole was cut into the ice and two carcasses were placed in parallel, 0.5 m apart (Figure 3); one of the carcasses was nude and the other clothed in a leather jacket (Figure 4). Prior to submersion, each carcass was intubated, and the lungs were inflated to mimic a scenario in which asphyxiation occurred with inhalation of minimal or no water. Each carcass was placed in an upright position (dorsal aspect up against the ice) and held in position against the bottom of the ice with a thin string with the snout facing west. The ice block that was cut out of the ice to facilitate submersion was then replaced to fill the hole. After 12 days, when the ice was 26-cm-thick, a third carcass was clothed and placed parallel with the other carcasses (Figure 3). Since the initial analysis of the first two carcasses indicated that detection of the carcasses was not dependent on the air volumes in the lungs, and it is possible for a victim to float even after inhaling water, the third carcass was not intubated, and the lungs were not inflated. Rather, the third carcass was held in the upright position against the bottom of the ice with thin string attached to the neck and hind quarters (seen on the far right of Figure 5).

Figure 4.

Two pig carcasses (one with clothing and one nude) under 5 cm of ice. Note the string was used to position the heads in line with each other.

Figure 5.

Ground penetrating radar (GPR) scan process; passes were always in the longitudinal (north-south).

GPR

The GPR system used in this study was the Spidar Network GPR system with Pulse EKKO PRO 500-MHz (15-cm-wide) and 1-GHz (7.5-cm-wide) antennas (Sensors & Software Inc., Radiodetection, Mississauga, Canada) (Figure 6). The available antennas enable GPR measurements of the subsurface using different frequencies. Attenuation through an electromagnetically lossy medium (such as ice) is linearly proportional to the frequency (i.e. higher frequencies will not penetrate as far and received signals will be more prone to artifacts and/or positional errors than lower frequencies). The trade-off is that lower frequencies provide lower target resolution information [10]. Although the limited depth penetration would be problematic in some ground burials that could be quite deep, we did not anticipate it to be a problem for under ice searches for drowning victims since they have likely broken through thin ice. We conducted experiments using both frequencies to demonstrate robustness to frequency selection for this application.

Figure 6.

Ground penetrating radar (GPR) units including transmitting and receiving antennas (left pair, 500 MHz; right pair, 1 GHz), data collection and transmission module, and battery (black case).

We placed the GPR system into a small sled so that we could manually pull it over the ice (Figure 5). During each pass, data were collected and stored within the GPR system itself and transmitted to a computer for real-time visualization. Data were later downloaded to a computer for subsequent analysis.

Data collection

To collect data over the target area (ice tank), we pulled the GPR sled over the ice in grid transects separated by 15 cm at approximately 0.3 m/s in the north-south direction (i.e. perpendicular to the long axis of the carcasses [1]); see Figure 5). To survey the entire ice surface, we completed 12 scans with the 500-MHz antennas and 12 scans with the 1-GHz antennas.

We performed the initial scans when the ice was 5-cm-thick using the 1-GHz antennas. We used the 1-GHz antennas initially since this frequency is most popular in controlled research and it has provided adequate detail of pig carcasses in previous research [1].

We conducted control scans without any carcasses under the ice (Day 0). We placed the first two carcasses up against the bottom of the ice and conducted the first experimental scans 2 days later (Day 2) after the pig insertion hole had re-frozen. We conducted subsequent scans on Day 11 (ice thickness 26 cm) after adding a third carcass up against the bottom of the ice, and again on Day 19 (ice thickness 42 cm encasing the pigs). Scans on Day 11 and Day 19 used both the 500-MHz antennas and the 1-GHz antennas to assess if the deeper penetration depth of the lower frequency band provided a search advantage when searching for a deeper target.

Data imagery and analysis

We used the software provided by the GPR manufacturer (Sensors and Software) to conduct preliminary data visualization. As we dragged the sled over the ice, the GPR measured the subsurface. By providing the software with the approximate sled speed of 0.3 m/s, the system software produced a radargram (also referred to as a B-Scan) of the subsurface. Based on the return signal time and the strength of the return, the software produced a two-dimensional image of depth on the vertical axis and approximated drag distance on the horizontal axis (Figure 7). Pixel intensity in these images indicates the presence of a scattering target [10].

Figure 7.

Single image scan at 500 MHz through 42 cm of ice. The three carcasses can be seen at positions of 2.43, 3.19 and 3.70 m.

In Figure 7, the left-most columns of the image show clear horizontal stripping that can be attributed to a consistent air/ice interface, air/water interface, tank bottom and aliasing from multiple reflections. As the columns proceed to the right (corresponding to the sled moving north along a transect), parabolic features emerge. These parabolas indicate potential subsurface targets. The beamwidth of the antenna picks up the targets before the antenna is directly overhead, and the algorithm translates this information into an over-estimated depth until the antenna is directly over top of the target. Positional errors occur in a radargram if the GPR is not being pulled at a consistent speed. For example, a portion of the left region of the figure is the result of the GPR being stationary for some initial period while the GPR data were being collected. Our testing setup was not perfectly consistent in synchronizing the start of the GPR scan with the motion of the sled along a transect, and, as a result, the longitudinal axis of the radargram is not necessarily meaningful. For the purposes of the target detection sought herein, this is not critical, as in practice a detection algorithm can be deployed in real time while the GPR is scanning, resulting in precise knowledge of the target location.

The radargram in Figure 7 clearly shows two targets near the surface, corresponding to the first two pigs at 2.43 and 3.19 m along the horizontal axis. Note that the vertical axis starts at −20 cm, which should coincide with the ice surface, as a result of default calibration of the device. A third target shows up approximately 3.70 m along the horizontal axis, at a depth of around 20 cm (or 40 cm from the top of the plot). This target aligns with the third pig, buried deeper below the ice.

A radargram is a standard analysis tool in GPR that, on its own, depends on visual assessment for target detection. Automating the presence of targets from radargram data is a rich area, including recent machine learning approaches [11]. In this work, we present a very simple image processing approach meant to make the final image used for target detection easier to interpret as follows. For each transect, we obtained a corresponding radargram. We averaged a sliding window of columns, with appropriate modifications for the window width at the boundary of the radargram. That is, if the radargram is represented by an matrix:

(1)

with length- columns , then the window-averaged radargram is also an matrix:

(2)

that has the th column

(3)

To detect the presence of an object at the th location along the tank, we calculate the (Pearson) correlation coefficient between columns in the averaged radargram . Given two columns and their correlation is calculated as follows:

(4)

where is the mean of the entries , in the column . The magnitude of the correlation coefficient is between and . Magnitudes near 1 indicate that the vectors and are strongly correlated, while low values suggest they are not related. Applying autocorrelation to neighbouring columns of a radargram will thus indicate if anything has changed as we slide the sled along the ice. On the left-hand-side of the image, neighbouring columns are practically identical, and will result in correlation values near 1. Once we approach the parabolic anomalies, neighbouring columns change, reducing the correlation value. If, however, this change occurs gradually from column-to-column, the correlation coefficient may still stay quite high. For this reason, we compare averaged columns in separated by a stride . Thus, for each column in , we compute a detection value :

(5)

A sample detection function , viewed as a function of longitudinal index , is shown in Figure 8 for averaging width and stride . In this view, targets are generally represented by two peaks corresponding to the change in radar signature at the boundaries of the target.

Figure 8.

Detection function obtained from a single radargram with width and stride .

By aligning adjacent passes and assigning colours to the correlation values, we obtained heat maps that indicated the presence of the pig carcasses beneath the ice.

Results

Figure 9 includes north-south cross sectional radargram reflection profiles for both 500-MHz and 1-GHz antennas at various ice thicknesses. Due to variations in the pull-rate of the sled, the horizontal axis (corresponding to the longitudinal direction) varies from scan to scan. With 5 cm of ice, the two pig carcasses, which are up against the ice but surrounded by liquid water, are clearly seen with the 1-GHz antenna. There were no obvious differences between the nude (left) and clothed (right) carcasses. With 26 cm of ice, the original two carcasses are encased in solid ice and the third carcass is visible up against the ice but surrounded by liquid water. With 42 cm of ice, all carcasses are encased in solid ice and visible with both 500-MHz and 1-GHz antennas. The 1-GHz antenna provided more detail allowing clearer demarcation between carcasses.

Figure 9.

Ground penetrating radar (GPR) scans at different ice thicknesses with 500-MHz and 1-GHz units. Two carcasses were under 5-cm-thick ice and a third (on the right) was under 26-cm-thick ice.

Figure 10 includes horizontal heat map profiles in the north-south axis for both 500-MHz and 1-GHz antennas at various ice thicknesses. We produced these heat maps by stacking correlation plots like that shown in Figure 8 for each grid transect. Per the previous discussion, the heat map shows the general outline of anomalies under the ice. With 5 cm of ice, the two carcasses are visible with the 1-GHz antenna, with the clothed carcass seeming to provide more signal. This prominence continues at all depths with the 1-GHz antenna but the difference is not obvious with the 500-MHz antenna. With 26-cm-thick ice, the third carcass is visible although the signal is less than for the other carcasses, and not clearly different than the signal for that area (north of the original two carcasses) with 5-cm-thick ice. With 42 cm of ice, the 1-GHz signal for the third carcass is more prominent and approaches the intensity seen for the other two carcasses. The third carcass is also visible with the 500-MHz antenna but not as prominent as with the 1-GHz antenna.

Figure 10.

Ground penetrating radar (GPR) heat maps at different ice thicknesses with 500-MHz and 1-GHz units.

Discussion

Summary of results

This study served as a proof of concept, demonstrating that a GPR can detect carcasses beneath and encased in ice. We demonstrated that while the 1-GHz antenna seemed to show more definition, both 500-MHz and 1-GHz GPRs could detect carcasses under ice up to 26-cm-thick, even when the carcasses were fully encased within the ice. Clothing had little effect on vertical reflections but enhanced the horizontal hot spot views. Using vertical reflections, carcasses that were up under the ice, but surrounded by liquid, were easily detectable under 5 cm of ice (carcasses 1 and 2) but not as prominent under 26 cm of ice (carcass 3). At both 5- and 26-cm ice thicknesses, the signals became more prominent once the carcasses were encased in ice compared to being in liquid water. Horizontal hot spots provided similar qualitative results, the most striking being the carcass under thicker ice became much more visible once it was encased in ice. While the primary purpose of this study was to detect carcasses beneath the ice, we also noted that ice thickness quantification was also possible using both the 500-MHz and the 1-GHz antennas.

Implications

The results of this study imply that GPR in the frequencies of 500 MHz–1 GHz is a viable tool for detecting drowning victims in the scenarios where extensive bottom searches are unsuccessful. While we saw resolution improvements using the 1-GHz antennas, both the 1-GHz antenna and the 500-MHz antenna could detect carcasses beneath thin ice (less than 30 cm). Since most drowning victims who are under ice broke through weak, thin ice, this strengthens the viability of this method.

Given the relative simplicity of the autocorrelation detection methods presented in this paper, immediate (in-field) feedback is feasible providing suggested areas for further investigations.

Limitations

While this study has demonstrated feasibility of GPR for detecting carcasses beneath the ice, there are some limitations that should be acknowledged. Firstly, the small tank we used for the trials eliminated much of the expected noise that may naturally occur in a river or lake. Second, no carcasses were placed under ice thicker than 26 cm. Most bodies in this scenario would either have fallen off an ice edge into open water and moved under the ice, or broken through thin ice; therefore, being up under thicker ice is unlikely. Finally, the first two pigs had inflated lungs (via intubation), while the third did not. This is unlikely to be responsible for the minimal signal for the third carcass under 26 cm of ice as the signal was clear once the carcass was fully encased in ice.

Future work

Future studies will deploy the GPR sensing system into more realistic settings such as natural rivers and lakes. To improve the feasibility, we will explore the possibility of using drones to drag the GPR across the ice, which would improve the safety for search and rescue personnel (eliminating the need for them to walk over thin ice to search for victims).

Future systems may include less expensive GPR systems, custom-built to be lighter and easier to drag with a drone. This custom system will use multiple antenna swaths to increase the scanning width and thereby decrease the search time over a larger body of water. The custom system will also include a localization system, eliminating the ambiguity of the drag speed. Future trials will also explore thicker ice, up to 90-cm-thick to determine the limits of GPR detection through ice. Further, blind search trials will be staged using GPS coordination to determine the effectiveness of this detection system.

Summary and conclusions

In summary, this study demonstrated the feasibility of GPR as a sensor for detecting drowning victims trapped below thin ice. We used pig carcasses to simulate child-sized human drowning victims under ice up to 26-cm-thick. We dragged GPR antennas at frequencies of 500 MHz and 1 GHz across the ice in parallel swaths and used a autocorreclation-based detection algorithm to identify areas where carcasses were likely located. Future work will focus on developing an end-to-end system that relies on autonomous drones to drag the GPR sensors across the ice for automated detection. This system could improve search effectiveness in scenarios where bottom searches have failed to locate a drowning victim.

Authors’ contributions

Gordon G. Giesbrecht developed the original question and protocol. Gordon G. Giesbrecht and Philip Ferguson conceived the study. Gordon G. Giesbrecht, Mitesh Patel, Darren Xie, Scott Murray, Vrushil Patel, Rafid Javid, Noah Wiens, Ian Jeffrey and Philip Ferguson designed the tests, carried out the study, and drafted and reviewed the manuscript. All authors contributed to the final text and approved it.

Compliance with ethical standards

This study was approved by Fort Gary Campus Animal Care Committee of the University of Manitoba (2021–25). All procedures performed conformed to the Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care Committee of the University of Manitoba.

Disclosure statement

The authors report that there are no competing interests to declare.

Funding

This work was supported by the Natural Science and Engineering Research Council of Canada (NSERC).