At noon, on a hot March day, in the flat semi-arid plains of Kachchh, a lone figure stood on a slight rise, fiddling with the controller of a drone that hovered a few feet off the ground. Soon, the whine of the drone faded as it lifted up into the sky and headed towards the sandy banks of a nearby pond, scanning the ground for mugger crocodiles (Crocodylus palustris).
As a basking crocodile came into view on the drone’s camera, it dropped down to about 10 feet above the reptile and began filming a 30-second video of the animal’s back, in a slow clockwise motion. Once done, the drone lifted up into the air again and moved on along the pond’s bank in search of another crocodile to film.
This exercise was at the core of an effort by researchers to develop a system that can identify individual mugger crocodiles. Brinky Desai, Ratna Ghosal and their team of collaborators from Ahmedabad University, Gujarat, have developed a model using artificial intelligence with these drone-collected images, to identify individual muggers from the pattern of scutes (hard bony plates) on the crocodiles’ backs.
“The scutes on the backs of crocodiles form a grid-like pattern that is unique to each individual, much like a fingerprint. We used these patterns and a deep learning approach using neural networks to create a biometric identification system for muggers,” explains Ghosal.
Identification is important
“Every year, there are many human-crocodile conflicts in Gujarat, especially in Vadodara; and more than half such encounters are fatal. These conflicts occur often in Vadodara city because people – especially labourers from other states – are unaware that muggers are found in the Vishwamitri river,” Desai tells Mongabay-India.
“In addition, cattle grazers and farmers in the rural areas of Vadodara (that contain nearly 90-100 villages) often try to take shortcuts by crossing rivulets where the crocodiles have territories. This leads to fatal attacks or severe injuries with people losing arms and/or legs,” she adds.
Very often, after such events, muggers are caught in traps and relocated to other areas. However, there is currently no way to know if the correct crocodile – that is, the crocodile which attacked or has had a history of attacking humans, was caught and relocated, according to Desai.
The model that Desai and Ghosal have developed with their collaborators, uses images of muggers to identify specific individuals – something that could be very useful in rescue and relocation efforts to minimise human-crocodile conflicts.
“There is a burden of human-wildlife conflict with many animals. Both human and wildlife are precious; so, any method that is able to unambiguously identify animals is very useful to deal with this problem,” says Karthikeyan Vasudevan from the Laboratory for the Conservation of Endangered Species at Centre for Cellular and Molecular Biology.
“This work is an example of how technology such as artificial intelligence and drones can be applied to address a problem that is important. However, one important caveat is that this method should be tested on large populations where inter-individual variations might be small,” he adds.
Desai and Ghosal are now in the process of developing an app that can be used for such purposes and believe that it would be very useful for the forest department and non-governmental organisations that carry out census studies.
“The forest department in Gujarat has been very supportive of our project and has shown great interest in using the app,” says Ghosal. “The app could help in generating better census data. Since muggers also migrate, it can sometimes be difficult to get accurate data based on visual estimations, especially as the same individual may be counted multiple times,” she adds.
When developed, this system can function similarly to how tiger stripe patterns from camera traps are used in tiger censuses.
Analysing the biology
On a broader level, identifying individual muggers is also important for understanding their biology. “Mugger crocodiles in India are distributed across many isolated pockets in 12 states, with fairly small populations. Many of these populations live in different and sometimes unique habitats with different challenges to their survival,” states Ghosal. Due to this, it is likely that there are many physiological and behavioural differences among mugger populations.
The study focused on three locations within Gujarat, which are radically different from each other. Desai explains that Vadodara is a highly polluted river ecosystem where crocodiles always have access to water bodies, but human-crocodile conflict is high.
Kachchh, on the other hand, is a semi-arid desert with highly seasonal water availability and barely any human-crocodile interaction. Anecdotal data indicates that crocodiles in Kachchh may migrate to other places during dry spells.
And finally, Anand – in the Charotar district – is a pond ecosystem with very low human-crocodile conflict, despite high populations of both. In Anand, the locals worship crocodiles as the vaahan (vehicle) of Varuna (god of the sea) and their attitudes towards crocodiles are primarily conservationist because of their religious beliefs.
How do these crocodiles adjust – both physiologically and behaviourally – to these diverse local environments?
“As of now, very little is known about this,” according to Ghosal. “We don’t know how these conditions affect the muggers’ growth, their reproductive cycles, nesting behaviours, parental behaviours and more,” she remarks. However, such data would be important for both, conservation efforts and effective ways to minimise human–crocodile conflicts.
To answer such questions on mugger physiology and behaviour, one needs detailed data obtained from close observations of several focal individuals over extended periods of time. Such tracking systems help in gaining a general understanding of how crocodiles live, grow, reproduce and behave in each of these distinctive settings.
How it works
Unlike land animals, for example, elephants – where one can identify an individual using traits like height, shape or presence of tusks, markings or notches on the hides and ears – there are few physical traits on crocodiles that allow identification.
“There really aren’t many physical features one can use to identify muggers because most of their bodies are either submerged in water or covered in mud,” says Desai. “One way to get past this limitation is to ‘tag’ certain individuals, but it is not only difficult to get permits to do this, but also very dangerous,” she states. This is because tagging for identification involves either clipping scutes or securing GPS systems on the backs of the crocodiles, both of which require the capture and handling of the animal.
“But based on several studies on captive crocodiles, we realised that the scute patterns on the muggers’ backs could be used for individual identification. Also, since most crocodiles bask in the sun during the day, capturing images of their dorsal areas (backs) was a feasible option,” explain Desai and Ghosal.
The scutes on each crocodile’s back form grid-like patterns. By marking the presence (coded as a 1) or absence (coded as a 0) of scutes, one can create a unique binary pattern, much like a fingerprint, to identify a particular individual.
However, doing this manually would be hugely challenging. So, Ghosal and Desai turned to several collaborators to see if artificial intelligence could be used to create an efficient recognition system.
This is where their collaboration with Mehul Raval proved to be crucial. Raval is a professor at the School of Engineering and Applied Sciences in Ahmedabad University who specialises in several aspects of deep learning, including surveillance and identification.
“The most difficult part for me was in understanding the language of ecology. But once I had a grasp of it, the underlying system of work in both the fields – ecology and computer science – which is based on scientific enquiry and experimentation, made this work very interesting for me,” says Raval.
The team collected images of crocodiles’ backs from a total of 19 locations in Gujarat – 10 in Kachchh, six in Charotar and three in Vadodara. They then used 88,000 images from 143 individual crocodiles to ‘train’ a convolutional neural network (a deep learning algorithm) known as YOLO-v51 to recognise mugger individuals. YOLO, in this context, is an abbreviation for the term ‘you only look once’ to highlight this algorithm’s speed, high accuracy, and learning ability to identify objects in real-time.
The trained model, when tested, was able to re-identify an individual nearly 89% of the time and could detect an unknown individual (a pattern that the model had not encountered before in its training) nearly 90% of the time. The model could also detect if an image belonged to a non-mugger species with 100% accuracy.
Fieldwork and algorithm
Getting to this stage in the development of a biometric identification system for muggers was not an easy task for the researchers. One of the most difficult parts was the fieldwork, according to Ghosal.
“Collecting data from Kachchh was relatively very easy since the land is quite flat. But in Anand, there are trees and thorny bushes under which the crocodiles bask – in such situations, it was impossible to get images as I couldn’t take the drone into such spaces. Not only could the drone get stuck, but the noise it made could also spook the crocodiles,” says Ghosal.
In addition to these challenges, Desai also had to fend off crowds of curious villagers fascinated by the new technology as well as angry birds – mostly raptors and crows – that would dive at the drone or flock around it and obstruct the drone’s camera.
“Once the drone disturbed a beehive just out of my line of sight. Visibility through the camera was ruined, and I almost panicked because I couldn’t directly see the drone itself. Thankfully, the drone comes with a feature that allowed me to direct it to return to the place from which it was launched. That feature really saved the day,” Desai recalls.
Another issue that the team faced was with no-fly zones within Vadodara city where they could go to very few places to collect data. “We could not fly anywhere near the airport or any government offices due to security concerns,” says Desai.
Apart from fieldwork, identifying the right algorithm and getting it to work was the next biggest challenge that the team faced.
“This work is a mix of ecology and artificial intelligence. We, as ecologists didn’t know much about AI and we had to find a collaborator who was not just an expert on artificial intelligence but also willing to learn some ecology and understand the limitations of how much and what types of data could be collected,” says Ghosal.
Currently, the model that has been developed still has some issues that need to be ironed out. For example, the model does not work very well if the images contain more than one individual, if the scutes are partially submerged or if the images are dark.
In addition, the memory and computational resources required to train the model on newly collected data is also quite large. Despite these problems, the team remains optimistic about rolling out an app-based on this work.
“It’s a work in progress, and we have some hopes and expectations that we will be able to deploy a version of the app in a testing mode for smartphones fairly soon – in about six months or so,” says Raval. “Meanwhile, all this image data on mugger crocodiles and details of our model training and testing systems are freely available online, so any interested party can work on it to speed up this process,” he adds.
This article first appeared on Mongabay.