AI ecologist: Developing algorithm for wildlife monitoring of the DMZ

CAS featured stories 2

AI ecologist: Developing algorithm for wildlife monitoring of the DMZ

Myung Ae Choi

The Korean Demilitarized Zone (DMZ) is often regarded as “the treasury of ecology”. This 248-km long ribbon of territory is left “untouched” since the armistice agreement in 1953 due to the prolonged Cold War politics in the Korean peninsula. However, other stories also emerge, arguing that the DMZ is not untouched but actually intervened with a range of human and nonhuman aArtificial Intelligence (AI) is becoming ever more popular from face-recognition applications on smartphones to security solutions. One underexplored but rapidly growing area for AI is its environmental use. Paired with remote-sensing technologies – trail cameras, drones, satellites – AI can improve the scope and the speed of environmental monitoring by analysing a large bulk of data almost instantaneously. Google has recently launched Wildlife Insight that develops ecological AI by collecting and analysing trail cam data across the world, whereas Microsoft runs a similar project called AI for Environment. We here at Korea Advanced Institute for Science and Technology (KAIST) have also been developing ecological AI – which we tentatively call “AI Ecologist” – that can identify and count the cranes in and around the Korean Demilitarized Zone (DMZ).

Cranes in the DMZ

The Korean DMZ is a 4-kilometer-wide and 248-kilometer-long territory across the Korean peninsula. It was established as a buffer zone at the end of the Korean War (1950-1953) to deter military conflicts between North and South Koreas. Remained as a no-man’s land for over 70 years, the DMZ is now transformed into a nature reserve, the “accidental” habitat for 91 endangered species. One of the wildlife in peril are the migratory cranes.

Red-crowned cranes (Grus japonensis) and White-naped cranes (Grus vipio) spend summer in Siberia and fly down to the DMZ and surrounding areas in winter (November to March). The grains from rice field provide enough food resources, while the restored wetlands inside of the DMZ offer safe resting places. The number of cranes increased from several hundreds in the 1990s to several thousand by now, making the DMZ and the vicinity as one of the most important wintering sites for these internationally endangered species.

Fig 1. Red-crowned cranes (left) and white-naped cranes (right). copyright: Yoo Seung-Hwa

This research project aims to document nonhuman species, including but not limited to endangered species, in selected sites around the DMZ. It differs from the existing ecological surveys in three ways. First, it focuses on ecological data produced at the local level, rather than another overview of the DMZ in general. Second, the scope of nonhuman species is more inclusive tCranes have been highly appreciated in Korean traditional culture, viewed as a symbol of long-life and good fortune. We had good population of wintering cranes but lost most of them during the past century through colonial hunting, urbanisation, and the Korean War. It was therefore a welcome surprise that the cranes started to come again to the DMZ and surrounding areas. However, the cranes in and around the DMZ are now faced with the new challenges presented by human encroachment into their habitats – relaxed border controls, greenhouses, roads and other infrastructure – in addition to changing climate. Still, ecological monitoring of the cranes is fairly restricted due to the security reasons if not the remaining landmines inside of the DMZ.

Counting cranes with AI

The research team is currently developing a citizen science platform for local residents to participate in environmental monitoring in their own neighbourhood. Prof Woontack Woo, Augmented Reality Research Centre at KAIST, is keen to design and implement an open data platform where local residents can identify the selected nonhuman species, and document and share their findings by using a Digital Twin. Prof. Buhm Soon Park, a historian at Center for Anthropocene Studies (CAS) and Graduate School of Science and Technology Policy, explores the affirmative potential of participatory research platforms for “information environmentalism”(Fortun 2004), through which local residents and researchers can enhance their knowledge and sensibility toward the environment. While sharing the enthusiasm for citizen science, Dr. Myung Ae Choi, an environmental geographer at CAS, is particularly interested in the geographies and modes of encounters, through which local rTo improve the capacity of crane monitoring where ecologists have limited access, we turned to AI and trail cameras. Ecologists monitor the changes in the crane population by painstakingly counting the number of the birds with their eyes and hands. It would help ecologists greatly if AI can provide the population estimate. We decided to apply the crowd counting algorithm, which is designed to count human population by analysing photographic images, to count the crane population. The crane ecologist in the team and his colleague generously shared 1,500 crane photos to train the AI.

Here in and around the DMZ, we have two different species of cranes, each of which have two distinctive age groups of juvenile and adult birds. To reflect these distinctive features, we trained the algorithm to be able to identify different species and age groups, and count their number respectively. This means that we have five classes – adult and juvenile red-crowned cranes, adult and juvenile white-naped cranes, and great white-fronted geese. After sets of training, the algorithm come to provide five class-specific density maps. By aggregating the weights assigned to the various colours of the dots, AI can provide the estimate population of each class.

Fig 2. Crane-counting algorithm

Fig 2 shows how the AI works. In the input image, we have two classes – adult and juvenile white-naped cranes. The ground truth informs that the image contains 39 adult white-naped cranes and 7 juvenile ones. Then the algorithm provides two density maps and the numbers: 42 for adults, and 4 for juveniles. Voila! Not yet perfect but it works.

Monitoring through trail cameras

The next step is to test and improve the crane counting algorithm to analyse trail camera data. Last winter, we have set up 13 trail cameras in the rice fields and river banks of Cheorwon, Gangwon Province, where the majority of cranes spends winter. The cameras are located outside of the DMZ to comply with the security restrictions. Crane-enthusiastic farmers, who have experiences in monitoring devices, have kindly allowed and helped us to install trail cameras within their rice fields. While we cannot, the cranes fly over the fortified lines, making use of both the wetlands of the DMZ and rice fields just outside of it. One camera was installed at the crane observatory that watches over the Hantan River (Fig 3). This particular camera is set up and being monitored in collaboration with the Graduate School of University of Zurich. It records the changing landscape and wildlife of Hantan River for the upcoming exhibition in Zurich as well as the AI training at KAIST.

Fig 3. Trail camera at the Hantan River Crane Observatory

From six-month operation, we have collected 97,000 photos and 23,858 short videos. We are at the moment training the crane-counting AI to analyse the trail camera images. From the trail run, we learned that the AI trained with human-taken images do not necessarily work strong with machine-taken images. Ecologists’ photos and trail camera photos have some distinctive features in terms of their modes of operation, camera angles and resolutions. Trail cameras do expand the scope of environmental monitoring as they allow close-up observation of wildlife in remote locations especially during the night.

However, these remotely produced images pose extra challenges for AI that aims to identify and count the wildlife. We are searching out and applying leading-edge AI technologies to improve the AI to solve these difficult problems.By developing AI for crane monitoring, we hope to illustrate the utility of, and the need for, wildlife monitoring assisted by remote-sensing devices and AI. Such technologies would help us better understand how wildlife persists in places such as the DMZ where humans have limited access.

Myung-Ae Choi is Research Assistant Professor at the Center for Anthropocene Studies, KAIST. She is an environmental and cultural geographer, looking at the political, cultural, and technological aspects of nature conservation, with specific regional focus on South Korea and East Asia. For her doctoral and post-doctoral research, Myung-Ae looked at ecotourism, dolphin and whale conservation, and crane conservation in the DMZ.

The original version of this essay is published on Nextrend Asia.