Working on Microsoft Azure platform, Mohanty and his colleagues used a Convolutional Neural Network model to come up with a solution that can identify and count penguins with a high degree of accuracy. The model can potentially help researchers speed up their studies around the status of penguin populations.

The team is now working on the classification, identification and counting of other species using similar deep learning techniques.

Building AI to save the planet

A long-time Microsoft partner headquartered in Hyderabad in India, Gramener is not new to leveraging AI for social good using Microsoft Azure. It was one of the earliest partners for Microsoft’s AI for Earth program announced in 2017.

“I believe that AI can help make the world a better place by accelerating biodiversity conservation and help solve the biggest environmental challenges we face today. When we came to know about Microsoft’s AI for Earth program over two years ago, we reached out to Microsoft as we wanted to find ways to partner and help with our expertise,” says Kesari.

While the program was still in its infancy, the teams from Gramener and Microsoft worked jointly to come up with quick projects to showcase what’s possible with AI and inspire those out there in the field. They started with a proof of concept for identifying flora and fauna species in a photograph.

“We worked more like an experimentation arm working with the team led by Lucas Joppa (Microsoft’s Chief Environmental Officer, and founder of AI for Earth). We built a model, using data available from iNaturalist, that could classify thousands of different species with 80 percent accuracy,” Kesari reveals.

Another proof of concept revolved around camera traps that are used for biodiversity studies in forests. The camera traps take multiple images whenever they detect motion, which leads to a large number of photos that had to be scanned manually.

Soumya Ranjan Mohanty, Lead Data Scientist, Gramener
Soumya Ranjan Mohanty, Lead Data Scientist, Gramener

“Most camera trap photos are blank as they don’t have any animal in the frame. Even in the frames that do, often the animal is too close to be identified or the photo is blurry,” says Mohanty, who also leads the AI for Earth partnership from Gramener.

The team came up with a two-step solution that first weeds out unusable images and then uses a deep learning model to classify images that have an animal in them. This solution too was converted by the Microsoft team into what is now the Camera Trap API that AI for Earth grantees or anyone can freely use.

“AI is critical to conservation because we simply don’t have time to wait for humans to annotate millions of images before we can answer wildlife population questions. For the same reason, we need to rapidly prototype AI applications for conservation, and it’s been fantastic to have Gramener on board as our ‘advanced development team’,” says Dan Morris, principal scientist and program director for Microsoft’s AI for Earth program.

Anticipating the needs of grantees, Gramener and Microsoft have also worked on creating other APIs, like the Land Cover Mapping API that leverages machine learning to provide high-resolution land cover information. These APIs are now part of the public technical resources available for AI for Earth grantees or anyone to use, to accelerate their projects without having to build the base model themselves.

Article Source