The snow leopard, the large cat native to the mountain ranges of Central and South Asia, is a highly endangered species. With an estimated estimated 3900-6500 individuals left in the wild, conservation efforts led by the Snow Leopard Trust are focused on preserving this iconic animal.
But the snow leopard is an elusive creature: given their range and emote habitat (including the highlands of the Himalayas), they are difficult to study. In order to gather data about the creatures, researchers have used camera traps to capture more than 1 million images.
But not all of those images are of snow leopards. It's a time-consuming process to classify those images as being of snow leopards, their prey, some other animal or nothing at all. To make things even more difficult, snow leopards have excellent camouflage, and can be difficult to spot even by experienced observers. To cut down on the 300 person-hours per camera survey (containing up to 100,000 images), Microsoft's Azure Machine Learning teamed up with the Snow Leopard Trust to build an automated classification system.
The system is based on a convolutional neural network (CNN) which used is to classify the camera trap images. Rather than training a CNN from scratch, transfer learning was applied to the ResNet-50 model, which has already been trained on images of objects, animals and people. The output from the upper layers of ResNet-50 is then used to build a traditional logistic regression model on human-classified images stored in Spark, using the open-source MMLSpark library. (You can see a similar example of transfer learning with MMLSpark here.)
This classifier is then made available to the researchers as an on-line service in Azure — "Snow Leopard Classification as a Service", if you will. All the researchers need to do is upload a file of images from the latest camera trap survey, and the system automatically identifies those that include a snow leopard. You can see the system in action in this presentation by Joseph Sirosh at the Strata Data conference earlier this month.
To learn more about the project, check out the blog post linked below.
Cortana Intelligence and Machine Learning Blog: Saving Snow Leopards with Deep Learning and Computer Vision on Spark