discussion / AI for Conservation  / 10 December 2015

Deep Learning Project Repository

Feel free to post links to projects you're aware of using Deep Learning to assist in conservation. 

Here's a few to get started.

Classification of 121 Plankton species from underwater imagery. The winning entry, and many of the top entries, used deep-learning based approaches. 

Identification of Right Whale individuals from aerial imagery. Many of the leading entries are using deep learning

Deep Learning for Large Scale Biodiversity Monitoring. Paper from Conservation Metrics, Inc. (I'm an author) describing their use of Deep Learning for bioacoustic and camera trap monitoring of endangered seabirds, invasive snakes, and more.

Forecasting and detecting deforestation from satellite images using Deep Learning. A collaboration between Orbital Insight (a startup applying deep learning to satellite imagery) and World Resources Institute (WRI) as part of their Global Forest Watch. 

Professor Kate Jones and her Biodiversity Modelling Research Group at UCL are exploring deep learning in conservation - particularly in their work monitoring bats. It sounds like they share your interest in deep learning and acoustics, David!

They give a great rundown of their work in this blog post published on the official blog of the Methods in Ecology and Evolution Journal. The post is worth a read in its entirety because it also looks to the future and discusses the big advances they're hoping to see in not only machine learning, but citizen science, calculating species abudances from acoustic data, and new biotelemetry and animal-borne data-logging techniques. 

Two of the projects mentioned in the article that focus on machine learning are: 

Bat Detectives and Machine Learning - Oisin Mac Aodha, Post Doc 

If you have ever tried to spot bats flying around at night you will know that it can be very difficult. However, bats leak information about themselves into the environment in the form of the sounds they make while navigating and feeding. These calls are often too high for us to hear, but we can use devices known as bat detectors to transform them into a form that we can record and listen to. Monitoring bat populations over wide areas or long periods can result in huge amounts of data which is difficult to analyse though. To address this problem, our group, along with Zooniverse, have setup a citizen science project called Bat Detective which asks members of the public help us find bat calls in audio recordings that have been collected from all over Europe (the infographic at the end of the post gives a bit more information on this). We have had an amazing response to date and our detectives have already located several thousand bat calls. However, to scale up monitoring, we need more automated methods of detecting calls. Using the analysis provided by our Bat Detectives, we are currently working on building algorithms that can automatically tell us if a recording contains a bat call.

To do this, we rely on an area of computer science called machine learning. One recent and exciting development in machine learning is the emergence of algorithms that learn directly from raw data. Previous to this, we had to first choose the important features of the data to present to the recognition algorithm. With these new methods, we think detection of bat calls in large amounts of data will become much easier in the next few years.


Automatic Species Identification Tools - Veronica Zamora-GutierrezPhD Student

In recent years, bioacoustic surveys have become a popular tool to monitor animal populations across the world, especially for bats. To use this method to monitor bat populations, we don’t just need to detect the calls in the recordings we make, we also need to reliably identify calls from different species or species groups. Identifying species from their calls is a challenging business in any situation. For bats it is further hampered by the lack of reference recordings for many regions and the lack of automatic classification algorithms. Tropical regions are of special concern as these megadiverse areas are experiencing more rapid environmental change and still have considerable knowledge gaps on basic species occurrences. With the collaboration of other bat researchers, we have collated the biggest and most complete bat call reference library for Mexico for my PhD (including the Mexican funnel-eared bat, pictured above). The availability of new material and new machine learning methods has allowed us to build accurate bioacoustic automatic identification algorithms for rapid biodiversity assessments of the bats in this megadiverse region for the first time. The next few years will be an exciting time for the development of biodiversity automatic recognition tools; we expect these methods (e.g. Microsoft’s Azure ML) to become more widely available to ecologists and easier to apply.

UCL is also has a deep learning project called engage - also involving some of the team from the Biodiversity Modelling Research Group at UCL. Their aim is to enable non-programming scientists to create systems that semi-automatically detect objects and events in their vast quantities of audio and visual data. Their projects include Bat Detective (mentioned in the previous post), and Underwater Classification - a collaboration with ZSL to classify the substrate types and organisms present in underwater imagery captured off Greenland. 

The team from Engage has also published an article, 'Putting the Scientist in the Loop - Accelerating Scientific Progress with Interactive Machine Learning', which identifies some of the opportunities and challenges in deploying machine learning in ecology research. 

The PAWS (Protection Assistant for Wildlife Security) project is a newly developed AI that takes data about previous poaching activities and outputs routes for patrols based on where poaching is likely to occur. These routes are also randomized to keep poachers from learning patrol patterns. Using machine learning PAWS can continually find new insights as more data is added.

Milind Tambe, a professor of computer science at the University of Southern California, leads the project. The core algorithm of PAWS is based on security games, a subset of game theory where a defender tries to optimize limited resources to prevent attacks. Using security game theory, Tambe has built algorithms used by Homeland Security, the Transportation Security Administration, and the Coast Guard that predicts where resources like agents and surveillance would best be placed to interfere with smuggling and terrorism.

Fei Fang, one of the Ph.D. students working on PAWS, had previously written a paper about how security games could be used to prevent wildlife crimes. Fang coined the phrase "green security games” to describe the field. But PAWS is the first time green security game theory has been used in the real world.

It's been tested in Uganda and Malaysia, which has helped the team refine the algorithms. 

We're looking to get a full case study for the WILDLABS.NET resources area, but in the meantime check out this article for more information about the project, and this paper for info about the field optimisation process.