Deep Learning for Large Scale Biodiversity Monitoring

Editor's Notes

• #2: Thanks, it's awesome to be here. I’m David Klein, I’m an AI researcher and developer. i’m so excited to be able to give you all a flavor of what we’re doing at conservation metrics to help safeguard global biodiversity using big data and deep learning.
• #3: Our planet is blessed with richness of plants and animals that has evolved over hundreds of millions of years.
• #4: We are but one of an estimated 8.7 million species on earth, and we’re reaching across ever increasing number of ecosystems to extract goods & services that are necessary to sustain our well being and growth.
• #5: Services as diverse as climate regulation, new medicines and pesticides, and of course harvested food for economic gain.
• #6: Not to mention the recreational, aesthetic, and even spiritual benefits of nature that we all enjoy.
• #7: But now, we are losing the earth’s biodiversity at an alarming rate.
• #8: In response, people and organizations all over the world are trying a host different things to protect it….
• #9: In doing so they are spending an estimated $20 billion annually, which is great, but...
• #10: There is an ugly secret for too many of these efforts …. we're often not sure on a relative basis how well they work. We all too often lack the data to scientifically evaluate the impact of a single conservation intervention, let alone to compare multiple alternatives.
• #11: Want ROI analysis of your conservation project? Sorry those are mostly nonexistent in the conservation sector.
• #12: Thus conservation funders are stuck on the left hand side of this continuum of funding models, relying on emotion and logic to decide what projects get funded, rather than ongoing data collection and hypothesis testing so critical for effective adaptive management.
• #13: So, why is conservation so far behind the data-driven decision making revolution? In part it’s cultural, but the main reason, we believe, is that the necessary data are difficult to get -sometimes dangerous - sometimes remote -
• #14: And almost always labor intensive, invasive, and…...
• #15: expensive!
• #16: Conservation Metrics’ mission is to move the conservation world to the right side of this funding continuum by harnessing recent technological advances in sensor networks, big data, and machine intelligence to provide affordable and effective measures of conservation outcomes.
• #17: Leveraging an increasing diversity of sensor types
• #18: ...and sensor platforms for measuring the natural world. These are wonderful because they’re cheap, work 24/7, and collect lots and lots of data… the challenge that we are addressing is how to manage and analyze all these data.
• #19: it's a true big data problem, especially for governments and conservation organizations with very limited budgets. If we take our current ongoing projects, which are have substantial data volume, and scale them up to the global need, we would be scaling from PB to EB every year. Obviously this amount would be impossible to manually analyze.
• #20: That’s why we are leveraging recent advances in AI, to automate the analysis of biodiversity sensor data.
• #21: You can think of it as a global biodiversity monitoring AI. That’s the vision.
• #22: In executing on this vision, we’ve experienced huge improvements in analysis throughput due to our integration of deep learning models for signal detection and classification. Deep learning of course refers mostly to modern, large scale neural networks with potentially many processing layers that execute sequentially, that can be trained carte blanche on data to perform complex tasks with high accuracy, tasks as diverse as image recognition, speech recognition, and text interpretation.
• #23: I chose deep learning, not only because of my long experience and interests in the field, but also deep learning systems have a number of desirable properties for the biodiversity monitoring application. These include * accuracy that that scales well with the amount of training data -- and we have a lot of data * extensibility such that the same algorithm can be applied to many disparate types of data * and perhaps most importantly, many proof points of world beating performance in complex image and audio recognition problems -- such problems are conservation metrics’ bread and butter I see many future applications for leveraging deep learning strengths.
• #24: Our software was developed in house, and has several features that have proven instrumental
• #25: including tools for data visualization, fast model training
• #26: and periodic auditing and refinement of the models. If you have any technical questions about our software or our model architectures, I’d be happy to talk to you later.
• #27: Once trained, our deep learning models are used to process data coming in from the field, and output the relative probability that a data sample contains, for example, a given species. This information is used to provide ranked pages of relevant results for our analysts. Here you see an unsorted page of audio spectrograms, showing the sound energy density as a function of frequency and time. These are from a monitoring project for an elusive and threatened seabird called the marbled murrelet. It contains sounds from birds we aren’t trying to monitor, robins and jays, and other non-target sounds such as saws.
• #28: And here’s the first page after sorting with our marbled murrelet model. We can’t listen to these here, but I can tell you that they do all contain marbled murrelet calls.
• #29: For images, we are building up similar capabilities, additionally using deep-learning models to produce per-pixel heat maps for highlighting and localizing objects of interest. No time for detail on this one, but we’re receiving millions of images for monitoring invasive brown tree snakes in Guam. Can you see the snake on the left? How about now?
• #30: … and that’s a really big snake!
• #31: What has the impact been so far? I’m going to give 3 examples about how this technology has been used to inform better conservation decision making.
• #32: I’d like to start with one of our first projects focused on a newly discovered species - the Bryan’s Shearwater. This surprising discovery was announced in 2011 when three individuals were found on Midway Atoll. But they weren’t breeding there. Where did they breed?
• #33: Around the same time, shearwaters in Ogasawara were found that genetically matched, and so the search for breeding sites began in the remote and largely inaccessible islands near Ogasawara. We sent 3 acoustic sensors to help extend the temporal scale of their surveys.
• #34: Calls were found using our software, and the data were used to narrow the search area. Finally the first nests for the species were found last winter. Now they can start active conservation projects on nesting sites.
• #35: Our second project is work we’ve carried out with the Queensland Parks and Wildlife Service in the Great Barrier Reef Marine Park. This park is enormous, and there are only a handful of biologists to monitor bird populations and over 900 islands they might breed on. Traditional population counts involve walking through breeding colonies with snowshoes and using fiber optic scopes to check underground burrows.
• #36: Using our approach and collecting data over a 9 month period, we found we were able to produce nest density estimates that were highly correlated with the traditional method, and we were able to do it while reducing costs and unintended impacts. We are now attempting to scale out and monitor many more of the 900 islands in the park - this wouldn’t have been possible without the use of sensor-data driven deep learning detectors.
• #37: I will close with a final example of a project that has grown rapidly over the last 3 years. Many species of birds and bats face threats from flying into man-made structures like transmission lines, wind turbines, or skyscrapers. Quantifying these impacts is often extremely difficult and costly given the huge areas covered.
• #38: For years we’ve known that endangered species were being killed by power transmission lines in Hawaii. We are now using sensors to listen to the sound produced by these collisions (shown here). This work is funded by the Kaua’i Island Utility Cooperative short-term HCP. The use of Deep Learning based classification techniques have allowed us to scale rapidly. This season, we will be collecting more than 80,000 hours of data from sensors deployed across the island. A traditional survey effort with field personnel at this scale would be impossible.
• #39: Our clients are now making data driven decisions about how to protect these endangered seabirds three years after starting the project
• #40: Such as these lasers being tested this season to alert birds about the transmission lines. Its an incredible privilege to be able to work to advance conservation with our partners. We now have a moral imperative to do this better cheaper and at scale.
• #41: Thank You! laysan