Call for Albatross/Seabird Drone Survey Data
The Wildlife Conservation Society, Duke University, and the University of Tennessee – Knoxville are working together to adapt and train a convolutional neural network (CNN) to detect a wide range of species in variable environments. To do this, we need data! We need tens of thousands of drone images with nesting albatross to input into the CNN to teach, train, and validate the algorithm.
Training a Transferrable Convolutional Neural Network to Detect Albatross Species Globally
For decades, researchers from the Wildlife Conservation Society traveled to two remote islands on the Falkland Islands (Islas Malvinas) to conduct ground surveys of the largest nesting black-browed albatross colony in the world. Due to the large size of the colony and steep terrain, researchers would spend weeks counting and extrapolate data from sample transects to estimate population size. In recent years, we have started to use an off-the-shelf Phantom 4 Pro quadcopter ($1,599.99) to design island-wide surveys and collect high-resolution imagery data that can be replicated each year. Furthermore, WCS partnered with researchers from the Duke University Marine Robotics and Remote Sensing Lab to develop a deep learning algorithm trained to detect nesting albatrosses with a computer model accuracy of 97%. This level of accuracy, coupled with the ability to easily replicate preprogrammed flights, presents a much easier, more accurate, and more consistent way to document population growth or decline in colonial nesting seabirds.
Automated Detection of Black-Browed Albatrosses and Southern Rockhopper Penguins
WCS and Duke University have recently partnered on a drone-based survey of black-browed albatrosses and Southern Rockhopper Penguins on Grand Jason and Steeple Jason in the Falkland Islands (Islas Malvinas). Using data from 12 drone surveys flown by WCS researchers across two years, (avg. resolution: five cm/pixel in 2018, one cm/pixel in 2019) Duke analysts created a model to automatically detect and count albatross. During 2018 and 2019, the model was able to detect a total population of 268,764 nesting albatross on Steeple Jason and Grand Jason, with an accuracy compared to manual counts dependent upon the survey area (the model population count and accuracy for the two largest bird survey areas were 133,075 at 2.0% and 57,360 at 9.4%). These first island-wide surveys will be the basis for understanding population dynamics of a species threatened by climate change. The results of this study are published in The Condor: Ornithological Applications.
Next Steps
This model, a deep learning algorithm called a convolutional neural network (CNN), has the capacity to be a transferrable model that can be trained and used interchangeably for researchers across the globe, which is why we are reaching out to leaders in the field to collaborate on this project. Our goals are:
- Apply CNN to other nesting black-browed albatross colonies.
- Train CNN to detect other species of albatross with similar color pattern and nesting habits.
- Assess CNN accuracy when transferred to colonial nesting seabirds.
- Assess suitability of CNN for transfer to other nesting animals.
How You Can Get Involved
Building upon the success of this model, WCS, Duke University and University of Tennessee – Knoxville are working together to adapt and train the CNN to detect a wide range of species in variable environments. To do this, we need data! We need tens of thousands of drone images with nesting albatross to input into the CNN to teach, train and validate the algorithm. We are seeking drone imagery that fits the following criteria:
- Flight path in overlapping parallel lines with sufficient overlap to create an orthomosaic
- Sensor angle nadir, but oblique imagery would be helpful as well
- Minimum 10,000 individuals per dataset
- Minimum 500 photos per dataset
- Minimum 16MP resolution RGB camera
- Flight height recommended between 50 and 100 meters, preferably below 80 meters
- Sufficiently large dataset if albatross images are in a mixed colony
This software will be designed to be open source and will be available to any researcher to improve albatross research globally. If you have data you would like to contribute to these efforts and have an interest in collaborating with WCS, Duke University and University of Tennessee on this project, email Dan Simberloff, tebo@utk.edu, Walter Sedgwick, wsedgwick@msn.com, Wieteke Holthuijzen, wholthuijzen@gmail.com, Wade Sedgwick, wsedgwick0@gmail.com, or Dave Johnston, david.johnston@duke.edu.
