New Wildlife Survey Technique

The Problem

Elephants are facing significant threats in the majority of the African countries where they reside. To design an effective conservation management strategy, it is necessary to have accurate population numbers and an understanding of herds geographical distribution patterns. Aerial surveys are the most commonly used method to survey elephant populations- there are several limitations with this technique which the use of satellite imagery for surveying can overcome: 1. Increasing frequency of surveys to update population numbers - Satellites can be tasked within 24hrs whereas it can take months, if not years, to get appropriate permissions to carry out aerial surveys. 2. Acquiring satellite imagery requires no permit hence it is possible to avoid lengthy national bureaucratic processes to gain civil aviation licenses to carry out a survey. It is possible to acquire imagery in countries where due to civil war it is not possible to conduct aerial surveys e.g. Mali- where elephant population numbers are unknown. 3. Surveying transboundary populations- Elephants reside in transboundary areas moving freely across national borders. Satellite imagery can be acquired of Transfrontier areas where populations are migrating across national borders - Aerial surveys are often only carried out in one country.

Our Proposal

Together with several machine learning experts at the University of Oxford Pattern Analysis and Machine Learning Research Group I am working to develop a new wildlife surveying technique using fine spatial resolution satellite imagery. I have a license to assess imagery from DigitalGlobe's Worldview 3/4 Satellites which is the world’s highest resolution satellite imagery at 31cm resolution. An elephant takes up a maximum of 20 pixels hence it is highly time consuming to locate them in an in an image as they move across vast landscapes. However, if identification is automated using a machine learning algorithm this can offer a new surveying technique that would be radically faster than conducting aerial surveys. I have identified wild elephant populations from GPS collar fixes and found imagery that is temporally and spatially synchronised. I used this imagery to develop training, test and validation data. This labelled imagery has now been fed into a convolutional neural network algorithm to automatically detect herds. The initial results are extremely promising and look set to provide a new elephant surveying tool. Using satellite imagery instead of aerial imagery provides an unprecedented opportunity to count populations in Transfrontier and hard to reach areas in close to real time without the need to undergo extensive bureaucratic processes to acquire aerial permits.

We Assume that...

I assume that this surveying technique will largely replace aerial surveys in the not too distant future.

I assume the cost of satellite imagery will continue to drop and resolution will increase so this solution can be easily applied by national park agencies.

Constraints to Overcome

Developing an extremely sophisticated algorithm via a convolutional neural network that will have a high level of detection accuracy.

Current Work

The CNN is undergoing further training on the University of Oxford supercomputer and we are awaiting the second round of results. The initial run was very promising at 30,000 iterations. We need to acquire grant money to scale this solution and carry out interviews with potential end-users so that the solution fits real-world needs.

Current Needs

We need funding to support our computing needs in terms of sending the programming to the Advanced Research Computing unit (ARC). Each project at the University of Oxford has a time limit on the supercomputer and the rest of the time must be covered by a grant. Once this is in place we can then optimise the initiations to train the algorithm as accurately as possible.