Prev: P1.6 Next: P1.10

P1.9: McWhirter, Paul Ross
Paul Ross McWhirter (Liverpool John Moores University)
Josh Veitch-Michaelis (Liverpool John Moores University)
Claire Burke (Liverpool John Moores University)
Steve Longmore (Liverpool John Moores University)
Owen McAree (Liverpool John Moores University)
Serge Wich Liverpool John Moores University


Theme: Machine Learning in Astronomy
Title: Saving endangered animals with Astro-Ecology

Conservation science is experiencing an unprecedented challenge in identifying and protecting endangered species across the world. The large stretches of land and sea require innovative solutions for the monitoring of endangered populations. Drones equipped with high resolution cameras with supporting data from satellites have helped to mitigate these challenges. Unfortunately, it is difficult to detect animals from optical images when they might only be a matter of a few pixels across from the perspective of an overhead drone. The Astro-Ecology research group at Liverpool John Moores University is working on deploying thermal infrared cameras on drones to detect animals from their body heat and therefore at night as well as during the day. In the thermal infrared band, animals appear as bright sources on a dark, colder background. Using these techniques, multiple populations have been classified and studied. As the project scaled up, the initial manual classifications of animal populations became unfeasible given the quantity of the collected thermal images. The nature of these images attracted astrophysicists to join the project as the automated detection and classification of bright sources on a dark background has been of particular interest to observational astronomy for decades. The application of readily available software such as Sextractor was surprisingly successful identifying the bright sources in the data. This success was short-lived as the thermal background could change strongly depending on the time of day and the local climate. Using the Moderate Resolution Imaging Spectroradiometer satellites (MODIS), a model of land surface temperature variation across daily and yearly timescales was produced. This allowed for future data gathering flights to be optimised for the local conditions prior to the expedition. The atmosphere also results in an absorption effect on the observed sources in the thermal infrared primarily due to water content as a function of the distance from the camera to the source and the air temperature. These sources of error must be corrected for the successful identification of an animal’s surface temperature. As the primary goal of this project is the detection and classification of animals, it is important that they be distinguishable from the background and have thermal profiles indicative of their species. This requires the development of an image pre-processing pipeline, similar to the calibrations applied to astronomical instrumentation, so that the detection and classification algorithms are applied to data with the expected thermal profiles where the temperature recorded by the instrument matches that of the target source. Individually modelling the thermal profile of every animal is quite difficult yet a large quantity of data has been collected containing these thermal profiles (once pre-processed). We are developing a machine learning pipeline based heavily on the approaches in modern computer vision which are simultaneously being employed within current astronomical image-based approaches. Our approach to this task requires a large set of labelled training data of multiple species. Whilst we have collected a large quantity of data, it would take a long time to manually identify the sources of interest. To address this problem we are using citizen science through the Zooniverse web portal. Using a detection method based on thresholding the thermal data based on the land surface temperature conditions, we identify sources with higher than expected temperature. The Zooniverse site requests users to perform a set of tasks designed to remove false positives, identify missed sources and classify the detected animals. Upon the completion of the citizen science workloads, the training data can then be applied to a selection of computer vision machine learning algorithms to produce models for the detection and classification of animals from their thermal profiles.

Link to PDF (may not be available yet): P1-9.pdf