Application of Deep Learning to Automated Species Identification Systems
Issue Date
2020-12-31Author
Khalighifar, Ali
Publisher
University of Kansas
Format
88 pages
Type
Dissertation
Degree Level
Ph.D.
Discipline
Ecology & Evolutionary Biology
Rights
Copyright held by the author.
Metadata
Show full item recordAbstract
Researchers monitor species’ populations for a variety of purposes, including surveillance of endangered or rare species for conservation goals, mitigation of public health threats caused by diseases vectored by insects, and threats to native populations by invasive species. Achieving each of these goals is essential to healthy ecosystems and human societies. For instance, species important in biodiversity conservation provide valuable ecosystem services, and disease vectors and invasive species impose heavy economic burdens, along with acute, negative impacts on human health and well-being. As such, the need for robust, efficient, and widely-applicable biodiversity monitoring techniques is at a premium. Traditional monitoring approaches generally involve labor-intensive data collection processes, with relatively short survey windows and limited spatial coverage. As such, these techniques are often unable to meet the increasing demands of global conservation and public health surveillance. Recent technology advances offer tools, such as audio recording devices, camera traps, and network systems, for transferring digital data, that have revolutionized aspects of biodiversity monitoring. However, while they can provide longitudinal, highly accurate data, these devices generate massive amounts of information that can overwhelm conservation biologists and public health providers. One solution is to apply real-time, automated systems to process and analyze such data streams. Deep-learning techniques can provide a cyberinfrastructure that can achieve real-time, automated species identifications deriving from such automated devices. They also provide the opportunity to explore, test, and discover unknown or overlooked evolutionary, ecological, and behavioral phenomena regarding target species or regions, making them potentially powerful tools for future basic research in biology. Lastly, as deep-learning techniques can be implemented in citizen-science platforms, they allow community members to participate in public health and biodiversity science in meaningful and actionable ways. Here, I present three examples of potential applications of such automated species identification systems. In each example, I explore the potential power of these tools by challenging an advanced deep-learning technique, TensorFlow Inception v3, to identify a set of taxa under different goals, all centered on accurate species identifications. The first chapter of my work delivers a comparison between a deep-learning-, image-based species identification system, and conventional classifiers for insect vectors of Chagas disease in Mexico and Brazil. The second and third chapters focus on the identification of acoustic signals made by taxa, using the trick of converting sounds to spectrograms: images representing auditory features. Specifically, the second chapter demonstrates successful application of a deep-learning model to a diverse clade of closely-related frog species in the Philippines, using single-note mating calls. The third chapter explores the potential for application of automated identification platforms to mosquito species using wingbeat patterns, with emphasis on participation of citizen scientists to improve surveillance of disease vectors.
Collections
- Dissertations [4713]
Items in KU ScholarWorks are protected by copyright, with all rights reserved, unless otherwise indicated.
We want to hear from you! Please share your stories about how Open Access to this item benefits YOU.