Abstract:

Only 4.5 percent of Singapore's land area is set aside as nature reserve. Despite this, Singapore is an important stopover point for more than forty species of migratory birds along both the Central Asian Flyway and the East Asian - Australasian Flyway. The birds are attracted to the wetlands and mudflats of the island-nation.

The Internet of Things (IoT) has afforded the investigation of geographies at scales that have hitherto not been practicable. The relatively low cost of open-source hardware means that localized spaces can potentially be saturated with IoT sensors, with the resulting datasets analysed through modelling in Data Science. Vey and Storring (2022) have appropriated the term ‘hyperlocal’ to refer to “place governance in a fragmented world”, by which they describe the distinct geographies of places at the scale of downtowns, waterfronts, and innovation districts.

This paper reports an independent research project conducted by a pair of high school students between April 2022 and March 2023, under the mentorship of a senior research scientist at the National Institute of Education in Singapore. It is inspired by Sustainable Development Goals 4 and 15 – namely ‘Quality Education’ and ‘Life on Land’ – of the United Nations. The paper documents the identification and evaluation of Deep Learning Models in ornithological identification.

In a presentation at the 2020 Geography Teacher Educators’ Conference hosted by the Geographical Association of the United Kingdom, Witt has described the hyperlocal as referring to “a very small geographical community, such as … a local patch”. Similarly, Matthews (2020) has suggested to us that the pandemic has catalysed renewed interest in the hyperlocal and in personal geographies.

This paper aims to explore the intersections of such geographies (of the hyperlocal and personal) through the lenses afforded by Data Science and the Internet of Things. Specifically, the capabilities of two neural network frameworks, VGG16 and InceptionV3, were identified due to their ability to build image classification models. The models were trained on publicly available datasets comprising more than 70000 images, representing more than 400 species. By the end of the project, the iteration of the model achieved a validation accuracy of 73.04% on forty locally sighted species. As an example of misidentification, an image of a juvenile Barred Eagle-Owl was classified as a White-headed Fish Eagle due to the similar white colouration. The bird prediction could be impacted by the lack of a variety of training data which represents the bird at all stages of its life, all while including images of both genders. A web-based service which linked to the external Singapore Birds Project was subsequently designed as a proof of concept for future scaling.

Keywords:

ornithology, deep learning, citizen science, sustainability education, biodiversity education, neural network.