Biodiversity Image Quality Metadata Augments Convolutional Neural Network Classification of Fish Species

The chapter investigates the role of image quality metadata in enhancing the classification accuracy of fish species using convolutional neural networks (CNNs). It begins by emphasizing the value of metadata in machine learning and deep learning tasks, particularly in the context of biological specimen image collections. The research focuses on the NSF-supported Harnessing the Data Revolution (HDR) project, Biology-Guided Neural Networks for Discovering Phenotypic Traits (BGNN), which aims to develop novel artificial neural networks for classifying fish species and extracting morphological data from images. The study identifies the limited attention given to image quality in existing metadata standards and highlights the need for a targeted metadata scheme to capture specimen image quality. The authors conduct an empirical analysis using a sample of 23,807 digital images of fish specimens, demonstrating that high-quality images significantly improve classification accuracy. They also identify specific quality annotations that are most important for classification accuracy. The chapter concludes with practical recommendations for assessing image quality in biodiversity image repositories to support machine learning analyses, making it a valuable resource for researchers and practitioners in the field.