Assessing The Use Case Of Neural Network Algorithms For Image Analysis In Biodiversity Assessment

Abstract

A subtype of artificial intelligence (AI), convolutional neural networks (CNNs) offer new opportunities for automated processing of image-based data in ecological research. This work examines the practical application of CNNs for ecological picture analysis to evaluate the impact of annotation at different taxonomic classification levels on model performance. We investigate the feasibility of manually annotating training data to different levels in order to evaluate the effect of different annotation strategies on CNN accuracy in ecological scenarios. We demonstrate that changes in annotation categories (animal, phylum, or morphology) have no effect on model performance, even when class counts differ significantly. As a result, annotators are forced to decide between investing the time and energy necessary to complete comprehensive annotation at the beginning of a project and enhancing model predictions at the end. These findings provide helpful suggestions for speeding up ecological research driven by AI without compromising model flexibility and performance.