WEIGHTED ENSEMBLE DEEP LEARNING FOR AUTOMATIC CLASSIFICATION OF SRI LANKAN BUTTERFLY FAMILIES

Abstract

Butterfly diversity plays a vital role within ecosystems. Traditionally, butterfly species identification has been a labor-intensive and error-prone task, demanding expertise and time from specialists. To address this challenge and contribute significantly to the conservation of endangered species while advancing taxonomy and biodiversity research, it is imperative to introduce automated algorithms. This study presents an innovative approach that leverages deep learning techniques, specifically an ensemble comprising VGG16, VGG19, and ResNet50 architectures, for the automated classification of butterfly families in Sri Lanka. Despite the country’s rich diversity of butterfly species, prior research on automated family-level identification has been scarce. We have curated a dataset containing 2,182 images representing five Sri Lankan butterfly families (Nymphalidae, Papilionidae, Pieridae, Hesperiidae, and Lycaenidae) and applied a weighted ensemble deep learning algorithm to capture distinctive features specific to Sri Lankan butterfly families. This marks the pioneering use of deep learning techniques for butterfly identification in Sri Lanka. We assess the performance of individual models, namely VGG16, VGG19, ResNet50, and the ensemble of these three methods. The weighted ensemble method stands out with an impressive accuracy of 95% when evaluated on augmented datasets. These results underscore the effectiveness of these deep learning models in accurately identifying butterfly families within the Sri Lankan context