Classification of Firefly in SARWAK using Artificial Intelligence

Abstract

Accurately classifying firefly species in their natural environments poses unique challenges. This study evaluates deep learning approaches for image-based firefly taxonomy using field-captured data. Images of fireflies from 14 species native to Sarawak, Malaysia, were compiled. Faster RCNN, SSD and YOLO object detection algorithms were trained to recognize firefly bounding boxes and predict taxa labels. Model accuracy and runtime were benchmarked. Faster RCNN achieved the highest average classification accuracy of 85% while maintaining reasonable speeds of 0.8 seconds/image. YOLO had the fastest runtime of 0.01 seconds/image but the lowest accuracy of 75%. SSD balanced both metrics with 80% accuracy at 0.2 seconds/image. Convolutional neural networks show promise for automated firefly monitoring. Expanding training data and integrating analyses of blinking behaviors may further enhance identification capabilities. This research indicates deep learning is a viable solution for large-scale, real-time firefly taxonomy using field-captured multimedia.