Deep learning pipelines for biodiversity monitoring: megadetector and vision transformer approaches with camera traps in Costa Rica

Abstract

This paper presents an automated pipeline for biodiversity monitoring using camera traps in Costa Rica. We introduce the cemede-redbioma-ct dataset, a new resource of local species images, and evaluate two alternative approaches: (i) a pipeline where a MegaDetector is used to extract bounding boxes, crop images, and then fine-tune vision transformers (DeiT, Swin, EfficientViT) for classification, and (ii) a pipeline where MegaDetector itself is fine-tuned for direct classification. Results show that DeiT achieved the best overall accuracy (82%), which, compared to other studies reporting results below 90%, is con sidered competitive. Challenges included severe class imbalance, low image resolution, motion blur, rain, low-light or nighttime conditions, and species appearing small or partially in frame, all of which reduced performance on rare or difficult species. The release of the dataset and the proposed pipeline support future research and practical applications in tropical biodiversity monitoring. Index Terms—biodiversity monitoring, image classification, DeiT, Swin, EfficientViT, MegaDetector, camera traps, cemede redbioma-ct dataset, tropical ecosystem.