# Copy‐Paste Augmentation Improves Automatic Species Identification in Camera Trap Images

## TL;DR

This paper shows that creating synthetic images by pasting animal segments onto new backgrounds improves AI's ability to identify species in new locations.

## Contribution

The study introduces copy-paste augmentation as a novel method to enhance AI generalization in biodiversity monitoring.

## Key findings

- Copy-paste augmentation improved AI generalization by 8% in identifying species in new locations.
- The method showed promise for improving performance on underrepresented species in long-tailed datasets.

## Abstract

Effective conservation requires effective biodiversity monitoring. The pace of global biodiversity change far outstrips the ability of manual fieldwork to monitor it. Therefore, technological solutions, like camera traps, have emerged as a crucial way to meet biodiversity monitoring needs. Camera traps produce vast amounts of data, and so AI is increasingly used to label images with species identities. However, AI struggles to identify species from new locations that are not part of the training data (‘generalisation’). Resolving this is crucial for the promise of automated biodiversity monitoring to be realised. Here we use ‘copy‐paste’ augmentation to help resolve the generalisation challenge. Copy‐paste augmentation refers to isolating animal ‘segments’ from existing images and pasting the segments onto novel backgrounds to create new, synthetic images that are then used as part of the training data. Theoretically, this could make a model agnostic to backgrounds and therefore more able to generalise to unseen locations. While generation of synthetic images is commonly used as an augmentation method in other fields, such as medicine, it has not been used before in biodiversity science. We found that copy‐paste augmentation improved the ability of AI to identify species in new, unseen locations by 8% ± 2%. There was species‐level variation in improvement, but the vast majority of species benefited from the approach. We found mixed results when using copy‐paste augmentation on models trained with very small numbers of images (1–8 per species). Copy‐paste augmentation improves the ability of AI models to generalise to new, unseen locations. Our method also shows promise for resolving the challenge of long‐tailed camera trap data. AIs perform poorly on species in the ‘long tail’ of these distributions because there are very few images to train on. Copy‐paste augmentation can help rebalance datasets by adding synthetic images of underrepresented species. Overall, our results suggest a promising role for augmentation methods that generate new, synthetic images in biodiversity science. Ecologists and conservationists must move beyond simple augmentation methods, such as image transformations, if we are to resolve key challenges in species identification AI.

Effective biodiversity monitoring requires AI to process the vast data from camera traps, but AI struggles to generalise to new locations. This study explores ‘copy‐paste’ augmentation—creating synthetic images by pasting animal segments onto new backgrounds—to improve AI's species identification. Results show an 8% improvement in AI generalisation, particularly for underrepresented species, highlighting the potential of advanced augmentation methods in biodiversity science.