# Self-Supervised Pretraining and Controlled Augmentation Improve Rare Wildlife Recognition in UAV Images

**Authors:** Xiaochen Zheng, Benjamin Kellenberger, Rui Gong, Irena Hajnsek, Devis Tuia  
**arXiv:** [2108.07582](https://arxiv.org/abs/2108.07582)  
**Date:** 2021-08-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.07582v1/

## Key points

- Wildlife censuses are important for assessing living conditions and survival risks of wildlife species.
- Traditional manual surveys are being replaced with automatic counts from UAV imagery and deep learning models.
- Supervised models are typically pretrained on large-scale curated datasets before fine-tuning on target imagery.
- The authors propose a self-supervised pretraining methodology using controlled augmentation techniques combined with MoCo and CLD to improve rare wildlife recognition in UAV images.
- Their method outperforms conventional supervised models trained on ImageNet by a significant margin, even when reducing the number of training animals to just 10%.
- This effectively reduces the amount of required annotations while still enabling high-accuracy model training in highly challenging settings.
- The proposed self-supervised pretraining methodology is a promising approach for improving rare wildlife recognition in UAV images and could significantly reduce the amount of required training data while achieving high accuracy models for automated animal censuses using aerial imagery.

## Summary

Wildlife censuses are crucial for assessing living conditions and potential survival risks of wildlife species. Traditionally, these surveys were conducted manually, but recently they are being replaced with counts derived automatically from UAV imagery paired with deep learning models. These models are typically supervised and pretrained on large-scale curated datasets such as ImageNet or MS-COCO before fine-tuning on target imagery. The authors propose a self-supervised pretraining methodology that uses controlled augmentation techniques combined with MoCo and CLD to improve rare wildlife recognition in UAV images. They evaluate their approach on Kuzikus dataset, which contains aerial images captured over a wildlife reserve in Namibia. Their results show that their method outperforms conventional supervised models trained on ImageNet by a significant margin. Moreover, they demonstrate that their approach still yields favorable results even when reducing the number of training animals to just 10%. This effectively reduces the number of required annotations while still enabling high-accuracy model training in highly challenging settings. In conclusion, the proposed self-supervised pretraining methodology with controlled augmentation techniques and MoCo and CLD is a promising approach for improving rare wildlife recognition in UAV images. The results suggest that this method could significantly reduce the amount of required training data while still achieving high-accuracy models for automated animal censuses using aerial imagery.