Wild animal surveys play a critical role in wild animal conservation and ecosystem management. Unmanned aircraft systems (UASs), with advantages in safety, convenience and inexpensiveness, have been increasingly used in wild animal surveys. However, manually reviewing wild animals from thousands of images generated by UASs is tedious and inefficient. To support wild animal detection in UAS images, researchers have developed various automatic and semiautomatic algorithms. Among these algorithms, deep learning techniques achieve outstanding performances in wild animal detection, but have some practical issues (e.g., limited animal pixels and sparse animal samples). Based on a typical deep learning pipeline, faster region based convolutional neural networks (Faster R-CNN), this study adopted several tactics, including feature stride shortening, anchor size optimization, and hard negative class, to overcome the practical issues in wild animal detection in UAS images. In this study, a kiang survey was conducted in UAS datasets (23,748 images) obtained by 14 flight campaigns in the eastern Tibetan Plateau. The validation experiments of our adopted tactics revealed the following: (1) feature stride shortening and anchor size optimization improved small animal detection performance in the animal patch set, increasing the F1 score from 0.84 to 0.86 and from 0.86 to 0.92, respectively; and (2) the hard negative class significantly suppressed false positives in the full UAS image set, increasing the F1 score from 0.44 to 0.86. The test results in the full UAS image set showed that the modified model with the adopted tactics can be applied to either a semiautomatic survey to accelerate manual verification by 25 times or an automatic survey with an F1 score of approximately 0.90. This study demonstrates that the combination of UAS and deep learning techniques can enable automatic/semiautomatic, accurate, inexpensive, and efficient wild animal surveys.

野生动物调查在野生动物保护和生态系统管理中发挥着至关重要的作用。在安全性，便利性和廉价性方面具有优势的无人机系统（UAS）已越来越多地用于野生动物调查中。但是，从UAS生成的数千张图像中手动检查野生动物既繁琐又效率低下。为了支持UAS图像中的野生动物检测，研究人员开发了各种自动和半自动算法。在这些算法中，深度学习技术在野生动物检测中表现出色，但存在一些实际问题（例如，有限的动物像素和稀疏的动物样本）。基于典型的深度学习管道，基于更快区域的卷积神经网络（Faster R-CNN），本研究采用了几种策略，包括特征步幅缩短，锚大小优化和硬负类，以克服UAS图像中野生动物检测中的实际问题。在这项研究中，对由青藏高原东部的14次飞行运动获得的UAS数据集（23,748张图像）进行了一次江河调查。我们采用的策略的验证实验表明：（1）特征步幅的缩短和锚点尺寸的优化提高了动物补丁集中小动物的检测性能，将F1分数分别从0.84提高到0.86和从0.86提高到0.92；（2）硬否定类别在整个UAS图像集中显着抑制了误报，将F1分数从0.44提高到0.86。完整的UAS图像集中的测试结果表明，采用所采用策略的修改后的模型可以应用于半自动调查以将手动验证速度提高25倍，也可以应用于F1分数约为0.90的自动调查。这项研究表明，UAS和深度学习技术的结合可以实现自动/半自动，准确，廉价和高效的野生动物调查。