Detecting 'poachers' with drones: Factors influencing the probability of detection with TIR and RGB imaging in miombo woodlands, Tanzania
Detecting 'poachers' with drones: Factors influencing the probability of detection with TIR and RGB imaging in miombo woodlands, Tanzania
复制标题
DOI：
10.1016/j.biocon.2019.02.017
复制
发表时间：
2019-05-01
影响因子：
5.9
通讯作者：
Wich, Serge A.
中科院分区：
文献类型：
--
作者：
Hambrecht, Leonard;Brown, Richard P.;Wich, Serge A.
保护生物学家越来越多地使用无人机来减少偷猎动物。然而，没有关于发现偷猎者的概率和影响发现的因素的已发表研究。在自愿受试者的实验环境中，我们评估了各种因素对偷猎者检测概率的影响：相机（可见光谱：RGB和热红外：TIR），树冠覆盖的密度，受试者与图像中心线的距离，受试者与背景的对比度，无人机的高度和图像分析师。我们手动分析了录像并标记了所有记录的主题检测。多层次模型被用来分析TIR图像数据和一般的线性模型方法被用于RGB图像数据。我们发现TIR相机比RGB相机具有更高的检测概率。在TIR图像的检测概率显着影响郁闭度，主题距离中心线和分析。在RGB图像的检测概率显着影响郁闭度，主题对比度对背景，海拔高度和分析。总的来说，我们的研究结果表明，TIR相机提高了人类的探测能力，特别是在一天中较凉爽的时候，但这受到厚厚的植被覆盖的严重阻碍。通过增加图像之间的重叠可以改善随着距图像中心线的距离增加而减少的检测的效果，尽管这需要在特定区域上进行更多的飞行。分析师的经验也有助于提高检测概率，但随着使用机器学习的自动检测的发展，这可能不再是一个问题。
Conservation biologists increasingly employ drones to reduce poaching of animals. However, there are no published studies on the probability of detecting poachers and the factors influencing detection. In an experimental setting with voluntary subjects, we evaluated the influence of various factors on poacher detection probability: camera (visual spectrum: RGB and thermal infrared: TIR), density of canopy cover, subject distance from the image centreline, subject contrast against the background, altitude of the drone and image analyst. We manually analysed the footage and marked all recorded subject detections. A multilevel model was used to analyse the TIR image data and a general linear model approach was used for the RGB image data. We found that the TIR camera had a higher detection probability than the RGB camera. Detection probability in TIR images was significantly influenced by canopy density, subject distance from the centreline and the analyst. Detection probability in RGB images was significantly influenced by canopy density, subject contrast against the background, altitude and the analyst. Overall, our findings indicate that TIR cameras improve human detection, particularly at cooler times of the day, but this is significantly hampered by thick vegetation cover. The effects of diminished detection with increased distance from the image centreline can be improved by increasing the overlap between images although this requires more flights over a specific area. Analyst experience also contributed to increased detection probability, but this might cease being a problem following the development of automated detection using machine learning.