Varela-Jaramillo Andrea, Winkelmann Christian, Mármol-Guijarro Andrés, Guayasamin Juan M, Rivas-Torres Gonzalo, Steinfartz Sebastian, MacLeod Amy
Institute of Biology, Molecular Evolution and Systematics of Animals, University of Leipzig, Leipzig, Saxony, Germany.
3Diversity, Quito, Pichincha, Ecuador.
Sci Rep. 2025 Jul 24;15(1):26884. doi: 10.1038/s41598-025-08381-9.
Population surveys are essential for conservation, but are often resource-intensive. Modern technologies, like drones, facilitate data collection but increase the analysis burden. Citizen Science (CS) offers a solution by engaging non-specialists in data analysis. We evaluated CS for monitoring marine iguanas, focusing on volunteers' accuracy in detecting and counting individuals in aerial images. During three phases of our Zooniverse project, over 13,000 volunteers contributed 1,375,201 classifications from 57,838 images; each classified up to 30 times. Using a Gold Standard dataset of expert counts from 4,345 images, we evaluated optimal aggregation methods for CS-inputs. Volunteers achieved 68-94% accuracy in detection, with more false negatives than false positives. The standard 'majority vote' aggregation approach (where the answer given by the majority of individual inputs is selected) produced less accuracy than when a minimum threshold of five volunteers (from the total independent classifications) was used. Image quality significantly influenced accuracy; by excluding suboptimal pilot-phase data, volunteer counts were 91-92% accurate. HDBSCAN clustering yielded the best results. We conclude that volunteers can accurately identify and count marine iguanas from drone images, though there is a tendency for undercounting. However, even CS-based data analysis remains relatively resource-intensive, underscoring the need to develop an automated approach.
种群调查对于保护工作至关重要，但往往资源消耗大。无人机等现代技术有助于数据收集，但增加了分析负担。公民科学（CS）通过让非专业人员参与数据分析提供了一种解决方案。我们评估了公民科学用于监测海鬣蜥的情况，重点关注志愿者在航空图像中检测和计数个体的准确性。在我们的“众包星系”项目的三个阶段中，超过13000名志愿者对57838张图像进行了1375201次分类；每人最多分类30次。我们使用来自4345张图像的专家计数的黄金标准数据集，评估了公民科学输入数据的最佳汇总方法。志愿者在检测中的准确率达到68%-94%，假阴性比假阳性更多。标准的“多数投票”汇总方法（即选择大多数个体输入给出的答案）产生的准确率低于使用五名志愿者（从总独立分类中）的最小阈值时的准确率。图像质量显著影响准确率；通过排除次优的试点阶段数据，志愿者计数的准确率为91%-92%。HDBSCAN聚类产生了最佳结果。我们得出结论，志愿者可以从无人机图像中准确识别和计数海鬣蜥，尽管存在计数不足的趋势。然而，即使是基于公民科学的数据分析仍然相对资源密集，这突出了开发自动化方法的必要性。
