Detecting presence and inferring absence are both critical in species monitoring and management. False-negatives in any survey methodology can have significant consequences when conservation decisions are based on incomplete results. Marsh birds are notoriously difficult to detect, and current survey methods rely on traditional labor-intensive methods, and, more recently, passive acoustic monitoring. We investigated the efficiency of passive acoustic monitoring as a survey tool for the cryptic and poorly understood Eastern Black Rail (Laterallus jamaicensis jamaicensis) analyzing data from two sites collected at the Tom Yawkey Wildlife Center, South Carolina, USA. We demonstrate two new techniques to automate the reviewing and analysis of long-duration acoustic monitoring data. First, we used long-duration false-color spectrograms to visualize the 20 days of recording and to confirm presence of Black Rail "kickee-doo" calls. Second, we used a machine learning model (Random Forest in regression mode) to automate the scanning of 480 consecutive hours of acoustic recording and to investigate spatial and temporal presence. Detection of the Black Rail call was confirmed in the long-duration false-color spectrogram and the call recognizer correctly predicted Black Rail in 91% of the first 316 top-ranked predictions at one site. From ten days of continuous acoustic recordings, Black Rail calls were detected on only four consecutive days. Long-duration false-color spectrograms were effective for detecting Black Rail calls because their tendency to vocalize over consecutive minutes leaves a visible trace in the spectrogram. The call recognizer performed effectively when the Black Rail call was the dominant acoustic activity in its frequency band. We demonstrate that combining false-color spectrograms with a machine-learned recognizer creates a more efficient monitoring tool than a stand-alone species-specific call recognizer, with particular utility for species whose vocalization patterns and occurrence are unpredictable or unknown.

中文翻译：

---

使用半自动分析长时间录音来进行东部黑轨检测

在物种监测和管理中，检测存在和推断出缺失都是至关重要的。当保护决策基于不完整的结果时，任何调查方法中的假阴性都可能产生重大后果。众所周知，沼泽鸟很难被发现，目前的调查方法依赖于传统的劳动密集型方法，并且最近还依赖于被动声监测。我们调查了被动声学监测的效率，该方法是一种针对神秘的，了解程度不高的东部黑铁路（Laterallus jamaicensis jamaicensis）的调查工具，用于分析在美国南卡罗来纳州汤姆·扬基野生动物中心收集的两个站点的数据。我们演示了两种自动化的技术，可以自动检查和分析长时间的声学监测数据。第一的，我们使用了长时间的假彩色光谱图来可视化记录的20天，并确认是否存在Black Rail“ kickee-doo”呼叫。其次，我们使用了机器学习模型（回归模式下的随机森林）来自动扫描480个连续小时的录音，并调查时空的存在。在长时间的假彩色光谱图中确认了对Black Rail呼叫的检测，并且呼叫识别器在一个站点的前316个排名最高的预测中有91％正确预测了Black Rail。从连续十天的录音中，仅连续四天检测到Black Rail通话。长时间的伪彩色声谱图可有效检测Black Rail通话，因为它们在连续几分钟内发声的趋势在声谱图中留下了明显的痕迹。当Black Rail呼叫是其频带中的主要声学活动时，呼叫识别器会有效执行。我们证明了，将伪彩色声谱图与机器学习的识别器相结合，可以创建比独立的特定物种的呼叫识别器更有效的监视工具，尤其适用于发声模式和发生情况不可预测或未知的物种。