BIGDATA: Collaborative Research: IA: BirdVox: Automating Acoustic Monitoring of Migrating Bird Species

目前对自然环境的生物声学监测需要人工处理，从录音中提取信息内容。因此，人工处理造成了一个基本的瓶颈，在这个瓶颈中，数据收集远远超过了提取相关和所需信息的能力。自然环境下生物声学物种自动分类研究大致可分为两类：一类是从音频片段中识别出一组预定义的已知物种，另一类是从连续音频流中提取出事件的物种。这两种分类技术都有其特定的问题——许多用于区分预定义物种的数据是在“工作室”条件下记录的，不能扩展到自然条件下，而连续音频流的处理会产生许多误报。为了克服这些挑战，我们将采取多层次的方法：分析由100个晚上的10个记录单元的整晚记录组成的数据集。建立一个网络软件，让公民科学家参与识别航班呼叫，为我们提供一个庞大而广泛的模型训练数据集。开发新颖的卷积深度学习网络，非常适合分析复杂的听觉场景。可视化模式检测和分类飞行呼叫的空间和时间，以产生关于鸟类迁徙的新信息。将模型生成的声学数据与雷达、视频和直接可视化的公民科学数据集进行比较，以产生最全面的夜间鸟类迁徙数据。结合鸟类发声的领域知识，让公民科学家参与开发大型良好注释的训练数据集，并采用新颖的深度学习方法，将最终解决自然环境中声学信号的机器分类问题。

Current bioacoustic monitoring of natural environments requires processing by humans to extract information content from recordings. Thus human processing creates a fundamental bottleneck in which data collection far outpaces capabilities to extract relevant and desired information. Bioacoustic research on automatic species classification in natural environments can be broadly divided into two groups: distinguishing a predefined set of known species from audio clips and extracting species as events that occur in a continuous audio stream. Both classification techniques have their specific problems--many of the data used distinguishing predefined species are recorded under "studio" conditions and not extensible to natural conditions, while processing of continuous audio streams generate many false positives. To overcome these challenges we will take a multi-tiered approach: Analyzing a data set consisting of full-night recordings from 10 recording units over 100 nights. Building a web-enabled software to engage citizen scientists to identify the flight calls, providing us with a large and extensive model training dataset. Developing novel convolutional deep-learning networks, which are well suited for analysis of complex auditory scenes. Visualizing patterns detected and classified flight calls in space and time to produce novel information about the bird migration. Comparing model-generated acoustic data with radar, video, and direct visual citizen science datasets to produce the most comprehensive accounts of nocturnal bird migration possible. The combination of domain knowledge in bird vocalizations, engaging citizen scientists to allow development of large well annotated training datasets, and taking a novel deep-learning approach, will finally resolve the machine classification of acoustic signals in natural environments.