A Citizen Science Based Approach to Automated Taxon Identification Using Computational Bioacoustics and Unsupervised Feature Learning

背景：该研究项目将研究自动分类单元识别（ATI）算法和方法设计的新方法，即物种的自动识别。ATI的应用领域很多，从快速评估生物多样性和监测生境健康和变化到早期识别农业害虫和克服分类障碍（解决缺乏训练有素的分类学家的问题）。ATI系统的传感器包括图像、生物声信号、雷达和声纳。该项目将集中研究生物声学信号，如唱歌的昆虫、鸟类，可能还有哺乳动物。开发一个成功的、健壮的和可扩展的ATI系统将在诸如生物多样性评估、栖息地质量监测、环境规划和公民科学等领域产生重大影响。目的和目标：a)为一些生物声学活跃的分类群开发光谱和时间特征提取方法。b)研究利用提取的特征集实现分类群最佳分离的新型人工智能技术。这将包括将人工神经网络与专家系统相结合，以增强和灵活的识别能力。该系统必须能够在高水平声干扰的自然环境中运行。c)评估为歌唱昆虫（包括蝉和直翅目昆虫（蚱蜢和蟋蟀））开发的方法。d)在智能手机上实施成功的算法，以创建和测试以公民科学为导向的物种识别方法。新颖性：该项目的新颖性体现在以下几个方面：(i)时变信号的新特征集，例如多尺度时域信号编码（MTDSC）；（ii）将更传统的人工神经网络（MLP，循环网络）与专家系统相结合，以封装非参数数据，如生物地理信息和物候；（iii）发展可扩展的架构；（iv）应用于智能手机，以市民科学为基础的应用。与EPSRC研究领域：人工智能技术保持一致-该项目将研究无监督机器学习技术与专家系统方法相结合的使用，以创建可扩展和鲁棒的生物声学信号识别。数字信号处理-项目的成功将强烈依赖于从生物声学信号中提取特征，这将需要DSP技术。音乐和声学技术-该项目将集中于生物起源的声学信号，分析和识别。涉及的公司和合作者：没有公司参与；我们可以通过林业委员会、野生动物信托基金和自然英格兰进入野外。我们希望随着项目的进展，能产生更多的合作。

Context:The research project will investigate novel approaches to the design of algorithms and methods for automated taxon identification (ATI), i.e. the automated identification of species. The application areas for ATI are numerous, ranging from rapid assessment of biodiversity and monitoring of habitat health and changes to early identification of agricultural pests and overcoming the taxonomic impediment (addressing the lack of trained taxonomists). The sensors for ATI system can include images, bioacoustic signals, radar and sonar. This project will concentrate on bioacoustic signals such as singing insects, birds and possibly mammals. Development of a successful robust and scalable ATI system will have significant impact in fields such as biodiversity assessment, habitat quality monitoring, environmental planning and citizen science.Aims and Objectives:a) To develop spectral and temporal feature extraction methods for a number of bioacoustically active taxa.b) To investigate novel AI techniques for optimal separation of taxa using extracted feature sets. This will include combining artificial neural networks with expert systems for enhanced and flexible recognition capability. The systems must be capable of operating in natural environments with high levels of acoustic interference.c) To evaluate the methods developed for singing insects including cicadas and Orthoptera (grasshoppers and crickets).d) To implement successful algorithms on smartphones to create and test citizen science oriented approaches to species identification.Novelty:Novel aspects of the project are in several areas: (i) new feature sets for time varying signals, e.g. multiscale time domain signal coding (MTDSC); (ii) combining more traditional artificial neural networks (MLP, recurrent nets) with expert systems to encapsulate non-parametric data such as biogeographical information and phenology; (iii) develop scalable architectures; (iv) application to smart phones for citizen science based applications.Alignment with EPSRC Research Areas:Artificial Intelligence Technologies - the project will investigate the use of unsupervised machine learning techniques combined with expert system approaches to create scalable and robust bioacoustic signal recognition.Digital Signal Processing - the success of the project will strongly rely on the extraction of features from the bioacoustic signals, this will require DSP techniques.Music and Acoustic Technology - the project will concentrate on acoustic signals of biological origin, their analysis and identification.Companies and Collaborators Involved:There are no companies involved; we have access to field sites via Forestry Commission, Wildlife Trusts and Natural England. We hope to engender more collaboration as the project progresses.