云南高原湿地生态环境音分类技术研究

The ecological environment audio signals are closely related to living environment of creatures and human beings. Research of classification and recognition on the ecological environment of audio data, for the ecological environment security and the monitoring, the analysis and study on the ecological environment and predicting the change of the ecological environment, has important significance. Since Yunnan attaches great importance to the protection of wetland resources and related research, the classification of ecological environment audio signals of wetland will provide a new perspective for this field. Relying on Yunnan plateau wetland resources, the project will carry on research including the technologies and methods of acquisition of wetland ecological environmental audio signals, the extracting of the stationary and non-stationary features of environmental audio signals，as well as multi-type features optimization ,selection and fusion technology. In order to ,at the backgroud of decreasing the numbers of training labeled samples,effectively improve the accuracy of classification of environmental audio signals and the generalization of classification model,the project will extensivelly explore the technologies of classification.That is how to effectively build the ensemble of the semi-supervised learning and active learning classification model.In additional, based on the many kinds of features extracted from wetland environmental audio signals, the classification model with multiple views and multiple classifiers will be researched. The classification result of wetland enviromantal audio signals will serve for establishing database of wetland ecological environment sound. The outcomes of this project is to provide the effective classifications of wetland ecological environment audio signals in Yunnan to extand our understanding the connection between the environment sound and the ecological environment change, to improve the wetland ecological environment monitoring, to help us give effective forecast the change of the wetland ecological environment.

生态环境音频与各种生物的生存环境、人类的生活环境密切相关。研究生态环境音频数据分类、识别技术，对于生态环境安全与监控、分析和研究生态环境、预测生态环境的变化都具有重要意义。云南高度重视对湿地资源的监测、保护与研究，而湿地环境音频数据的分类识别将为该地区的湿地生态环境的研究提供新的视角。 项目将依托云南高原丰富多样湿地资源，研究湿地生态环境的音频信号采集，信号平稳特征和非平稳特征的提取方法，多类型特征优化选择与融合技术；探讨半监督学习和主动学习集成的分类模型，以及基于多类音频数据信号特征构造多视图下的多分类模型；研究在训练样本较少的情况下，有效地提高环境音分类的精度和分类模型的泛化性。 通过项目研究，完成对选择湿地区域采集的环境音分类识别，初步建立云南高原湿地生态环境音的数据库。为探索云南湿地生态环境声音与生态环境变化之间的联系，湿地生态环境监测，预测生态环境变化提供有效音频分类信息。

生态环境音频与各种生物的生存环境、人类的生活环境密切相都关。研究生态环境音频数据分类识别技术，对于生态环境安全与监控、分析和研究生态环境、预测生态环境的变化都具有重要意义。湿地环境音频数据的分类识别为云南地区的湿地生态环境的监测和保护研究提供新的视角。. 项目主要以云南高原湿地生态环境音为研究对象，自主研究设计了野外湿地环境音频信号采集设备；针对环境音频信号平稳和非平稳特征，提取了MFCC和CELP等系列特征的基础上，研究了短时傅里叶变换STFT声谱、小波变换声谱和EMD分解的稀尔波特-黄变换声谱分析；研究单一和多评价度量属性的分类最佳特征选择方法；基于多样性和差异性度量的选择集成分类和子特征空间融合集成；针对训练样本少的问题，研究了在熵优先采样EPS方法的基础上，提出基于支持向量机SVM_EPS和多视图的MV_EPS的主动学习方法，建立分类模型，将研究的动态多分类器选择方法，选择性集成与半监督融合的AD M-training方法，有效提高了环境音在少训练样本下的分类精度；基于上述研究成果搭建完成“云南高原湿地环境音”YNW_ESR数据库，包括音频信号库，特征库和模型库。同时利用项目研究成果创建了网站资源平台“林智信息处理组—云南高原湿地生态环境音”，该平台提供云南湿地环境音资源的共享，实现检索湿地环境音信息，特别是湿地鸟类的信息，平台可下载音频信号文件、特征及声谱图信息，并提供上传音频文件进行在线识别服务。.项目的研究，完成了云南湿地环境音采集、特征提取和分类识别，建立了云南高原湿地生态环境音数据库。为深入研究云南湿地生态特性提供了音频分析识别的研究方式，为湿地生态环境监测及预测生态环境变化提供有效的音频分类信息。