Research On Environmental Sound Recognition In The Field Of Urban Ecolog

Assessing the impact of human activities on biodiversity through the analysis of environmental soundscape characteristics is a cutting-edge research topic in urban ecology.By identifying the ecological sound and studying the correlation between different categories of sound,it is possible to better analyze the impact of human activities on biodiversity.Applying deep learning methods to classify urban ecological sound can effectively improve work efficiency,but there are still problems such as a lack of training data,model selection difficulty,and implementation challenges.In this paper,the identification of seven types of environmental sounds,including human-sound,insect-sound,bird-sound,bird-human sound,insect-human sound,bird-insect sound,and silent segment,were studied as an example of urban ecological sound recognition.The main research contents are as follows:(1)A method for assisting sample selection was proposed to address the problem of manual sample annotation when creating a training dataset.First,Mel Frequency Cepstral Coefficients(MFCC)and Inverted Mel Frequency Cepstral Coefficients(IMFCC)were used to fuse features to obtain new features,MFCC＿IMFCC.Then,the effectiveness of this feature in environmental sound classification was verified by combining the K-Nearest Neighbors(KNN),Random Forest(RF),and Support Vector Machine(SVM)algorithms,respectively.Finally,the PCA algorithm was used for feature dimensionality reduction to reduce feature redundancy and improve the efficiency of the algorithm.Experimental results showed that when SVM was combined with the PCA algorithm,with only 100 training samples per class,the average precision,recall rate,and F1 score of the seven types of environmental sound were 72.00%,71.54%,and 71.61%,respectively,and the training and prediction times were 439 ms and6187ms,respectively.This algorithm can effectively assist sample selection and annotation work.(2)Through studying the impact of human activities on biodiversity using acoustic information,it is necessary to efficiently and accurately classify a large amount of highprecision environmental sound collected.However,there is currently no deep learning solution specifically designed for this application.To address this problem,based on recorded sound data from the urban forest in Guangzhou,a deep learning model called Eco Env Net was designed and constructed to accurately identify environmental sound.The performance of this model was compared with classic network models such as Res Net,Mobile Net,and Efficient Net.Experimental results showed that the proposed Eco Env Net had the best overall performance among all the compared models,with the accuracy of 93.81% for identifying seven types of environmental sound.This work provides practical experience for the application of deep learning technology in ecological research and offers new technical references for further exploring the relationship between human activities and biodiversity.(3)As researchers in the ecological field lack sufficient time and energy to keep up with the constantly evolving algorithm models and implement code when classifying environmental sounds,it negatively affects the efficiency of research work.In order to facilitate the use of deep learning technology for environmental sound classification,this paper designs and implements an integrated graphical interface platform for environmental sound recognition,which includes functions for dataset management,model management,and application management,and integrates the above-mentioned data filtering tool and Eco Env Net model.In this platform,the data management module allows users to create and upload their own sound datasets for subsequent model training and data classification;the model management module allows users to train classification models according to their needs;and the application management module allows users to create prediction tasks based on trained models and predict the real data.Through testing the environmental sound data collected from the Daifushan area in Guangzhou,it was found that the platform can run stably and has high robustness.