Research On Monitoring Of Grazing Information For Sheep Based On Acoustic Signals

As the concept of precision livestock farming and grassland ecological conservation is strengthened,the free-range livestock farming industry is eager to rely on advanced technologies for the scientific management of livestock and pastures.Monitoring essential information of grazing sheep,such as real-time location,grazing behavior(biting,chewing,chewing-biting),grazing activity(foraging,ruminating,resting),growth condition of foraged grass(very dense,relatively dense,sparse,hereafter referred to as grass condition)and foraging intake,is the primary data support to realize the scientific management.At present,the monitoring methods of real-time location and grazing activity are relatively mature.However,the studies related to grazing behavior,grass condition,and foraging intake are not perfect enough,and it is necessary to conduct more systematic and in-depth research.Among various methods for monitoring grazing information,the acoustic method has become one of the most promising methods due to its advantages of non-interference with livestock life and comprehensive functions.Therefore,the following studies were carried out based on acoustic method with grazing sheep as the research object:(1)Identification of grazing behavior.The purpose of grazing behavior identification is to mark the acoustic signal fragment(hereafter referred to as the event)generated by grazing behavior in the collected acoustic signal for further classification.In this study,using feature extraction techniques and parameter optimization methods,an excellent recognition algorithm was constructed based on the collected acoustic signals from three typical weight sheep foraging multiple types of grass and ruminating at night.The results showed that the algorithm's accuracy was 96.16%,which was suitable for typical sheep foraging a variety of grass and had comprehensive functions and high precision.(2)Classification of grazing behavior.The purpose of grazing behavior classification is to classify the identified events as biting,foraging-chewing,chewing-biting,or ruminating-chewing.The best classification model was derived by comparing the performance of the typical deep neural network(DNN),convolutional neural network(CNN),and recurrent neural network(RNN)models.The results showed that the RNN model had the highest classification accuracy of 95.61%,CNN was slightly lower at95.03%,and DNN was 84.04%.The improved RNN model in this study had a stronger classification ability compared with the typical CNN model.(3)Classification of grass condition.Acoustic signals of sheep grazing were collected under three typical grass conditions.According to the grass condition types,the acoustic signals were divided into several samples,and six kinds of form samples were constructed from each sample.The log-Mel features of the form samples were fed to the CNN model or RNN model to identify the grass conditions.In the preprocessing of the form samples,this study compared two methods of uniform sample length and examined the effects of different sample uniform lengths on model accuracy.The results showed that FCB samples(fixed-length chewing and biting sequentially connected)combined with the RNN model had the highest accuracy(90.24%).The method of filling or truncating the sample's waveform to unify the sample length was more accessible to obtain the model with high accuracy than scaling the sample log-Mel image.The uniform lengths of samples had a significant effect on the model accuracy.(4)Forage intake estimation.Twenty-nine sets of trials were executed with sheep of three weight levels,two grass species,and three levels of moisture content combined.Segment samples were constructed from every trial,and many explanatory variables(44in total)related to chewing were extracted from each segment sample.The statistical analysis method was used to reveal the effects of sheep weight class,grass moisture content class,and grass type on the slope of the explanatory variables.Intake was predicted using all explanatory variables,all explanatory variables&factor variables(sheep weight,grass moisture content,grass NDF content,grass ADF content),respectively.The results showed that the influence of grass moisture content class on the slope of the variable was approximately equal to the sheep weight class and less than the individual sheep differences.Furthermore,adding factor variables increased the R~2 score of the intake estimation model substantially and reduced the number of variables used.The R~2 score of the model was further increased by introducing sheep weight into the intake,and the number of variables used was further reduced.When forage intake was FMI?W(fresh matter multiplied by the square of the sheep bodyweight value),the intake could be accurately predicted with an R~2 score of 0.9727 using all explanatory&factor variables.