Application Of Convolutional Neural Network In Underwater Acoustic Signal Recognition And Birds Counting

The ecological protection and restoration of islands and coastal wetlands are inseparable from the support of various data,such as meteorological data,hydrological data,and biological data.The acoustic tomography inversion method calculates the temperature field and flow field of the sea area through the different of the acoustic signal propagation time between different underwater stations.In order to obtain the propagation time,the commonly extraction method is to use the correlation between the signals to identify the signal and extract the signal arrival time.However,this method requires the repeated operation of shifting the signal and the correlation calculation of the signals between different stations.The calculation process is tedious and inefficient and cannot match the needs of specific scenarios.Convolutional neural networks based on deep learning can effectively improve the computational efficiency of signal arrival time extraction.Convolutional neural network performs self-learning on the characteristics of different site signals,realizes multi-category recognition,reduces calculation steps,and effectively responds to the complex noise environment in the ocean.A well-designed model structure can further reduce computing costs,quickly extract signal arrival time and meet application scenarios that require strict computing time such as rapid verification and real-time display.This article processes and recognizes underwater acoustic signals in ocean acoustic tomography according to three main steps: dataset production,model design,and training.The accuracy of 1D U-net segmentation model and GAP Net classification model on the acoustic tomography experimental signal testset reached 87.65% and 86.26% respectively.Although it has practical application value to a certain extent,its accuracy is still insufficient compared with the relevant calculation results.The neural network model of supervised learning has a strong dependence on data,and transfer learning can effectively overcome the negative factors of too small or incomplete dataset in the task.As an application supplement to transfer learning and project needs,based on transfer learning,object detection and density estimation are used to count the bird targets on the island and well statistical results with an average counting errors of 6.70 and 15.11 have been obtained.