Sound-spectrogram based automatic bird species recognition using MLP classifier

Abstract

Bioacoustics play a major role in the field of ornithology, ecology, animal behaviour study, habitat monitoring, species conservation and design of deterrent system. This work focusses on the design and implementation and performance evaluation of an automatic, more efficient and flexible bird sound based recognition system for classifying eight species of popular Eurasian birds the standard online annotated databases. A dedicated virtual instrument tool with effective GUI is developed that acquires, pre-processes the sound samples and generates statistically evaluated short term Fourier transform spectrogram based feature matrix, suited for characterization of vocalization patterns of bird species. Using the well-labelled feature data of sound records, multi-layer perceptron artificial neural network (MLP-NN) classifier model is designed, trained, tested and optimized using feedforward-backpropagation supervised learning algorithm. Various experiments, following a systematic approach, are conducted to optimize the structure of MLP with respect of number of neurons in the hidden layer, epochs and learning rate for attaining enhanced recognition accuracy (96.1%), recall (82.6%) and precision (84.5%). The performance of the optimal model is also analysed in terms of recognition capabilities of individual bird species that indicate promising results. A few of the birds are recognized accurately and precisely when present as compared to the others. The tendency of the model to wrongly identify or miss bird species also remained low. The model performance over the unseen dataset also remained satisfactory with cross-validation classification accuracy of 81.4%. The system being scalable, can easily be reused in future to retrain the model over the large set of sound samples from real world recordings with improved acoustic features for achieving very high classification accuracy and reliability.