2D convolutional neural network for the detection of Asian Koel (Eudynamys Scolopaceus) vocalizations

Abstract

Acoustic activity detection plays a vital role for automatic wildlife monitoring which includes the study of ecology, populations and habitats assessments. Birds are one of the few wildlife species to be monitored as their population and distribution are expected to change due to climate changes in order to conserve the ecosystem, diversity and seasonal population changes. Monitoring animals based on sound (bioacoustics) monitoring involves continuous observation to capture rare events. Several existing bird sound classification devices records sounds at point reading and processed the data off-line that involves complex Convolution Neural Network (CNN) architecture which takes longer time in the processing stages as data needs to be acquired before being processed. This approach is not applicable on the real-time monitoring. Therefore, this project investigates the best architecture that can be implemented to lower the complexity of algorithm for a bird sound classification. Data training with bird sound from all over the world and non-bird sounds will be done in optimizing the algorithm. In precise, this project focuses on a bird sound classification with low resource CNN to classify an Eudynamys Scolopaceus bird species. The bird sound detection will be assessed on the Xeno-Canto dataset which is a dataset containing bird vocalization samples and Urban8k that is shared openly are used for training and testing. Data segmentation is done on each of the samples with 16kHz sampling frequency of 25% overlapping to avoid data loss. Segmented samples are then converted into spectrograms and fed into MobileNet CNN and Bulbul CNN Architecture for training and testing. A set of testing samples were used to predict the accuracy of each model and prediction results were presented in a confusion matrix. Results from both comparisons showed that MobileNet has a higher accuracy of 80% than Bulbul CNN with 64%. Further development and optimization of model architecture with the use of more training samples can be done in the future towards achieving a higher accuracy in classifying the bird sound.