Convolutional Neural Network-based Bioacoustic Tarsier Detection for Wildlife Monitoring

Philippine Tarsiers are known to be sensitive especially to loud noises, physical contact, and bright lights; and with the decreasing population of the Philippine tarsiers, conservation methods and wildlife monitoring are necessary to alleviate this problem. One way to monitor them is by using their bioacoustic characteristics. Together with a machine learning model called Convolutional Neural Networks or CNN, a noninvasive method for monitoring them is possible. For this to happen, the training of the ML model is necessary. To do this, the researchers gathered an audio dataset of the Philippine Tarsiers, then the researchers converted the gathered data into mel spectrograms then annotated and pre-process the dataset in order for the CNN model to tell apart three types of sounds, namely: background noise, a single tarsier calls, and multiple tarsier call. There were a total of 3 runs made; and with each run, a different and updated fine-tuning method was used in order to match the needs of the model. The latest accuracy of the model reached up to 96.50 percent correctness. A Raspberry Pi 5 with 8 GB RAM was used together with the TensorFlow software in order to train the model. A shotgun microphone was also connected to the Raspberry Pi 5 together with a 20x4 I2C LCD screen for the feedback. With this, it is possible to design and run a non-invasive wildlife monitoring device for the tarsiers and can be used for future research in monitoring animal census and studying their behaviour.