부산대학교 도서관

Bird sound classification is an important task in the implementation of automatic farm monitoring systems, as many birds are harmful threats to crop farms, especially maize, sorghum, and rice farms. The ability to automatically classify and detect bird pests can improve the monitoring system for farmers, by detecting and notifying them in case of the presence of bird pests on the farm, in near real-time for their quick decision-making and damage minimization by taking action to scare birds away. There are many studies on the classification of bird sounds. However, there are few studies on bird sound classification in the context of farm monitoring. In this paper, we present a method for bird sound classification based on grey-level co-occurrence matrix (GLCM) features. GLCM is a statistical approach that measures the texture properties of the image. In the context of sound classification, sound waves can be represented as spectrogram images, which can then be processed using GLCM features. To measure the texture of spectrograms, four GLCM features namely contrast, correlation, energy, and dissimilarity were calculated and used. The feature vectors obtained from the sound signals of two bird species, namely nightingale (Luscinia megarhynchos) and mousebird (Colius striatus), were utilized as inputs for classification using two different algorithms: Light Gradient Boosting Machine (LightGBM) and support vector machines (SVM). These two bird species are known to cause damage to crops in Bugesera, the eastern province of Rwanda. The LightGBM classifier exhibited promising results, achieving an accuracy of 90.6%, surpassing the 85% accuracy obtained with SVM in classifying nightingale and mousebird sounds. The superior performance of the LightGBM model suggests its potential usefulness in farm monitoring tasks.