SUMMARY
Forest has an important role on earth. The need to monitor forest from illegal activities and the types of animals in there is needed to keep the forest in good condition. However, the condition of the vast forest and limited resource make direct forest monitoring by officer (human) is limited. In this case, sound with digital signal processing can be used as a tool for forest monitoring. In this study, a system was implemented to classify sound on the Raspberry Pi 3B+ using mel-spectrogram. Sounds that classified are the sound of chainsaw, gunshot, and the sound of 8 species of bird. This study also compared pretrained VGG-16 and MobileNetV3 as feature extractor, and several classification methods, namely Random Forest, SVM, KNN, and MLP. To vary and increase the number of training data, we used several types of data augmentation, namely add noise, time stretch, time shift, and pitch shift. Based on the result of this study, it was found that the MobileNetV3-Small + MLP model with combined training data from time stretch and time shift augmentation provide the best performance to be implemented in this system, with an inference duration of 0.8 seconds; 93.96% accuracy; and 94.1% precision.