We investigate the audio classification of five insect species and a diverse class of environmental sounds using short audio recordings with a low sampling rate. First, we use a large Convolutional Neural Network (CNN) that achieves over 93% accuracy. Building on this, we investigate the development of smaller, more efficient models tailored to resource-constrained devices.
 
By using different techniques, we develop a compact CNN with a model size of less than 2 MB that achieves an accuracy of almost 88%. By further reducing the model size to below 40 kB, we still achieve an accuracy of 80%. These lightweight models are designed for use on resource-constrained microcontrollers and enable non-invasive and scalable monitoring of insect populations in real-world environments.

Utilizing AI for the classification of insect sounds, particularly those of grasshoppers, is a promising method to monitor biodiversity non-invasively in the field. Our project focuses on the development and implementation of a sustainable TinyML model. This model aims to efficiently classify grasshopper species based on their songs in real-time on a microcontroller with minimal computational resources.