This study aims to achieve early and reliable monitoring of wood-boring pests, which are often highly concealed, have long lag times, and cause significant damage to forests. Specifically, the research focuses on the larval feeding vibration signal of the emerald ash borer as a representative pest. Given the crucial importance of such pest monitoring for the protection of forestry resources, developing a method that can accurately identify and interpret their vibration signals is paramount.
We introduce MelSPPNET, a self-explaining model designed to extract prototypes from input vibration signals and obtain the most representative audio segments as the basis for model recognition. The study collected feeding vibration signals of emerald ash borer larvae using detectors, along with typical outdoor noises. The design of MelSPPNET considers both model accuracy and interpretability.
Experimental results demonstrate that MelSPPNET compares favorably in accuracy with its similar non-interpretable counterparts, while providing interpretability that these networks lack. To evaluate the interpretability of the case-based self-explaining model, we designed an interpretability evaluation metric and proved that MelSPPNET exhibits good interpretability. This provides accurate and reliable technical support for the identification of emerald ash borer larvae.
While the work in this study is limited to one pest type, future experiments will focus on the applicability of this network in identifying other vibration signals. With further research and optimization, MelSPPNET has the potential to provide broader and deeper pest monitoring solutions for forestry resource protection. Additionally, this study demonstrates the potential of self-explaining models in the field of signal processing, offering new ideas and methods for addressing similar problems.