AUTHOR=Wu Yaqin , Cheng Lijun , Hao Wangli , Niu Jianjun , Li Yuze , Chang Yan , Lv Jia , Li Xuru 

TITLE=Toward non-invasive early pest surveillance: cross-modal adaptation using PLMS acoustic-visual representation and pre-trained transfer learning

JOURNAL=Frontiers in Plant Science

VOLUME=Volume 16 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/plant-science/articles/10.3389/fpls.2025.1610163

DOI=10.3389/fpls.2025.1610163

ISSN=1664-462X

ABSTRACT=Pest infestations pose significant threats to agricultural productivity and ecological balance, making early prevention crucial for effective management. Toward non-invasive early-stage pest surveillance, this study introduces a novel cross-modal adaptation paradigm, leveraging the comprehensive bioacoustic repository, InsectSound1000 database. Firstly, the methodology initiates with adaptive audio preprocessing, where raw signals are filtered using the low-pass filter to remove high-frequency interference, followed by the downsampling operation to prevent aliasing and reduce computational complexity. Secondly, Patch-level log-scale mel spectrum (PLMS) spectrograms are proposed to convert acoustic signals into visual representations, refining time-frequency patterns through patch-level hierarchical decomposition to capture low-frequency and localized spectral features. The logarithmic transformation further enhances subtle low-frequency insect sound characteristics, optimizing feature analysis and boosting model sensitivity and generalization. Next, the PLMS acoustic-visual spectrograms undergo data augmentation prior to being processed by the pre-trained You Only Look Once version 11(YOLOv11) model for deep transfer learning, facilitating the efficient extraction of high-level semantic features. Finally, we compare the proposed algorithm with traditional acoustic features and networks, investigating how to balance preserving the frequency content of the signal and meeting computational requirements through optimized downsampling. Experimental results demonstrate that the proposed method achieves an Accuracy@1 of 96.49%, a Macro-F1 score of 96.49%, and a Macro-AUC of 99.93% at the 2500Hz sampling rate, showcasing its superior performance. These findings indicate that cross-modal adaptation with PLMS spectrograms and YOLOv11-based transfer learning can significantly enhance pest sound detection, providing a robust framework for non-invasive, early-stage agricultural pest surveillance.