Top

Improved PointNeXt for Acoustic Recognition of Stored-Grain Insects under Granary Monitoring Conditions

Authors

Files

PDF

Abstract

Acoustic monitoring provides a non-destructive route for detecting stored-grain insects, yet field granary recordings are usually affected by grain-bed attenuation, sensor coupling variation, fan noise, grain-settling interference, and short-duration biological impulses. This study proposes an improved PointNeXt model for insect acoustic recognition under practical granary monitoring data conditions. Candidate acoustic events are transformed into sparse time-frequency point clouds through ridge-preserving denoising, adaptive thresholding, local saliency estimation, and multi-scale point sampling. A residual acoustic PointNeXt backbone is modified with density-aware neighborhood aggregation, channel recalibration, and a lightweight temporal voting head to strengthen weak-event representation and suppress isolated mechanical disturbances. Experiments are organized on a simulated granary acoustic dataset containing five target classes and background events, with grain depths from 0.4 m to 2.0 m, 1.0 s windows with 50% overlap, and signal-to-noise ratios from -5 dB to 20 dB. The proposed model reaches 96.8% accuracy, 95.9% macro-F1, and a 2.8% false-alarm rate, improving macro-F1 by 4.7 percentage points over the original PointNeXt and by 7.6 percentage points over a ResNet spectrogram baseline. Under low-SNR recordings, the method maintains 92.6% macro-F1 and reduces missed detections of weak larval feeding bursts by 38.4%. The results indicate that point-cloud modeling of acoustic events can preserve sparse discriminative structures while limiting redundant spectral background, although broader field validation remains necessary before commercial deployment.

PDF

Details