Asynchronous Acoustic–Image Classification for Unattended Ground Sensors with Domain Adaptation and Bidirectional Prototype Distillation

Acoustic and image sensors in unattended ground sensor systems often operate at different sampling rates and use different trigger mechanisms and fields of view, so their observations are asynchronous and lack instance-level correspondence. In addition, web images used to supplement scarce field annotations exhibit a substantial domain shift from target-domain images. This paper proposes a three-stream joint learning framework that integrates reliability-weighted image-domain adaptation, reliability-aware dynamic bidirectional prototype distillation, and balanced complementary modality dropout. A lightweight shared backbone processes labeled acoustic spectra, labeled mixed-source images, and unlabeled target-domain images. A main classifier and two auxiliary classifiers estimate target pseudo-label confidence and agreement. Their consensus is used to construct reliability weights for marginal alignment, while confidence guides target-sample selection and weighting for class-conditional alignment. Acoustic and image class prototypes transfer category-level knowledge without constructing artificial instance pairs, while direction-specific distillation weights are dynamically assigned according to classification reliability, prototype availability, and prototype stability. Across three independent runs with different random seeds, experiments on datasets comprising 148,038 acoustic spectra and 167,754 images yield acoustic, image, and overall accuracies of 91.97±0.08%, 95.16±0.14%, and 93.72±0.07%, respectively, exceeding the highest corresponding baseline mean accuracies by 1.52, 1.30, and 2.31 percentage points. The deployment model contains 123.40 K parameters and requires 0.47 MiB in FP32, making it a compact classification model with potential for deployment in resource-constrained field environments.