Acoustic-Hyperspectral Multimodal Generation and Fusion for Cotton Drought Stress Detection Based on UAV Remote Sensing
TL;DR
A deep detection framework that integrates acoustic emission signals with hyperspectral images to address the limitations of single-modal characterization and susceptibility to noise interference in detecting drought stress in cotton is proposed.
Abstract
With the growing demand for early monitoring of crop moisture in precision agriculture continues to grow, this article proposes a deep detection framework that integrates acoustic emission signals with hyperspectral images to address the limitations of single-modal characterization and susceptibility to noise interference in detecting drought stress in cotton. The study first constructs a multiscale time–frequency branch and a regional–global spectral branch to capture, respectively, the transient pulse characteristics induced by cavitation and the long-range dependencies of the canopy spectrum. Subsequently, a dual second-order attention module is introduced to enhance edge information and spatial high-frequency details through the synergistic use of channel gradients and structural tensors. Finally, a deep adaptive fusion mechanism is designed to achieve dynamic weighted allocation of bimodal features. Experiments show that the model achieves an overall accuracy of 97.01%, significantly outperforming the baseline models, even in the more challenging task of early detection of mild drought, the model maintains a high accuracy of 95.47%. In summary, the method proposed in this article provides a highly robust technical approach for the nondestructive diagnosis of crop water deficit.