Acoustic Phenotyping: Harnessing Deep Learning for Early Detection of Hydric Stress in Crops

Climate change and water scarcity pose critical threats to global food security, necessitating the evolution of precision agriculture beyond traditional soil moisture sensors and satellite imaging. While current remote sensing techniques detect water stress after visible morphological changes (wilting) occur, they often fail in the critical "pre-symptomatic" phase. This paper introduces a novel, non-invasive diagnostic framework-Acoustic Phenotyping-that leverages the biophysical principle of acoustic emissions (AE) from xylem cavitation. By deploying high-frequency ultrasonic microphones and utilizing advanced Deep Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) architectures, we demonstrate that distinct acoustic signatures are emitted by plant stems under tension hours before visible wilting or canopy temperature changes occur. Our proposed model processes raw audio signals transformed into Mel-spectrograms, achieving a classification accuracy of ۹۴.۷% in distinguishing between three hydric states (Optimal, Moderate Stress, Severe Stress). This method offers a game-changing, real-time capability for automated irrigation systems, significantly reducing water consumption while preserving crop yield and quality. Furthermore, we discuss the primary engineering challenges, including ambient noise filtering and signal attenuation, providing viable solutions for field-scale deployment.

کلیدواژه ها:

Plant Bioacoustics ، Acoustic Emission ، Deep Learning ، Precision Irrigation ، Hydric Stress ، Convolutional Neural Networks