Speaker: Fernando David Modrego Arceo
Abstract: Plant bioacoustics is an emerging field that suggests that plants not only respond to sound stimuli but also emit detectable acoustic signals, particularly under stress conditions. Recent studies have revealed airborne ultrasonic emissions produced by plants subjected to drought or physical damage, which opens new possibilities for noninvasive monitoring of their physiological state. However, the approaches used so far to classify these signals have been of low complexity. This work proposes a methodology based on artificial intelligence and deep learning to classify plant acoustic emissions under water and mechanical stress. For this purpose, the signals acquired by Khait et al. [1] were employed, and a comprehensive processing and classification system was developed, including spectral transforms combined with Gradient Boosting Machines (XGBoost) [2] and deep convolutional neural networks (CNN) with a ResNet-50 backbone [3]. The results show a significant improvement in classification accuracy compared to traditional models such as support vector machines (SVM). In particular, the CNNbased models achieved an average accuracy of 79% in multiclass classification (distinguishing between stressed and non-stressed plants and plant types) and 92% in a one-vs-rest class verification system. This study validates the feasibility of using artificial intelligence for automatic detection of plant stress via acoustic signals and lays the foundations for future applications in precision agriculture, crop phenotyping, and sustainable environmental monitoring.