Abstract
Drought poses a significant challenge by inducing water stress in crops, urgently calling for effective monitoring and early intervention. In this work, we conducted a comprehensive 38-day investigation of the bioimpedance of water stressed tomato stems under controlled environmental conditions. A total of 8000 measurements were meticulously collected across a frequency spectrum spanning from 100 Hz to 10 MHz. The collected data was categorized into control, early stress, and late stress groups, corresponding to distinct phases of water stress treatment. To analyze the data, we employed established equivalent circuit models, including Cole, Randles, and double Cole. Subsequently, we evaluated the performance of eight machine learning algorithms widely-used in predicting the water stress stages of the plants. Remarkably, the Cole model, in conjunction with a multi-layer perceptron (MLP) algorithm, demonstrated robust performance, achieving an impressive F1 score of 0.89. In fact, in the control group, 378 out of 388 instances were accurately identified. However, misclassifications were observed in the early and late stress groups, with 31 out of 179 and 14 out of 208 instances mislabeled, respectively. Furthermore, we scrutinized the Cole model’s circuit parameters over time, providing insights aligned with the plant physiological behavior documented in the literature. Our findings suggest that MLP models trained on stem bioimpedance data hold promise as a valuable technique for estimating water stress in tomato plants, offering an essential tool for proactive management by farmers.