Biospeckle imaging enables non-destructive observation of dynamic physiological activity in plant tissues; however, the relative sensitivity of different biospeckle activity maps to water stress and their implications for data-driven classification remain insufficiently understood. This study systematically evaluates multiple biospeckle activity mapping approaches for water stress analysis in maize (Zea mays L.) leaves and examines how their characteristics influence deep learning–based classification performance. Maize plants were subjected to three irrigation levels (0%, 50%, and 100%) over a 7-day experimental period. Stomatal conductance was measured as an independent physiological reference, and a microfluidic phantom experiment was conducted to verify the physical response behavior of the biospeckle imaging system. Temporal variations in biospeckle activity were statistically analyzed, followed by deep learning–based classification using representative two-dimensional convolutional neural network models. Statistical analysis revealed that biospeckle activity exhibited stress-dependent responses, with severe water stress (0%) being consistently distinguishable, whereas moderate and well-watered conditions (50% and 100%) showed partially overlapping patterns. These trends were consistent with stomatal conductance measurements. Deep learning models trained on different biospeckle activity maps achieved classification accuracies of up to 0.73 and macro-averaged F1 scores of 0.73, with notable differences in performance depending on the selected activity representation. These results suggest that while traditional statistical parameters show limited linearity, the proposed deep learning-based biospeckle analysis could serve as a useful tool for water stress classification. By capturing complex spatial-texture features, this study presents a potential data-driven approach for precision plant phenotyping.
Lee et al. (Fri,) studied this question.