Machine-learning-assisted photoacoustic phenotyping of storage energy and thermal diffusivity in three rice varieties under alternate wetting and drying irrigation and contrasting substrates
Abstract
Drought stress poses significant challenges to rice production, necessitating precise, non-invasive phenotyping tools for identifying resilient genotypes. The present study evaluated storage energy (SE) and thermal diffusivity (TD) in three rice varieties (F-67, F-68, and F-2000; 120 plants, with 20 per variety × substrate combination) grown under alternate wetting and drying conditions in two contrasting substrates: control soil-substrate and coconut peat substrate. The term SE denotes the fraction of absorbed light energy stored in photosynthetic products. In contrast, TD is a measurement of the rate of heat propagation through leaf tissue. This parameter is directly linked to tissue structure and leaf water content. Measurements employed the photoacoustic technique with frequency-modulated light (21–47 Hz) to probe depth-dependent leaf energy dynamics. Statistical analyses, including analysis of covariance with days after transplanting as a developmental covariate, revealed significant substrate effects on both SE and TD, with coconut peat-grown plants exhibiting lower SE and TD, consistent with reduced water stress. Treatment period significantly influenced both parameters, with prolonged water deficit producing the most pronounced physiological responses. After excluding sampling day, optimized machine learning (ML) classification achieved an overall treatment-period accuracy of 70%, although balanced accuracy and macro-F1 indicated more modest multiclass discrimination. Substrate classification remained limited at approximately 58% accuracy, thereby corroborating the modest effect sizes observed in linear models. Principal component analysis and unsupervised clustering provided partial support for treatment-related physiological groupings, although substantial overlap remained among groups. Notably, varieties showed no significant differences in leaf-level SE or TD despite marked morphological divergence in tillering, leaf area, panicle traits, and biomass allocation, suggesting that varietal responses to reduced water and nutrient supply manifest through differential resource partitioning rather than alterations in leaf-level SE. These findings suggest that SE and TD may serve as potentially valuable non-invasive indicators of water-deficit and substrate responses, pending validation across broader genotype panels and environments. Furthermore, ML integration with traditional statistics could provide multi-faceted phenotyping frameworks for drought tolerance assessment in rice breeding programs.