Hierarchical machine learning diagnosis framework based on a multimodal flexible wearable system for early identification of plant stress with suppressed microclimate cross-interference

Early and accurate plant-stress diagnosis remains essential for sustainable, data-driven agriculture. Current wearable systems based on spectral or microenvironmental sensing primarily measure exogenous signals and do not directly capture plant water status. Consequently, microclimate fluctuations can confound diagnosis by obscuring the distinction between environmental perturbations and physiological stress. Here, we present TRI-Wear, a multimodal flexible wearable system based on an “environmental driver–water flux–water stock–biochemical response” framework. TRI-Wear integrates temperature and humidity (T&H) sensors, in situ transmission spectroscopy, and a liquid-metal strain sensor to non-destructively monitor transpiration dynamics. It also incorporates stem radial micro-deformation, termed the “plant pulse,” as a proxy for turgor-pressure fluctuations. We developed a hierarchical machine-learning framework in which T&H and strain signals screen environmental disturbances and a Spectral Diagnostic Model (SDM) identifies biochemical stresses, including nutrient deficiencies and pest infestations. Within the monitored tomato cohort, the SDM achieved a mean accuracy of 99.3% under sample-level stratified 5-fold cross-validation across five biochemical classes. This value is a cross-validated within-cohort estimate rather than an independent-plant test result. TRI-Wear provides high-confidence warnings before visible symptoms and offers a route for translating flexible plant electronics toward agricultural Internet of Things (IoT) applications.