Sensing water stress from plant motion: A structured and low-cost approach to digital plant phenotyping

Water stress is a major target for digital phenotyping, yet many sensing solutions rely on costly hardware or indirect physiological proxies. This study asks whether subtle plant motions captured with consumer-grade camera can provide a scalable indicator of plant water status. We present a structured digital phenotyping pipeline that converts time-lapse sequences into optical-flow fields and extracts the newly-developed morpho-kinematic (MK) features that summarize the canopy structure along with magnitude, directionality, and temporal dynamics of canopy movements. The experimental program follows an iterative, data-driven trajectory. Early hyperspectral and open-field trials on zucchini highlighted how high-dimensional-low-sample settings and environmental confounders can dominate water-stress signals and destabilize validation. A subsequent pilot trial based on time-lapse orthophotos exposed operational bottlenecks making segmentation and sample tracking unreliable. These lessons, at the cost of a smaller but higher-quality sample size, motivated a pivot to side-view perspective on lettuce to isolate informative tissues. Within this new setting MK features were formalized and the most stable cross-experiment classifier configuration – i.e., biologically driven acquisition + Adaptive Linear Opinion Pooling (ALOP) – reached 0.91 balanced accuracy. This was obtained by iteratively testing for robustness and transferability across sequential controlled-environment trials, ablations and additive studies. Subsequently, a semi-controlled transfer step (tunnel) identified new bottlenecks concentrating, again, in the segmentation step. Nonetheless, a reproducible short-horizon predictable component (next day feature increment), based on current state + irrigation decision, was demonstrated. Thereafter, the associations between motion signatures, agronomic performance, and secondary metabolites showed incomplete but viable. Finally, we demonstrated the loop logic into which embed the sensing pipeline (time-lapse and MK features), the state estimator (ALOP), the predictor and the planner for supporting irrigation decision in a digital-tween-inspired framework.