Beyond Leaf Images: A Personalized Multimodal Temporal AI Framework for Pre-Symptomatic Plant Health Monitoring and Disease-Risk Early Warning

Abstract

Artificial intelligence has advanced plant disease recognition, mainly through image-based deep learning. Its weakness is timing: a camera records a visible symptom, a model matches it to a known class, and a label is returned. This confirms disease but says little about the earlier period, when a plant is already changing yet shows no clear lesion. This paper proposes PlantPulse, a personalized multimodal temporal framework for that pre-symptomatic window. It draws on environmental, soil, optional electrophysiological or impedance, thermal and volatile organic compound signals, together with periodic RGB images. Rather than judging every plant against a population model, PlantPulse learns a reference pattern for each plant during a healthy period, scores new observations against it, and uses a temporal anomaly layer to detect departures. A risk layer then uses environmental and physiological context to separate normal variation, plausible abiotic stress, possible disease risk and cases in which the evidence is insufficient, while an uncertainty layer allows the system to abstain rather than mislabel unusual signals as disease. The paper defines a common data representation, candidate statistical and deep learning models, three multimodal fusion strategies and an evaluation protocol built around warning lead time, false alarm burden, calibration, missing modalities, cross-environment testing and leakage control. PlantPulse is presented as a research design rather than a validated system; its value must be established through longitudinal trials with independent ground truth. The aim is continuous monitoring in which images remain important but are no longer the only evidence.