Integrating Leaf-Image Deep Learning, Meteorological Risk-State Discovery, and Stochastic SICR Dynamics for Potato Early and Late Blight: A Data-Science and Intelligent-Systems Framework

Abstract

Potato early blight and late blight create substantial agronomic risk, yet leaf-image diagnosis, weather-driven warning, and transmission modeling are often treated as separate tasks. This paper develops an integrated data-science and intelligent-systems frame-work that combines convolutional neural network-based symptom recognition, meteorological environmental suitability modeling, AI-based weather-risk-state learning, and a stochastic susceptible-infected-carrier-recovered transmission mechanism. Leaf-image evidence is drawn from the Potato Disease Leaf Dataset. ResNet50 transfer learning provides a cross-source generalization baseline, whereas Xception fine-tuning provides the stronger image-classification branch. Weather-risk modeling uses daily temperature, relative humidity, and rainfall from Chahar Right Front Banner, Ulanqab, situated within an expanded three-growing-season record (2023–2025). Disease-specific environmental suitability indices are computed using Gaussian temperature, sigmoid humidity, and rain-fall saturation responses; K-means clustering identifies recurring weather-risk states, and a cross-validated logistic surrogate classifier learns threshold-defined warning labels. Image performance is strongly protocol-dependent: ResNet50 reached 66.32% accuracy under a stricter cross-source test, whereas fine-tuned Xception exceeded 95% under the source split, so the two are not directly comparable. August 2025 showed markedly higher suitability than August 2024, and K-means recovered a 16-day high-risk state. Because the humidity-complete window is limited to two August periods (62 days), these warning-state results are demonstrative rather than broadly generalizable. The study contributes a transparent, modular, and reproducible early-warning framework while explicitly distinguishing surrogate warning-state learning from field-validated disease-incidence prediction.   Received: 4 June 2026 | Revised: 10 August 2026 | Accepted: 28 August 2026   Conflicts of Interest The authors declare that they have no conflicts of interest to this work.   Data Availability Statement The data that support the findings of this study are openly available on Zenodo at https://doi.org/10.5281/zenodo.22105735. This deposit comprises the leaf-image dataset that supports the image-classification results (the Potato Disease Leaf Dataset, with early blight, late blight, and healthy-leaf categories), the daily meteorological records for Chahar Right Front Banner (Ulanqab City, Inner Mongolia) used in the environmental suitability analyses, the derived ESIs, and the analysis code. The redistribution conditions of the original meteorological sources were verified prior to deposition.   Author Contribution Statement Hongzhi Zhang: Conceptualization, Methodology, Investigation, Resources, Data curation, Writing – original draft, Writing – review & editing, Project administration, Funding acquisition. L. P. Varlamova: Validation, Writing – review & editing, Supervision. Jianan Chen: Software, Formal analysis, Data curation, Visualization.