Weather-Driven Multi-Disease Wheat Disease Forecasting for IoT-Oriented Early Warning Systems

Abstract

Early detection of wheat diseases such as Fusarium head blight, stripe rust, Septoria tritici blotch, and stem rust requires continuous monitoring of environmental conditions that drive pathogen development. This work investigates a weatherdriven forecasting framework for short-horizon disease early warning using hourly ERA5-Land data as a practical proxy for field microclimate sensing. A preprocessing pipeline imposes temporal consistency on hourly weather records and derives features linked to disease physiology, including humidity-duration metrics, dewpoint proximity, wetness proxies, rainfall accumulation, radiation conditions, and shortterm temperature-humidity interactions. Disease-specific weak labels are generated from literature-based biological rules, allowing supervised forecasting without fieldconfirmed event annotations. Several learning approaches, ranging from baseline classifiers to lightweight temporal deep models, are evaluated for forecasting near-future disease risk at 3 h, 6 h, and 12 h horizons across multiple locations. The results show that weather-driven forecasting can provide strong short-horizon performance for multiple wheat diseases, with boosting-based classical models outperforming temporal deep learning models in the current setup. Although the current implementation is weather-first rather than a deployed sensor system, the proposed pipeline is aligned with future IoT-oriented early warning and can support later integration with field microclimate nodes and distributed sensing architectures.