Work

Detecting wheat powdery mildew through the integration of two-dimensional correlation spectra and dual-mode deep learning

Year: 2026

Type: article

Abstract: Accurate and timely monitoring of wheat powdery mildew (WPM) severity is essential for precision crop management and yield loss mitigation. Hyperspectral technology offers a non-destructive detection capability, as the series of symptoms produced by the disease induces specific spectral responses. However, accurately decoupling pathological features from complex, overlapping canopy signals and elucidating the dynamic physiological mechanisms driving spectral variations remained a common challenge in current research. To tackle these challenges, this study proposed a novel framework integrating two-dimensional correlation spectra (2DCOS) with a dual-modal deep learning architecture, aiming to explore pathologically sensitive features and enhance detection accuracy. Field canopy hyperspectral data, crop biochemical parameters, and disease severity were investigated over a three-year period (2022–2024) in Xinxiang, China, to demonstrate the validity of the proposed framework. Utilizing disease severity as external perturbations, 2DCOS was employed to enhance spectral resolution and reveal deep-seated pathological information. Subsequently, a dual-modal network incorporating a squeeze-and-excitation (SE) attention mechanism was constructed to synergistically fuse 1D spectral vectors (capturing intensity features) and 2D correlation images (capturing inter-wavelength coupling and textural patterns). The analysis of asynchronous spectra based on Noda's rules successfully reconstructed the chronological order of WPM-induced stress. Results indicated that pigment degradation (650–690 nm) manifested prior to structural collapse (760–930 nm) during the mild infection stage, whereas structural deterioration and leaf desiccation became the dominant drivers of spectral variance in severe stages. Based on these dynamic insights, the 760–930 nm region was identified as the most robust sensitive window. This was attributed to the presence of distinct auto-peaks across the mild-to-severe stages, which effectively filtered out unstable signals in the visible region and interference from transient water stress. Validations across three growing seasons (2022–2024) demonstrated the superiority of the proposed method. The dual-modal network achieved the highest estimation accuracy with a validation R 2 of 0.88, an RMSE of 9.13%, and an NRMSE of 29.88%, outperforming optimal machine learning models (R 2 = 0.79, RMSE = 13.67%, NRMSE = 44.76%) and single-modal CNN architectures (R 2 = 0.86, RMSE = 11.02%, NRMSE = 36.08%). Furthermore, the model exhibited strong robustness across different wheat cultivars and inter-annual environmental variations. This study not only elucidates the biophysical mechanisms of canopy spectral responses to varying WPM infection levels, but also provides a theoretically grounded and highly precise deep learning solution for field-scale disease monitoring.

Source: Artificial Intelligence in Agriculture

DOI: 10.1016/j.aiia.2026.04.007