Integrating hyperspectral features and physiological parameters for quantitative estimation of fig rust severity

Fig rust caused by Phakopsora nishidana is a destructive foliar disease that impairs leaf photosynthetic capacity and seriously endangers fig yield and production stability. Traditional visual diagnosis of fig rust severity relies on manual observation, which is labor-intensive, highly subjective, and incapable of large-scale, quantitative disease evaluation. Although hyperspectral remote sensing has been widely adopted for plant disease monitoring, few studies have integrated hyperspectral reflectance features and leaf physiological parameters for quantitative estimation of fig rust severity, restricting the accurate and refined monitoring of this disease.

This study took leaf lesion coverage as a continuous quantitative indicator of fig rust severity, and collected hyperspectral reflectance data (400-1000 nm) and leaf physiological parameters from 336 fig leaf samples covering five disease severity levels. We systematically analyzed the variation characteristics of leaf physiological indexes under different disease degrees, and comprehensively compared the performance of different spectral preprocessing methods, feature wavelength selection algorithms, and machine learning regression models for fig rust severity prediction.

With the aggravation of fig rust infection, leaf chlorophyll and water content decreased significantly, while anthocyanin content increased remarkably; flavonol content fluctuated across different disease levels and only presented a weak correlation with lesion coverage. The combination of Savitzky-Golay smoothing and first derivative (SG+FD) was identified as the optimal spectral preprocessing method, and the competitive adaptive reweighted sampling (CARS) algorithm outperformed the successive projections algorithm (SPA) in feature wavelength screening. The pure spectral CARS-SVR model yielded an R² of 0.849 and an RMSE of 4.393, while the SVR model based solely on physiological parameters achieved an R² of 0.778 and an RMSE of 5.372. The fusion SVR model integrating spectral and physiological information obtained the best prediction performance, with R², RMSE, and RPD reaching 0.924, 3.067, and 3.641, respectively. Ablation analysis confirmed that chlorophyll content was the dominant factor improving model prediction accuracy. Additionally, model prediction reliability declined under high disease severity, with the severity of heavily infected leaves generally underestimated, showing uneven model performance across different infection levels.

The fusion of hyperspectral and leaf physiological data can effectively improve the accuracy and stability of quantitative fig rust severity estimation. The proposed method provides a reliable leaf-level technical solution for precise monitoring of fig rust, and offers valuable theoretical and methodological references for the intelligent quantitative evaluation of crop foliar diseases.