Purpose To achieve early detection of rice leaf blast and nitrogen deficiency at the canopy scale using hyperspectral data from rice field plots.
Methods Based on nitrogen stress gradient experiments and natural field infection, hyperspectral images of nitrogen-deficient and naturally infected rice canopies were captured using an outdoor hyperspectral imaging system. Spectral reflectance features of moderate and mild infection, nitrogen-deficient, and healthy rice canopies were extracted and analyzed. For the preprocessed hyperspectral data, three feature extraction methods were applied, including principal component analysis (PCA), vegetation index (VI), and competitive adaptive reweighted sampling (CARS), and linear discriminant analysis (LDA) and support vector machine (SVM) classifiers were used to construct early detection models for rice leaf blast and nitrogen deficiency.
Results The model constructed using PCA alone did not show significant classification performance. The models based on VI showed slight improvement over the full-spectrum models, but the enhancement was not substantial. The model constructed using features selected by the CARS method demonstrated the best discriminatory ability. Further dimensionality reduction of the extracted CARS features by PCA, five principal components were got and then used to construct the SVM and LDA models, which had overall classification accuracies of 95.51% and 96.15%, with Kappa coefficients of 0.91 and 0.92, respectively, demonstrating high classification consistency.
Conclusion This study successfully achieves effective identification of rice leaf blast and nitrogen deficiency at the canopy scale using a limited number of feature variables. The results provide new insights and theoretical foundations for remote sensing monitoring of rice diseases and nutrient stress. The findings can support large-scale rice pest and disease monitoring and precision spraying, offering significant applications in the field of precision agriculture.
DownLoad: