Spectroscopic detection of rice leaf blast infection from asymptomatic to mild stages with integrated machine learning and feature selection

Rice blast is considered as the most destructive disease that threatens global rice production and causes severe economic losses worldwide. A detection of rice blast infection in an early manner is vital to limit its expansion and proliferation. However, little research has been devoted to spectral detection of rice leaf blast (RLB) infection, especially at the asymptomatic or early stages. To fill the gap, this study aimed to examine the feasibility of detecting RLB infection from leaf reflectance spectra at asymptomatic, early and mild stages of disease development. Greenhouse experiments were conducted over two consecutive years to collect hyperspectral data (350–2500 nm) on various days after inoculation (DAIs) for the three infection stages. These hyperspectral data were processed to select disease specific spectral features (DSSFs). Such DSSFs were then used to feed the machine learning based sequential floating forward selection (ML-SFFS) methodology for determining the optimal feature combination (OFC) and overall accuracy (OA) in the detection of RLB at various infection stages.

The results demonstrated that the rice plants displayed considerable biochemical and spectral variations and this pattern of variations existed consistently during plant-pathogen interactions. A multivariate pool of DSSFs comprising two reflectance bands, fourteen SIs, and five continuous wavelet coefficients, were determined for revealing the dynamic response of RLB infection across two years. The combination of 2 to 4 spectral features selected by the ML-SFFS algorithm was sufficient to identify infected leaves with classification accuracies over 65% and 80% for the asymptomatic and early infection stages, respectively. The OA could rise up to 95% for the mild stage. Compared to the use of all DSSFs with a support vector machine (SVM) classifier, the SVM-based SFFS (SVM-SFFS) algorithm prevailed in the classification accuracy up to 10% over the sampling period. Our results demonstrated the feasibility of accurate classification of RLB infected samples by ML-SFFS. This study suggests that reflectance spectroscopy has great potential in the pre-visual detection of RLB infection and airborne or spaceborne imaging spectroscopy is promising for the mapping of early occurrence and severity levels of RLB infection at large scales.