Rice blast disease, caused by Magnaporthe grisea, has posed a serious threat to global rice production in recent years. However, the conventional diagnostic techniques cannot fully detect the visible symptoms during the asymptomatic incubation period. In this study, a rapid and non-destructive diagnostic framework was established for the symptomatic and pre-symptomatic stages. Fourier transform infrared photoacoustic spectroscopy (FTIR-PAS) was also used to detect early metabolic and structural changes in the infected rice leaves. Spectral data ( 2 000-900 cm-1) of the sampling sites were collected from the healthy (H0-H3), symptomatic (D0-D4), and incubation-stage leaves using FTIR-PAS with depth-resolved scanning. The moving mirror velocity was also adjusted to 0.3, 0.4, and 0.6 cm/s. Background signals (D0/H0) were subtracted to isolate the pathogen-specific spectral features. Key absorption bands, 1 050 cm-1 (C-O stretching), 1 550 cm-1 (amide II), and 1 650 cm-1 (amide I), were analyzed to identify the disease-induced alterations. According to the intensity ratio of 1 650/1 050 cm-1 (R1650/1050), the diagnostic model was developed and then validated using Principal Component Analysis (PCA), Pearson correlation, and spatial gradient mapping. The original spectra revealed that in healthy samples (H1-H3), there were much lower absorption peak intensities at 1 550 and 1 650 cm-1 of the shallow-layer spectra at 0.4 and 0.6 cm/s, compared with the high absorption intensity in the deep-layer spectra at 0.3 cm/s. The 0.3 cm/s measurement also penetrated the cellular layer to detect the amide groups, while 0.4 and 0.6 cm/s were confined to the epidermis, thus lacking amide-related signals. In diseased samples (D1-D3), the absorption intensity at 1 550 and 1 650 cm-1 in 0.4 cm/s spectra increased significantly, closely resembling the 0.6 cm/s profiles. Furthermore, the absorption intensity of 0.3 cm/s spectra was progressively intensified and converged with 0.6 cm/s data, as the sampling sites approached the lesion center (D1 to D3, and D4 as the center). The rice blast infection reduced the epidermal thickness, thus leading to the spectral similarity between shallow and deep layers. There was a positive correlation between the degree of thinning and pathogen load. At the same time, the Background-subtracted spectra demonstrated that the weak absorption peaks were observed at 1 550 and 1 650 cm-1 in the healthy samples, comparable to baseline levels, whereas the diseased samples exhibited the strong peaks in the 0.867 spectral correlation with Magnaporthe oryzae. The pathogen-specific signals were verified in the infected leaves via FTIR-PAS. Furthermore, the magnitude of the R1650/1050 ratio was identified as a diagnostic marker for the rice blast. The total accuracies of 81%, 87%, and 77% were achieved with the F1-scores of 0.84, 0.91, and 0.80 at thresholds of 0.4, 0.5, and 0.6, respectively. The optimal diagnostic performance was obtained at the threshold of 0.5, with the precision and recall reaching 89% and 92%, respectively. Latent-stage diagnosis was also applied using background-subtracted spectra. The co-varying patterns of the 1 050 and 1 650 cm-1 bands mirrored the symptomatic stages. The R1650/1050 threshold of 0.5 was then achieved with 80% accuracy and an F1-score of 0.81, indicating the robust diagnostic efficacy for early infection detection. The high accuracy was combined with the interpretability using physiochemical biomarkers, particularly for early rice blast diagnosis. Unlike opaque machine learning, the R1650/1050 ratio was directly linked to the spectral changes in the pathogen activity and structural degradation. The simple configuration and high speed can offer actionable insights for the field deployment in precision agriculture. Multi-spectral indicators can also be integrated to validate the model's robustness and scalability under diverse environmental conditions.
- Article type
- Year
- Co-author
Pesticides can be often regulated to detect the deposition constitutes on the crops in fields. Nevertheless, the current research cannot focus on the specific difficulties in practical application. Fortunately, attenuated total reflection Fourier transform infrared spectroscopy (ATR-FTIR) technology can be expected to probe the presence of pesticides on the leaf surface without damage to the samples. This study aims to evaluate the in situ non-destructive sensing of pesticide deposition using infrared ATR-FTIR together with chemometric analysis. The examples were also taken as the broad-spectrum fungicide boscalid and the economic crop cotton. 140 standard samples of cotton leaves were then fabricated with known amounts of pesticide deposition. The spectral data was acquired after experiment. Subsequently, the preliminary analysis was conducted to determine the differences among the samples using principal component analysis (PCA). The great variations in spectral characteristics were attributed to the differences in pesticide deposition. Distinct patterns and trends were also identified from the data. After that the interval partial least squares (iPLS) and correlation analysis were employed to screen out the wavelength variables with strong correlations. This procedure aimed to identify the specific wavelengths the most relevant to pesticide deposition. Finally, a prediction model was established using the optimal variables and partial least squares regression (PLSR). The deposition amount of pesticide was predicted using the spectral data. The outstanding performance of prediction model was achieved with an R2 value of 0.83 and an RMSE of 1.21 μg/cm2 in calibration, while an R2 value of 0.86, an RMSE of 1.18 μg/cm2, an RPD of 2.3, and an RPIQ of 2.0 in validation. These results demonstrated the high accuracy and reliability of the prediction model on pesticide deposition. In addition, a classification model was also built with the probabilistic neural network (PNN), in order to distinguish whether the samples were with a higher deposition yield than the limit of detection (LOD) of the PLSR model. The PNN classification model shared an overall accuracy of 95% to discriminate whether the pesticide deposition amount on the samples was greater than the LOD of PLSR model. Therefore, it was feasible to classify the samples using PNN model. The prediction of PLSR model was better performed for the preliminary screening and classification. The results demonstrated that the ATR-FTIR shared the successfully perceived boscalid deposition on the surfaces of cotton leaf. Five well-defined peaks of absorption were observed at 3400, 2800, 1720, 1450, and 1160 cm-1, corresponding to the O-H, C-H, C=O, aromatic, and C-O stretching vibrations, respectively. These specific peaks were also resulted from the overlapping between the pesticide and leaf surfaces. The detection mechanism was enhanced from the valuable peaks information about the chemical interactions between the pesticide and the leaf. Among them, 272 wavelength variables were selected as the inputs that highly correlated with the pesticide deposition among the five well-defined peaks. The quantitative model was then established using PLSR. The most relevant spectra were selected to further improve the accuracy and performance of the model. The low LOD (3.54 μg/cm2) of PLSR model was calculated and validated by the RMSE value, indicating the sensitivity to detect even small amounts of pesticide deposition. Therefore, the ATR-FTIR technology can be expected to accurately detect the pesticides with different amount of deposition. This finding can provide the theoretical foundation and data support to in situ and non-destructive perception of its original deposition or residual amount in other crops. The promising potential can be offered for the pesticide monitoring and management, thus enabling more efficient and sustainable agricultural practices.
京公网安备11010802044758号