Fig 1.
Spectral graphs using mean pixel intensities according to genotype and days after salinity treatment.
Average pixel intensity for each image at a given wavelength was used to create spectral graphs for each day (x-axis) and for each genotype (y-axis). Non-zero pixel intensities were used to prevent bias due to higher number of pixels with zero (0) value during the earlier stages of plant growth.
Fig 2.
Spectral graphs of the data cleaned by smoothening through gap-segment derivatization, presented by genotype across different durations of salinity stress.
Spectral data used to create Fig 1 was cleaned and smoothened using a moving average (gap-segment) and 1st degree derivatization, which emphasized the peaks. This approach also removed the random changes in pixel intensity, which lead to smoother curves that emphasized real differences between treatments.
Fig 3.
Spectral graphs based on wavelengths showing high magnitude of differences between treatments, presented by genotype across different durations of salinity stress.
This dataset was generated by calculating the quartiles of pixel intensities in each wavelength, after which only the wavelengths that have a difference of more than the 3rd quartile between stress and control were selected. This approach intended to remove the wavelengths that may be non-contributory for the creation of PLSR models, which can sufficiently distinguish between salinity and control. However, this method also removed a substantial amount of data, which appeared to undermine the predictive power of the model.
Fig 4.
Empirical measurements of tissue Na+ and K+ contents across different durations of salinity stress as measured by flame photometry analysis.
Plants were sampled at three evenly spaced time-points (6, 12, and 18 DAS). Na+ content was stable under control condition but markedly elevated under salinity stress. Tolerant genotypes (i.e., donor parent Pokkali, FL510, FL478) had lower Na+ content compared to susceptible genotypes (i.e., recipient parent IR29, FL454, FL499) throughout the duration of the experiment. This trend implies that the difference in Na+ content started before the earliest sampling time-point (6 DAS). All genotypes showed a decreasing trend in K+ accumulation. There was little difference in K+ content in first sampling time-point, which increased with exposure time. These trends were similar across genotypes, implying that in terms of K+ accumulation, different genotypes appeared to have the same response regardless of their inherent tolerance or sensitivity to salinity.
Fig 5.
Trends in ion accumulation predicted by PLSR and the corresponding linear regression of the PLSR model created from the mean image pixel intensity.
The accuracy of the model was tested by creating a scatterplot between the samples not included in training the PLSR model, and their corresponding Na+ (A) and K+ (B) content predictions. While the models for both ions had satisfactory R2 values, the data-points tend to cluster in two groups. The model was further used to predict ion accumulation trends for Na+ (C) and K+ (D). Images taken during the last day of imaging were used in conjunction with the empirical ion content values that were determined by flame photometry. Overall, the trends in Na+ and K+ accumulation were opposite. However, the trends are also opposite from what is expected based on flame photometry. Despite this result, there is a definite deviation between salinity and control for both Na+ and K+.
Fig 6.
Trends in ion accumulation predicted by PLSR and the corresponding linear regression for the dataset cleaned by gap-segment derivatization.
The initial dataset used for creating the ion accumulation model in Fig 2 was smoothened and derivatized to the first degree. This approach removed the inconsistent peaks in the spectra, creating a more consistent data distribution. The regression scatterplots for testing the model accuracy are shown for Na+ (A) and K+ (B). For this dataset, the spread of data-points in the regression plot for Na+ was improved compared to the previous dataset, while the regression plot for K+ had the same clustering as previously observed. The predicted ion accumulation trends for Na+ (C) and K+ (D) are shown. As with Fig 2, the patterns of ion accumulation predicted by the PLSR model showed an opposite trend to what was expected based on flame photometry data. The scales for the prediction graphs had also changed compared with the initial dataset, as the predicted values were higher in general for this dataset.
Fig 7.
Trends in ion accumulation predicted by PLSR and the corresponding linear regression graphs for the dataset that included only the wavelengths with high mean difference between control and salinity.
The dataset used in Fig 2 was filtered for wavelengths with substantial mean difference between control and salinity. Difference in the means of control and salinity across genotypes were calculated and the third quartile values were used as threshold for selecting the wavelengths. In total, 67 out of 243 bands were used for the PLSR model. As the wavelengths no longer form a continuous selection of points like the other datasets in Figs 2 and 3, no additional smoothening and derivatization was necessary. The regression scatterplot for Na+ (A) also showed an improved spread compared to the plots in Fig 2. However, its R2 value was lower compared to that in Fig 3. The regression scatterplot for K+ (B) showed clustering of points into two groups in spite of high R2 value. Trends in ion accumulation for Na+ (C) and K+ (D) are shown, which appeared to be more similar to those shown in Fig 2.