Hyperspectral remote sensing of yellow mosaic severity and associated pigment losses in Vigna mungo using multinomial logistic regression models

Abstract

Yellow mosaic disease (YMD) has been a serious threat to blackgram cultivation especially during post-monsoon season. Visual assessment of disease severity is qualitative and time consuming. Rapid and non-destructive estimation of YMD by hyperspectral remote sensing has not been attempted so far on any of its hosts. Field studies were conducted for two seasons with eight blackgram genotypes having differential response to YMD. Comparison of mean reflectance spectra of the healthy and YMD infested leaves showed changes in all the broad band regions. However, reflectance sensitivity analysis of the narrow-band hyperspectral data revealed a sharp increase in reflectance from the diseased leaves compare to healthy at 669 (red), 505 and 510 nm (blue). ANOVA showed a significant decrease in leaf chlorophyll (p < 0.0001) with increase in disease severity, while no such relationship was observed for relative water content. By plotting coefficients of determination (R2) between leaf chlorophyll and percent reflectance at one nm wavelength interval, two individual bands (R571; R705) and two band ratios (R571/R721; R705/R593) with highest R2 values were selected. These bands showed a significant linear relationship with SPAD chlorophyll readings (R2 range 0.781–0.814) and spectrometric estimates of total chlorophyll content (R2 range 0.477–0.565). Further, the relationship was stronger for band ratios compared to single bands. With optimal spectral reflectance ratios as inputs, disease prediction models were built using multinomial logistic regression (MLR) technique. Based on model fit statistics, reflectance ratios R571/R721 and R705/R593 were found better than the individual bands R571 and R705. Validation of MLR models using an independent test data set showed that the overall percentage of correct classification of the plant into one of the diseased categories was essentially same for both the ratios (68.75%). However, the MLR model using R705/R593 as dependent variable was of greater accuracy as it gave lower values of standard errors for slopes (βG range 9.79–36.73) and highly significant estimates of intercept and slope (p < 0.05). Thus the models developed in this study have potential use for rapid and non-destructive estimation of leaf chlorophyll and yellow mosaic disease severity in blackgram.