Diagnosis the dust stress of wheat leaves with hyperspectral indices and random forest algorithm

摘要

The dust stress of wheat leaves were diagnosed with hyperspectral technology. A four-level dust stress (i.e. severe, moderate mild stress and normal status) experiment was conducted using the wheat in growing season as the study object. The spectra of wheat leave which treated with different stress intensity levels were collected by AvaSpec-2048×14-USB2 Spectrometer, and 32 samples were obtained for each level. All of the samples were divided randomly into two groups, one group with 96 ones used as calibrated set, and another with 32 ones as prediction set. The spectra data were pretreated by the methods of S.Golay smoothing and baseline correction. Using vegetation indices (VIs) designed for dust pollution stress diagnosis (i.e. dust stress normalized index, DSNI) as the model input variables, and the values of stress level as the output variables, the hyperspectral diagnosis models of dust stress intensity were established by random forest classification algorithm. And then 32 unknown samples were predicted by the diagnosis model. The result showed that the prediction accuracy was 87.5%, indicating it was feasible to diagnose the dust stress intensity with the new designed VIs.