Hyper-spectral estimation of wheat biomass after alleviating of soil effects on spectra by non-negative matrix factorization

Hyper-spectral technology has been proven to be an effective method for the fast and non-destructive monitoring of crop biomass. However, the biomass estimation accuracy of this method is limited due to the effects of background factors, such as soils and water. In this study, a spectral separation method, non-negative matrix factorization (NMF), was proposed to alleviate the effects of soil on spectra. With the application of the NMF method, pure vegetation spectra were extracted from the field-observed spectra of wheat canopy, which were collected in four growing seasons from the tillering to the booting stages of wheat. Then, prediction models of wheat biomass (WB) were established and validated using the extracted spectra with the partial least squares regression (PLSR) method. The results showed that the NMF method could effectively separate the vegetation spectra from the mixed canopy spectra. Based on the extracted vegetation spectra, the WB prediction accuracy could be greatly improved with an increase of 31.7% for the R2p and an increase of 46.6% for the ratio of performance to deviation (RPD) as compared to the original spectra, indicating that the NMF method could significantly improve the performance of the WB prediction model. This method has potential application in the estimation of biomass using remote sensing technology.