| China has a huge demand for high-quality rice.The prediction of rice grain quality plays an important role in cultivation management and harvest of rice.Therefore,it has become a problem to find an efficient and accurate method for predicting rice grain quality.The development of hyperspectral remote sensing provides useful data for crop growth parameters monitoring and prediction.Based on the practical needs of rice hierarchical acquisition and harvest,this study focuses on the prediction of rice grain quality parameters at different level（rice powder,dry panicle,fresh panicle and canopy）.The results can provide technical support for prediction of crop quality using satellite hyperspectral data,and provide important reference for national and local agricultural departments to formulate rice cultivation,production and harvest strategies.Using the traditional band combination method as the contrast,and taking the organs and canopies reflectance spectra as the basic data,this method consists of the following five steps:1)continuous wavelet transform;2)extraction of sensitive wavelet features;3)analysis of common characteristics;4)application of sensitive wavelet features of organs spectra;5)prediction of rice quality.Finally,prediction model of rice grain quality was established at organs and canopies level.The results show that,compared with spectral index,the prediction model based on sensitive wavelet features can effectively improve the accuracy of prediction model,and the independent verification result is better than spectral index.Based on the spectra of rice powder and dry panicle,the protein sensitive wavelet feature WF17105,（rice powder:R2=0.56;dry panicle:R2=0.55）and amylose sensitive wavelet feature WF2037,6（rice powder:R2=0.59;dry panicle:R2=0.62）were extracted.Otherwise,the protein sensitive wavelet coefficients WF1710,5 and amylose sensitive wavelet coefficients WF2037,6 of rice powder and dry panicle can be applied to fresh panicle spectra,but not to canopy scale;the protein sensitive band 2089 nm and amylose sensitive band 1321 nm at the fresh panicle level can be applied to the canopy spectra at the full ripe stage.WF740,3 of canopy spectra at heading stage can estimate GPC better（R2=0.60）with RMSE=0.38%;WF 1320,4 of fresh panicle spectra at dough ripe stage can estimate GAC better（R2=0.58）with RMSE=2.53%.Fresh panicle and canopy are different from dry panicle and rice powder,the prediction model can not eliminate the differences of GAC among varieties,so it needs to be predicted by varieties.The correlation between WF 1594,3 of fresh panicle spectra and GAC of varieties with low starch content is R2=0.70,RMSE=1.18%,and that between WF2066,4 at canopy scale and GAC of varieties with low starch content is R2=0.67,RMSE=0.47%.The prediction accuracy of variety of high starch content was higher than that of low starch content,but the underestimation was serious.The sensitive wavelet features extracted in this study has high stability and great application value.The results broaden the application scope of continuous wavelet analysis and provide technical support for the prediction of rice grain quality.Overall,this study systematically explored the problems existing in the prediction of grain quality of rice and wheat,and proposed a series of solutions,mainly including extraction of sensitive wavelet features of rice quality based on the continuous wavelet analysis,constructing the prediction model;finding the common characteristics of rice poweder and dry panicle spectra,on this basis,analyzing response to the change of grain quality and the applicability of the sensitive wavelet features in fresh panicle and canopy spectra.Our result would advance the application of continuous wavelet analysis in crop quality parameter sensitive feature extraction,reduce the uncertainty of quality parameter prediction,and provide scientific support for large-scale prediction of rice grain quality. |
PDF Full Text Request