Research On Hyperspectral Characteristic Information Fusion And Modeling Of Rapeseed

Hyperspectral remote sensing technology has become one of the important technical means of precision agriculture because of its advantages of large amount of information,convenience,non-destructive and high degree of automation.Many scientists and technicians have carried out a lot of research work in crop hyperspectral remote sensing growth monitoring,field crop disaster prediction and evaluation,yield estimation and so on.The main research idea is to extract hyperspectral sensitive bands,to construct characteristic spectral index,and to model the biochemical parameters of crops through certain data processing tools,statistical methods and mathematical operations.This method,which only utilizes one or two characteristic band information in hyperspectral data,is simple and easy to implement.And it has obvious advantages,but on the other hand,wasting a lot of hyperspectral characteristic information.In recent years,with the development and application of cloud computing,internet of things,large data and artificial intelligence technology,how to dig useful information in hyperspectral data and further optimize the inversion accuracy and reliability of spectral application model has become an urgent research area.Taking rapeseed as the research object,this paper studies the combination of hyperspectral characteristic bands,the dual-parameter fusion,the multi-parameter fusion,and the application of machine learning algorithm in hyperspectral information fusion.And its aim is to provide some theoretical reference and accumulation of practical experience for the research field of hyperspectral agricultural characteristic information fusion.1.A method to construct characteristic parameters of information accumulation in hyperspectral continuous characteristic bands is proposed.It uses the mathematical transformation of Sigma to accumulate the information of hyperspectral characteristic bands information of the normalized vegetation index NDVI,and constructs the derivative parameters of the A_VNDVI series.The research shows that the cumulative A_VNDVI derivative parameters have a good correlation with the leaf area index of rapeseed,among which A_GRNDVI and A_GNDVI have the best inversion performance,and the determinant coefficients R² are 0.7252 and 0.729,respectively.The LAI regression equation of each cumulative vegetation index and its prototype vegetation index improves in fitting coefficient to some extent,and the range is generally about 1.5%.The determinant coefficient of the regression model of index A_BNDVI is 4.74%higher than that of BNDVI,and the increase is obvious.The LAI estimation model of rape canopy was also validated.And its result shows that the inversion models of A_GDNVI,A_NDVI,A_GBNDVI and A_GRNDVI had good applicability.The anti-interference analysis of the model shows that the anti-interference ability of A_VNDVIparameters has been greatly improved by introducing more hyperspectral characteristic information,but the sensitivity of the model has also decreased accordingly.Therefore,when the measurement background is complex or the interference information is strong,A_VNDVIcan be selected for LAI inversion and estimation.2.A method to construct new parameters of information accumulation in characteristic bands between hyperspectral remote sensing data is proposed.The method selects sensitive bands according to certain intervals in the hyperspectral characteristic band,constructs the interval-band cumulative vegetation index IAVI,and then models the SPAD of rape canopy leaves.The results show that the inversion models of IA_NDVI,IA_RVI,IA_GNDVI and IA_BNDVI all have the higher performance than those of the corresponding pre-derivation vegetation indices.Among them,IA_GNDVI has the greatest increase,whose R² increases by 0.0257 and whose RMSE decreases by 0.117.Compared with BNDVI,IA_BNDVI has more obvious optimization range of performance,△（R²）and△RMSE are 0.0232 and-0.051,respectively.In the validation data analysis,IA_NDVI and IA_GNDVI are better than NDVI and GNDVI respectively,and IA_BNDVI is better than BNDVI in R² and RMSE.Only in RE,it is slightly higher than BNDVI model.Sensitivity and anti-interference analysis show that IAVI model has certain advantages in anti-interference.Therefore,IAVI characteristic parameter inversion is worth considering when there are complex background of hyperspectral measurement and many interference factors.3.Four weighted fusion methods are proposed according to the commonly used evaluation indexes of statistical models:the reciprocal root mean square error method（RR）,the weighted determination coefficient method（RS）,the weighted optimal root mean square error method（OR）and the weighted optimal determination coefficient method（ORS）.On the above basis,six spectral characteristic parameters such as S_Dr and R_rb were selected,and eight spectral characteristic parameters such as RVI and NDVI were selected to carry out the LAI dual-model fusion of rape canopy and SPAD multi-model fusion of canopy leaf.The result shows that the dual-parameter fusion model has obvious advantages over the single-parameter regression model in overall performance.Among the 15 fusion models analyzed,12 models have better determinant coefficients and root mean square error than single parameter regression models.Among them,the?（R²）of the leaf area index OR/ORS regression model of F(S_D,ρ₂))was 0.0904 and its reduced range of?（R²）was0.0617.In the fusion process,the regression model can achieve the maximum simultaneous decision coefficient and the minimum RMSE value,thus achieving the integration of OR and ORS fusion models.On the premise that independent variables and dependent variables have a certain correlation,the stronger the correlation between the characteristic parameters involved in fusion,the less significant the fusion effect is.On the contrary,when the correlation is weak,the fusion model may have more advantages.The result of multi-parameter fusion shows that 19 of the 21 multi-parameter fusion models studied have higher accuracy and reliability than the single-parameter regression model before fusion,and the adjustment determinant coefficient Adj＿R² of the model has been improved in varying degrees,while the RMSE value has been reduced accordingly.In addition,the result of model validation also shows that OR/ORS regression model performs better than RS and RR models in training,but may be weaker than RS and RR models in generalization performance.4.Dual-parameter and multi-parameter information fusion regression models of rapeseed canopy leaf SPAD are constructed based on BP neural network,support vector machine（SVM）and partial least squares multiple linear regression（PLS-LR）.The comparison analysis shows that BPNN fusion model has high accuracy,good generalization performance and obvious advantages on the whole.And RR,RS and OR/ORS fusion methods also have good fusion effect.In some parameter groups,the performance of the fusion model is even better than that of BPNN model,showing high application value,while PLS-LR and SVM methods are relatively poor on the whole.In the initial stage of multi-parameter fusion（dual-parameter fusion,tri-parameter fusion and four-parameter fusion）,redundant information plays an obvious role,and the fusion effect is often more significant.The increase of decision coefficient and the decrease of root mean square error of fusion model are obvious,while the optimization range of model performance in the follow-up fusion process is gentle as a whole.Therefore,with the advancement of the fusion process,the fusion characteristic parameters should be selected more carefully.