Quantitative Study Of Spatial Prediction Accuracy Based On Spectral Images

Hyperspectral imaging has been used for the inspection of an increasing number of agro-products and render their spatial distributions of quality indices.Compared with conventionalchemicalvaluesofregions,predictionofspatialdistributionof quality-representing indices over a surface enriches the spectra-based information a new dimension of space,making it possible for people to"see"the distribution of the detected values on the sample surface and their coverage areas,a unique capability among all agro-inspection techniques.At present,prediction functions are trained based on regional spectra and then be applied on pixel spectrum to get the corresponding spatial distributions in the form of prediction maps of different quality-attributes.Though,the accuracy of such spatial prediction has not yet been fully studied.In fact,it is impossible to verify the results of spatial prediction.Because only an average value can be obtained in one area of the sample using conventional physicochemical methods.Consequently,the values at individual pixels in a spectral image are usually not available for validation,in other words,the spatial prediction results can not be verified by comparison with their measured values.Considering the aforementioned,a method of evaluating the accuracy of the spatial prediction by measuring the error of data acquisition system is used in this work to lay the foundation of validating the spatial accuracy of spatial predictions using spectral images.The“spectra”data in Matlab2012a is used as a basic data source,and different spatial distribution samples are generated by random fill of computer.The hyperspectral images of99%,75%,50%,and 20%standard reflectivity calibration plates are collected through the acoustooptic tunable filter?AOTF?acquisition system under fixed exposure and variable exposure mode in the range of 550¹000nm,and are used to study the signal to noise ratio?SNR?of the system.The results show that compared with fixed exposure,the SNR of spectral images can be effectively improved by variable exposure acquisition mode,especially on the two sides of spectrum.The system noise in the two acquisition modes has different forms of error distribution,which provides a reference and guidance for how to increase the error of the sample hyperspectral image to make it closer to the real situation.The prediction functions of partial least squares?ALL-PLS?and genetic algorithm partial least squares?GA-PLS?based on regional data are applied to spatial prediction on the premise of ensuring regional accurate prediction.When there is no error,the root mean square error of spatial prediction accuracy is less than 0.1476,the coefficient of determination is greater than0.9900,and the accuracy is very high.When there is the error of variable exposure mode,the accuracy of spatial prediction has decreased(accuracy:RMSE_ALL-PLS=0.6924,R²_ALL-PLS=0.6724,RMSE_GA-PLS=0.8290,R²_GA-PLS=0.6244).The accuracy of spatial prediction is poor under the error of the fixed exposure mode(accuracy:RMSE_ALL-PLS=1.7329,R²_ALL-PLS=0.5321,RMSE_GA-PLS=2.1047,R²_GA-PLS=0.2154).In order to further study the influence of system noise on spatial prediction,the error of spectral imaging system and the error of physical and chemical detection instruments are gradually added to the ideal sample data.The results show that the spatial prediction accuracy of the two functions decreases rapidly with the increase of spectrum and instrument error.When the systerm error reaches a certain degree?the spectrum error:the front?0.03,0.1?,the middle?0.03,0.03?,the posterior?0.03,0.1?;the instrument error:b=0.2?,the accuracy of spatial prediction is below the critical value.The spatial prediction results can no longer accurately reflect the actual distribution of the sample and lose the value of application.By measuring the error of the data acquisition system,the accuracy of spatial prediction of the agro-product qualities based on the spectral images can be evaluated quantitatively,which provides a reference for further research.The essence of the technology of using spectral imaging technology to detect the agro-product qualities,which is that the rule of spatial prediction with the change of system noise is obtained through the characteristic curve of the regional spectrum,is studied.It is not restricted by specific physicochemical detection methods and types of detection objects.When two conditions are satisfied:1.the physicochemical data used in the prediction model of spectral imaging is aimed at the"region";2.the spatial resolution area of the predicted results is less than the area of the physicochemical detection,the spatial reliability law of this study can be applied.