Greengage Quality Prediction Based On Spectral Images And Pixel-wise Reliability Appraisal Method

Greengage,a fruit with high nutritional value,is rich in many kinds of amino acids and trace elements needed by human body.Along with the study on utilization of edible value and medicinal value of greengage at home and abroad,greengage has gradually been utilized in intensive processing instead of traditional way.According to the diversities of ingredient quality,greengage will be divided into different groups with different purposes and processing technology.Spectral imaging is one of the most important tools in inspecting greengage quality nondestructively.Nevertheless,since the reference values of pixels can not be detected while predicting greengage quality based on spectral imaging,it is very difficult to verify and evaluate pixel-wise prediction precision.Evaluation method for pixel-wise prediction is proposed and researched in the paper based on three aspects such as mean,range and priori knowledge.The three corresponding functions was designed and the pixel-wise prediction precision was evaluated quantitatively.The spectral images of greengage samples were collected by a hyperspectral imaging system based on AOTF,and preprocessed by ROI creation,reflectance correction,characteristic spectrum extraction,spatial filtering,spectral filtering.In addition,Brix and PH values were detected by traditional methods,used as references of model establishment and verification.The corresponding linear and nonlinear prediction model were established based on all bands.The PLS and RBF prediction models of greengage ingredient were compared.The results show that PLS model has better properties of regional prediction than RBF model.Pixel-wise prediction of this two model is evaluated by the evaluation method proposed in this paper.Although PLS model's consistency of priori knowledge is poorer than RBF model's,but its precision of mean and reliability of range are higher than RBF model's(PLS model of Brix:RMSE=0.6195,P=0.1848,Q=0.0748;PLS model of PH:RMSR=0.07,P=0.1737,Q=0.0321;RBF model of Brix: RMSE=1.1652,P=0,Q=0.1118;RBF model of PH:RMSE=0.1152,P=0,Q=0.1843).Besides,70 simplified models were established by GA-PLS algorithm with different selected bands from 1 to 10.Experiment results showed that these models had a good property of regional prediction.Its correlation coefficient is greater than 0.6 and root mean square error is less than 0.95.On this basis,models were further applied to pixel-wise prediction and the pixel-wise prediction results were evaluated by the evaluation method proposed in this paper.Researches showed that the model based on spectral images in 6 bands whose center wavelengths were 1?5?33?65?68?91nm,preprocessed by mean filtering,was the most reliable.The accuracyof mean was 0.6465,the reliability of range was 0.0014,and the onsistency of the priori knowledge is 0.3523.Examples of Pixel-wise Reliability Appraisal for models based on all bands and selected bands prove that evaluation method for pixel-wise prediction proposed from mean,range and priori knowledge is feasible.It has great significance to greengage intensive processing and the level of prediction and visualization.