Research Of Model Updating And Reliability-Improving Approaches For Deep Learning-Driven Spectral Analysis Of Internal Fruit Quality

The research and application of nondestructive fruit perception and quality control techniques for postharvest treatment are crucial for establishing a sound system for the quality evaluation of agricultural products.Visible/near-infrared spectroscopy-based techniques can effectively capture molecular vibration information with features of rapidity and accuracy;thus,they have become popular analytical tools in fruit quality control.However,the acquisition of spectral data is often impacted by biological and instrumental variations,which can significantly degrade the predictive performance of calibrated models.Developing a new model is generally necessary to provide satisfactory analytical accuracy under such interference,which has become a significant obstacle for the widespread application of spectroscopic techniques in on-site fruit quality detection.This study aimed to reduce the need for training new models and collecting labeled training samples under biological or instrumental differences and improve model performance and reliability.A deep learning-based model updating approach was developed to maintain predictive accuracy despite variations in harvest seasons,orchards,or spectrometers.A reliability-improving method was proposed to reduce prediction uncertainty for samples under different spectral sensing postures.The effectiveness of these proposed methods in on-site fruit quality control applications was further examined.The main research objectives and conclusions are as follows.(1)To improve spectral analysis for fruit samples harvested in different seasons,the Deep Net Updating(DNU)approach was developed based on Convolutional Neural Network(CNN)and transfer learning strategy.This approach was designed to maintain model performance despite interseason variations.The model updating experiment was conducted on Cuiguan pear,Rocha pear,and Mango datasets.The results indicated that the DNU approach reduced predictive error by at least 9.2 %,17.5 %,and 11.6 % on three datasets compared with three baseline methods.As the sample size in the updating set increased,the accuracy of the DNU method improved gradually,while conventional methods fluctuated.The DNU approach consistently outperformed other methods in multi-season model updating scenarios on the dataset that included four seasons’ data.The effect of cumulative data on multiple seasons was further explored.The results demonstrated that increased data variability of training samples facilitated improved model prediction performance for new seasons.(2)This chapter investigated the differences in samples obtained from various orchards and examined the necessity of updating the model to handle such variations.A Navel orange dataset with samples collected from four orchards was used for this experiment.The model extrapolation analysis demonstrated that the CNN model developed using multi-orchard samples had advanced generalization capabilities,while model updating was required for models developed on singleorchard data.The DNU approach provided superior performance in model updating between different orchards,resulting in at least a 12.2 % reduction of test error compared with the three baseline methods.Monte Carlo cross-validation showed that the DNU method improved predictive stability under five random data partitions and significantly outperformed Slope Bias Correction and Recalibration methods at a 95 % confidence interval.(3)To address the need for standard samples to be scanned simultaneously on different instruments when using conventional calibration transfer methods,this chapter evaluated the performance of the standard-free DNU approach in the calibration transfer analysis.The experiment was conducted on the Mandarin orange and Valencia orange datasets,with samples scanned on two spectrometers.The standard-free DNU approach reduced at least 36.4 % of test errors compared with four typical standardization methods,such as Piecewise Direct Standardization,in the calibration transfer experiment.The DNU demonstrated stable performance under different calibration and target instrument selections,while standardization methods fluctuated.Furthermore,the spectral data collected on the instrument with advanced calibration accuracy could improve modeling performance on other devices.A Gradient-based Class Activation Mapping approach was leveraged to understand the working mechanism of the DNU approach.The feature visualization results showed that the DNU approach effectively maintained the critical spectral features captured by the model developed on the calibration instrument and fine-tuned the model weights to adapt to the target instruments,providing superior calibration transfer accuracy without the need for standard samples.(4)This chapter aimed to improve the uncertainty of model prediction for fruit samples in different detection postures or directions caused by the heterogeneity of internal tissue structure.The study utilized spectral data collected from five random directions of Mandarin orange and Valencia orange samples.The impact of changes in the width or complexity of neural network models on prediction uncertainty was evaluated based on the bias-variance trade-off in classical machine learning theory.The results showed that increased model width or decreased complexity could not reduce prediction variance.To address this issue,the Uncertainty-Constrained Convolutional Neural Network(UCCNN)was proposed to improve the model convergence by introducing a variance penalty term in the loss function.The UCCNN approach effectively enhanced model reliability with an average 27.1 % reduction in prediction variance compared to the CNN model in the two datasets.The feature visualization of the hidden layers demonstrated that introducing uncertainty constraints can effectively aggregate input data distribution with similar label values in the feature space.Therefore,UCCNN can reduce prediction uncertainty and improve model reliability under spectral fluctuations in different detection directions.(5)The proposed model updating and reliability-improving approaches were applied to onsite orange grading systems to validate their effectiveness on external test samples in practical application.The performance of the DNU approach was verified on twelve detection channels.The results demonstrated that the DNU-based modeling process reduced spectral data acquisition by 68.9 % while increasing the error by 10.2 % compared with the CNN model on two batches of external test samples.The reliability improvement offered by UCCNN was validated by detecting samples under five-time random postures.The results showed that UCCNN decreased model variance by at least 38.0 %,with a 1.4 % reduction in predictive error.Overall,the proposed model updating approach can significantly reduce data acquisition and re-modeling requirements with limited accuracy loss.Furthermore,the reliability-improving method substantially decreases predictive uncertainty across different detection directions with minor accuracy changes.These approaches demonstrate good feasibility and potential in on-site fruit grading applications.