Search

The soluble solids content (SSC) of pears is mainly composed of sugars, organic acids, and other soluble substances and is one of the important indices used to measure the sweetness and quality of pear juice. The SSC of pears is mainly composed of sugars, organic acids, amino acids, esters, alcohols, phenols, flavonoids, and other compounds, and different groups within these compounds have different characteristic absorption peaks corresponding to different characteristic wavelengths. Traditional methods such as genetic algorithm (GA) and competitive adaptive reweighted sampling (CARS) models used for screening characteristic wavelengths are mainly based on statistical methods, and characteristic wavelengths are selected by finding the wavelengths related to the changes in the concentration of the target analytes. By ignoring the molecular structure and chemical properties of the target analytes and disregarding the influence of the groups of the compounds in the target analytes on the spectral characteristics, wavelengths that are not related to the target analytes may be selected, thus affecting the accuracy of the analytical results. In this paper, a partial least squares (PLS) model was established based on the characteristic wavelengths of CARS, GA, and LASSO algorithms, and the best least absolute shrinkage and selection operator (LASSO) was selected and compared with the characteristic wavelengths selected by group weighted fusion (GWF). The LASSO regression was validated by 10-fold cross-validation to select the appropriate regularization parameter, and the 33 characteristic wavelengths correlated with the SSC of pears were selected in the full spectral range, and the 9 characteristic wavelengths corresponding to the group response were weighted and fused and input into the PLS regression model. Using an established model, the coefficient of determination (R2) and the root mean square error (RMSE) of the calibration set were 0.992 and 0.177%, respectively, and the R2 and RMSE of the test set were 0.998 and 0.128%, respectively. The R2 of our LASSO–GWF–PLS prediction model was improved from 0.975 to 0.998, indicating that the LASSO–GWF–PLS method has very good prediction ability for detection of SSC in pears. [ABSTRACT FROM AUTHOR]

Simple Summary: This paper is mainly based on the tea disease leaves for image classification research, using a combination of convolution, iterative module and transformer in the form of a combination of the traditional convolution for local feature extraction advantage and transformer for global feature extraction potential. The optimal cut size, small sample training ability, anti-interference ability and generalization ability of the model are demonstrated through five sets of experiments respectively. Also at the end of the class activation map visualization was performed to clearly see the model's classification basis on tea leaves. The results show that the model in this paper is able to accurately capture the location of leaf diseases, which further validates the effectiveness of the model. Tea diseases are one of the main causes of tea yield reduction, and the use of computer vision for classification and diagnosis is an effective means of tea disease management. However, the random location of lesions, high symptom similarity, and complex background make the recognition and classification of tea images difficult. Therefore, this paper proposes a tea disease IterationVIT diagnosis model that integrates a convolution and iterative transformer. The convolution consists of a superimposed bottleneck layer for extracting the local features of tea leaves. The iterative algorithm incorporates the attention mechanism and bilinear interpolation operation to obtain disease location information by continuously updating the region of interest in location information. The transformer module uses a multi-head attention mechanism for global feature extraction. A total of 3544 images of red leaf spot, algal leaf spot, bird's eye disease, gray wilt, white spot, anthracnose, brown wilt, and healthy tea leaves collected under natural light were used as samples and input into the IterationVIT model for training. The results show that when the patch size is 16, the model performed better with an IterationVIT classification accuracy of 98% and F1 measure of 96.5%, which is superior to mainstream methods such as VIT, Efficient, Shuffle, Mobile, Vgg, etc. In order to verify the robustness of the model, the original images of the test set were blurred, noise- was added and highlighted, and then the images were input into the IterationVIT model. The classification accuracy still reached over 80%. When 60% of the training set was randomly selected, the classification accuracy of the IterationVIT model test set was 8% higher than that of mainstream models, with the ability to analyze fewer samples. Model generalizability was performed using three sets of plant leaf public datasets, and the experimental results were all able to achieve comparable levels of generalizability to the data in this paper. Finally, this paper visualized and interpreted the model using the CAM method to obtain the pixel-level thermal map of tea diseases, and the results show that the established IterationVIT model can accurately capture the location of diseases, which further verifies the effectiveness of the model. [ABSTRACT FROM AUTHOR]