Dr. P.K. SharmaSILPA MANDALT-113642024-10-012024-10-012023https://krishikosh.egranth.ac.in/handle/1/5810215309MACHINE LEARNING BASED NON-DESTRUCTIVE SYSTEM FOR PESTICIDE RESIDUE DETECTION IN APPLE ABSTRACT Apples are popular worldwide for their nutritional value. However, they face challenges from fungal diseases like apple scab and powdery mildew. Preventing these diseases is crucial for quality during storage. Though Carbendazim, an effective fungicide, is commonly used, its residues in harvested fruit raise health, safety, and environmental concerns. Consumer demand for food safety drives the need for swift and accurate measurement of pesticide residues, particularly in raw agricultural products. However, traditional methods are time-intensive, costly, and environmentally unfriendly. Application of appropriate chemometrics along with spectroscopic and imaging techniques can identify the pesticide residue concentration in apples. In this study, a visible-near infrared spectroradiometer (350-2500 nm), hyperspectral camera (398-1003 nm) and digital camera were used to detect pesticide residue with varying Carbendazim levels (control, half dose, recommended dose and double dose). After acquisition of the reflectance data using spectroradiometer, pre-processed was carried out using Standard normal variate (SNV). The sensitive band selection techniques used were Wrapper forward selection, Wrapper Backward elimination and Random Forest feature selection. The residue content of carbendazim was predicted using Classification and Regression Trees (CART), Artificial Neural Network (ANN) and Support Vector Machine (SVM) using these selected wavelengths. Combining RF-feature with SVM performs better compared to CART and ANN with MAE, RMSE, correlation(r) and MAPE for training dataset were 0.34, 0.75, 0.91 and 17.35 respectively, whereas for testing data set were 0.52, 0.64, 0.87 and 16.49 respectively. In second technique HSI camera operated in reflectance mode was used to capture hyperspectral images (hypercubes) of apples. References from 'dark' and 'white' backgrounds were used to correct these hypercubes. Before extracting spectral data and creating models to estimate pesticide levels in apples, both spatial and spectral pre-processing of the hypercubes was carried out. The hyperspectral images were pre-processed with multiplicative scatter correction (MSC), standard normal variate (SNV), Savitsky-Golay (SG) smoothing. The potential wavelength (613nm) was selected based on PCA loadings to distinguish between pesticide treated and untreated apples. The maximum classification accuracy of 93.4% was obtained using PLS-DA with SNV pre-processed data. Across all the combinations of preprocessing and classification models, the best efficiency (97.3%) was exhibited by ANN model with raw data. ANN model exhibits slightly lower error rates and higher sensitivity, specificity, and precision compared to PLSDA across all classes. Digital image processing technique was also used for classification of pesticide contaminated apple. After acquisition of image, ROI (region of interest) was selected. Statistical and Textural features were extracted in spatial domain from the images using Gray-Level Co-occurrence Matrix (GLCM). Dimension reduction was done using the extracted features and found that homogeneity, entropy and mean contributes strongly to its interpretation of pesticide concentration. Classification models were developed using different classifiers like SVM, Decision tree, Gradient Boosting, Random Forest, Gaussian NB and ANN. Among these Artificial Neural Network (ANN) classifier exhibited the most superior performance in identifying the pesticide residue concentration of Carbendazim in red delicious apple, attaining an accuracy of 0.92 and with high average precision, recall, and F1-score. From the results obtained in the present study confirmed that hyperspectral imaging and digital imaging along with suitable classification technique can be used to separate the pesticide contaminated apples. The developed models were validated using unknown concentration of pesticide. Result showed that for developed classification model accuracy of ANN using hyperspectral imaging was more compared to ANN using digital imaging. While validation RF featured SVM developed model showed R2 value of 0.89 for prediction of carbendazim residue. The study also confirmed that the spectroscopy technique with proper pre-processing method can be used for prediction of Carbendazim residue. Hence, based on the entirety of the research, it is deducible that the amalgamation of machine learning methodologies with non-invasive measurement techniques such as hyperspectral imaging, digital imaging, and spectroscopy enables accurate prediction of carbendazim residue in apples. This approach demonstrates significant potential in enhancing the precision and efficiency of pesticide residue detection in the agricultural domain, particularly in ensuring the quality and safety of apple produce. Keywords: Apple, Carbendazim, Hyperspectral Imaging, Digital imaging, Spectroscopy, Machine learningEnglishThesis