Abstract
The reliable and real-time assessment of fruit quality and ripeness from the field to the table throughout all the phases of harvesting, handling and transport is extremely important in order to meet production and consumer demands, and at the same time drastically reduce food waste. To reach these objectives, there is an urgent need for fast, reliable, cost-effective and portable non-destructive techniques allowing a real-time quantitative-based high-throughput decision making process. Among non-destructive techniques, electrical impedance spectroscopy (EIS) has proven to be a particularly suitable method, allowing a correlation between the measured bio-impedance and the fruit physio-chemical changes. Nevertheless, the use of such technique in the context of fruit quality control is still at an early developmental stage, as it is limited by several technical constraints. EIS is in fact a technique which: (i) is still almost entirely relegated to a laboratory use, due to the need to use bulky and expensive instrumentation, (ii) currently uses non-optimized and perfectly suited electrode systems (e.g. ECG electrodes) to achieve the electrical contact with the fruit and (iii) lacks data analysis methods able to correlate the fruit quality with the bioimpedance measurements. The solution to such issues requires the employment of an interdisciplinary approach, covering all the research areas related to the bioimpedance context ( extit{i.e.} instrumentation, electrode materials, data analysis)and organically developing a new set of methods and instrumentation, strictly dedicated to fruit quality analysis.
This thesis focused on the development of EIS methods for the non-destructive estimation of fruit quality evolution, mainly during its on-plant and post-harvest ripening. Such methods were tackled with a four-fold approach by developing: (i) a novel portable impedance analyzer, (ii) dedicated contact electrodes, (iii) a new equivalent circuit model and (iv) machine learning classification method for fruit ripening discrimination. First of all, the need to novel portable systems to carry out on-field bioimpedance analysis lead to the development of the \emph{FruitMeter}, an AD5933-based portable impedance analyzer, designed specifically for fruit quality monitoring throughout its entire supply chain. This study presents the complete design, development and validation of the systems, which operates in the 10 Hz −- 100 kHz frequency range with a good measurement accuracy as compared with a bench-top impedance analyzer. Its portability, low cost and easiness of use, coupled with the high firmware customizability, low form factor and accuracy, allow fruit quality monitoring directly on-plant and during its transport, storage and processing, paving the way for its employment in smart agriculture. Secondly, the technological advancements in the field of printing techniques and conductive inks allowed the fabrication of a custom-made electrode based on a novel TPU-CNF-Natural rubber ink spray coated on a stretchable fabric substrate. The fabricated electrode was thoroughly characterized with electro-mechanical stretching and twisting tests, FTIR analysis and its performance finally compared to a state-of-the-art ECG electrodes for EIS fruit measurements. Results showed that such electrode is able to reliably measure fruit bioimpedance while being highly resistant to mechanical stresses, confirming the validity of such technique for the realization of electrodes for a larger number of specific fruit geometries. Thirdly, to allow data reduction and interpretation we also worked on the fitting on equivalent circuit models, which is a common technique used in bioimpedance studies, for both data reduction and interpretation. The proposed novel low-complexity equivalent circuit model developed in this work showed a comparable goodness of fit with the models found in literature and was able to fit the variation of the impedance curves during the ripening. In particular, this circuit is able to model both the flow of current in the fruit and the mechanisms occurring at the electrode-electrolyte interface, while the trend of the circuit extracted parameters explained well the physiological changes occurring during the fruit ripening and senescence. Finally, supervised machine learning classification approaches were employed to develop strawberry ripening status discrimination models. In this work, a total of 923 strawberry fruit were measured directly on-plant at different ripening stages by means of bioimpedance data (20\,Hz -- 300\,kHz). Starting from these data, six of the most commonly used supervised machine learning classification techniques, extit{i.e.} Logistic Regression (LR), Binary Decision Trees (DT), Naive Bayes Classifiers (NBC), K-Nearest Neighbors (KNN), Support Vector Machine (SVM) and Multi-Layer Perceptron Networks (MLP), were employed, optimized, tested and compared. KNN and MLP networks, showing a good classification accuracy, can be considered as promising alternatives for real--time estimation of strawberry ripeness directly on--field. Furthermore, such models were employed to develop a complete feature selection and optimization pipeline, not yet available for bioimpedance data analysis of fruit.