Smell detection of whole fruit in a natural environment using an electronic nose (e-nose)

The smell of fruit is a dynamic concept that changes and is used to determine the type and quality of fruit. However, the detection of fruit smells using an electronic nose (e-nose) faces numerous obstacles, as fruit samples are often destroyed, isolated in a reaction chamber, and not often exposed to their natural environments, making it unpracticable for storage and a real-life scenario. With special attention to natural indoor and outdoor conditions for day and night, polyethylene (PE) bag packaging, and packed and unpacked fruit samples recorded in the office and home, a robust smell dataset of whole fruit forms of apples, bananas, mangoes, oranges and pears was developed using an e-nose smell inspector. The preprocessed smell dataset contains 7 classes of smell with 11 different conditions each, 47 features, including temperature and humidity, and 3555 samples, for a total of 234630 observations. The Pearson correlation analysis between features and temperature and humidity as influencers of smell shows that ch36 is the most indicative feature for the emitted smell of fruit. The experimented results of classifier algorithms on the fruit smell dataset demonstrated that Random Forest (RF) is best suited for fruit smell detection for having the highest accuracy of 0.99 compared to Decision Tree (DT) at 0.95, Artificial Neural Network (ANN) at 0.93, K-Nearest Neighbor (KNN) at 0.93, Stochastic Gradient Descent (SGD) at 0.31, Logistic Regression (LR) at 0.29, Support Vector Machines (SVM) at 0.23, and Gaussian Naive Bayes (GNB) at 0.14. Available online at https://github.com/larrywal-express/fruit-smell-analysis