Machine learning-based nutritional profiling of Indian mango varieties using RGB image features

Mangoes are very common fruits predominantly found in south-east Asian countries, commonly consumed due to their sweet taste profile, their cultural significance and nutritional content with significant quantities of Vitamin A, Vitamin C, potassium, magnesium, etc. Along with this, they are also known for their antioxidant, flavonoid and sugar content. The traditional methods used for the measurement of these are considered to be destructive, time consuming and unsuitable at larger scales. This paper proposes and compares the use of 3 machine learning models namely a custom baseline CNN, EfficientNetB3 and DenseNet121 for the estimation of the nutritional profile of the Badami, Banganapalli, and Mallika mango varieties with the use of high definition RGB Images. 39 mango samples were used to obtain 329 images using a RGB camera that captured images from multiple angles around the vertical axis. Mango samples underwent biochemical laboratory procedures in order to obtain the flavonoid, antioxidant and sugar levels. These values were paired with features like skin color, texture, and shape which were extracted from the images using machine learning models. Deep Convolutional Neural Network was identified as the optimal model which utilized the DenseNet121 architecture achieving R² values of 0.72 for antioxidants, 0.67 for flavonoids and 0.73 for total sugars. This indicates promising and reliable performance of the model with a notable correlation and low error rate. This study provides a versatile template which could be employed in the analysis of a wide range of crops and agricultural products with easy scalability and adaptability.