Deep Learning Approaches for Mango Species Classification: Performance Metrics and Model Comparisons

Mangifera indica, generally referred to as mango, is one of the most economically and nutritionally important fruits. Conventional methods for the identification of mango species are clumsy and prone to high manual errors, highlighting the need for automated systems. The following study focuses on deep learning models to bring improvements in mango species classification by considering both custom-designed Convolutional Neural Networks and pre-trained models with transfer learning to take on the challenge. The dataset includes images of eight varieties of mangoes, amounting to 3,048, collected from Pabna of Bangladesh. Several preprocessing and augmentation techniques were done to enhance the efficiency of the training. The proposed methodology is based on custom CNNs and nine pre-trained models, such as DenseNet, EfficientNet, MobileNet, VGG, and Xception for mango species classification. Some of the metrics used for the evaluation of models include accuracy, the F1 score, precision, recall, and error rates including MSE and RMSE. The results showed very good performances using the custom CNN and Xception models with accuracies of 96.94% and 99.78%, respectively. By contrast, ResNet50 and MobileNetV2 exhibited weak performance with overall low accuracy and higher error metrics. The study concludes that deep learning models, especially Xception and custom CNN, could be very promising for practical applications of mango automated sorting and quality control. Further work will be focused on the exploration of more advanced models, such as Vision Transformers, by expanding the dataset in order to enhance the robustness of the model for real-time agricultural scenarios.