Mango Category Detection Using Systemic Data Augmentation and Deep Learning Approach

In the agricultural industry, precise mango type categorization is essential for quality evaluation, grading, and post-harvest management. Using an enhanced MobileNetV2 architecture, this work proposes a deep learning-based method for the automatic categorization of six mango varieties: Chaunsa (Black), Chaunsa (White), Dosehri, Fazli, Langra, and Sindhri. Using depthwise separable convolutions and a unique fine-tuning technique, the suggested model, called DC-MobileNetV2, enhances classification performance while preserving computational economy. With a balanced dataset, a thorough analysis was carried out, contrasting DC-MobileNetV2 with a number of cutting-edge CNN architectures, including ResNet101V2, ResNet152V2, Xception, InceptionResNetV2, InceptionV3, and VGG16. With weighted F1-scores of 95.02 % and macro and overall accuracy of 95.00 %, the suggested model outperformed all baseline models. The model's dependability was further validated using ROC analysis, which showed that AUC scores for all classes ranged from 0.99 to 1.00. The robustness and discriminative capacity of the model were demonstrated by the confusion matrix analysis, which showed few misclassifications, especially among closely related mango types. With potential uses in automated sorting and quality control systems in the agricultural sector, our results imply that Deep Convolutional MobileNetV2 is a very efficient and portable solution for real-time fruit categorization tasks.