Creative Commons source reading · Machine learning for performance
Automated Mango Variety Classification Using Deep Feature Extraction and Machine Learning Classifier Integration
Ahmad I; Khaliq A; Siddique B; Gouda M; Huang T; Tao J; Qiu Z
Originally published in Foods · 2026-01-23
From the article
Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.
Abstract
AbstractManual mango variety classification is time-consuming, error-prone, and contributes significantly to post-harvest losses in developing economies. This study aims to develop a computationally efficient and highly accurate artificial intelligence framework for automated mango variety classification suitable for real-time applications. Eight deep transfer learning models were evaluated as feature extractors and combined with ten classical machine-learning classifiers. Model performance was assessed using accuracy, log loss, memory usage, training time, and inference latency. The hybrid models EfficientNetB0–Linear Discriminant Analysis (LDA) and ResNet50–Logistic Regression achieved 100% test accuracy while reducing inference time by up to 330 times compared to full Convolutional Neural Network (CNN) models. These findings…
3.1. Results of Manual Assessment of Mango Cultivars
The human assessment showed significant variations in classification performance across the different mango cultivars, as shown in Table 3. Chaunsa (White) and Langra provided the greatest accuracy levels of 0.960 ± 0.026 and 0.947 ± 0.021, respectively, which can be explained by their specific morphological features. On the other hand, the lowest accuracy was recorded using Dosehri (0.823 ± 0.040), probably because it was similar to Anwar Ratool. The levels of precision differed significantly, where Fajri scored almost perfectly (0.968 ± 0.027), and Sindhri scored significantly lower (0.756 ± 0.013), meaning that Sindhri was often misclassified as different cultivars. Recall values tracked the accuracy trends, indicating that the high and wild cultivars were consistent among those cultivars that share high accuracy. Overall, the…
4. Conclusions
This research involved a two-phase approach to selecting machine learning models. In the initial phase, eight distinct machine learning models were trained and assessed. Following the evaluation results, the models that met the criteria were combined with ensemble decision networks to create a hybrid model. This model was developed for an on-farm, real-time automated system to classify mango varieties, aiming to improve post-harvest management of mangoes. In the initial phase of model development, experimental findings revealed that DenseNet201 and ResNet50 achieved the highest test accuracy of 0.99 and the lowest log loss of 0.02. Meanwhile, the test accuracy of the other models also exceeded 0.96, confirming the effectiveness of these deep transfer learning models for real-time classification of mango varieties. After evaluating the…
Editorial reading guide
- What data created and tested the model, and was evaluation independent?
- What is the cost of a false positive, false negative or out-of-distribution decision?
