Deep Learning-Based Detection of Orange Diseases Using MobileNetV2 for Enhanced Agricultural Diagnostics
Pratham Kaushik, Pooja Sharma
Abstract
The proposed work will classify the four most prevalent orange diseases: Blackspot, Canker, Fresh, and Greening Citrus, using a deep learning approach. In total, a dataset of 1,090 images was gathered from Kaggle and divided into training and validation sets to successfully train the models and present the results. This work is based on MobileN etV2 with transfer learning to use effective pre-trained features coming from large-scale datasets adapted for illness classification. The architecture contains additional layers with Dropout and Average Pooling in order to avoid overfitting and improve generalization. This model was trained for more than 10 epochs with lots of 32 and yielded an overall accuracy of 95% on the validation set. Precision, recall, and F1-score were some of the important performance metrics that proved high reliability in classification for all disease categories, while the Greening Citrus achieved very high accuracy. The analysis of the confusion matrix and the classification report reveals the following: The model succeeded to capture clear patterns of diseases, though slight misclassifications in Blackspot and Fresh classes exist. This work has shown the capability of deep learning and can enhance proper development in automated orange disease detection; instead, this study is for agriculture for which timely and correct diagnosis is very important. Improvements to be made in the future involve making the datasets more diverse and mobile-compatible implementations for its accessibility and enabling farmers to handle fruit diseases more effectively.