Automated Detection and Classification of Orange Diseases Using DenseNet121: A Deep Learning Approach
Pratham Kaushik, Pooja Sharma
Abstract
Disease detection in oranges is an important procedure to be carried out for maintaining the quality and quantity of the citrus crop. Deep learning methods, particularly CNN, have cropped up to classify plant diseases. This work investigates the performance of DenseNet121, which is one of the state-of-the-art CNN architectures in classifying orange diseases into four classes: Fresh, Citrus canker, Black spot, and Greening Citrus. Using transfer learning, tuning a pre-trained DenseNet121 model is done, on the publicly available dataset consisting of 232 infected and healthy orange images, with 58 images in each class. This leads to a strong increase in the overall model accuracy, with an overall value of 96%, as it proves capable of engendering proper discrimination among the classes of the diseases. Precision, recall, and F1-score for all categories are above 95%, revealing that the performance from the different categories is relatively consistent and reliable. Among them, Fresh reached 0.97 for precision and 0.96 for recall, while Black Spot reached the highest recall of 0.97. This very strong performance of classification suggests a big potential of the model in practical applications for real-world agriculture, where early detection of diseases with the best possible accuracy is of the highest importance. The results validate that DenseNet121 is quite valid in this disease detection context and this is supposed to be a scalable, automated system to improve the methods of monitoring and intervention in the citrus industry. Moreover, the model could be used in real-time systems and comparisons with other deep learning architectures may be made to better optimize the performance. This dataset can be downloaded from Kaggle at for further research and experimentation in this domain.