Performance of Optimized CNN Models in Classification of Citrus Limon Diseases

Abstract

This study presents a comprehensive evaluation of multiple convolutional neural network (CNN) architectures for automated citrus fruit disease classification using a dataset of 2,032 labeled images. To enhance data quality and improve model generalization, the dataset was preprocessed through normalization and augmented using geometric and intensity-based transformations. Three model configurations were implemented and compared: a standalone DenseNet121 architecture, a fused VGG19-EfficientNetB7 model, and a fused MobileNetV2-InceptionResNetV2 model. Each architecture was trained under identical experimental conditions to ensure a fair performance comparison. The DenseNet121 model achieved high precision and recall values, indicating strong feature extraction capability; however, its validation accuracy exhibited noticeable fluctuations, suggesting instability and reduced robustness when exposed to unseen samples. The VGG19EfficientNetB7 fusion model demonstrated competitive accuracy in selected data partitions, benefiting from deep hierarchical feature learning, but its performance lacked consistency across the entire dataset, pointing to limited generalization. In contrast, the MobileNetV2 InceptionResNetV2 fusion consistently delivered superior results, achieving an overall classification accuracy of 94% with balanced class-wise performance and minimal misclassification errors. The strong performance of the MobileNetV2InceptionResNetV2 model can be attributed to the complementary nature of its architecture, which effectively combines lightweight feature extraction with deep, multi-scale representation learning. Overall, the findings underscore the advantage of hybrid fusion architectures in improving stability, accuracy, and generalization for citrus disease recognition, highlighting their potential for deployment in real-world precision agriculture and intelligent crop monitoring systems.

Topics & Concepts