EfficientNetV2B3 and NasNetmobile Fusion method in Classification of Citrus Diseases
Rhowel M. Dellosa, Isaac Angelo Dioses
Abstract
This paper describes a hybrid deep learning system that combines the EfficientNetV2-B3 and NASNetMobile architectures to automatically detect and classify citrus leaf illnesses, notably blackspot and citrus-canker. The hybrid model improves multiscale feature extraction and classification performance by leveraging EfficientNetV2-B3's compound scaling and NASNetMobile's neural architecture search capability. The dataset, which consisted of tagged citrus leaf images, was divided into training and validation sets, with model performance measured across 30 epochs. According to the experimental data, the suggested model attained a 95% classification accuracy while displaying consistent convergence and little overfitting throughout the training phase. While traditional deep learning models may generate somewhat superior accuracy, the hybrid approach offers a computationally efficient and scalable alternative ideal for real-time agricultural environments with restricted processing resources. The results show that the integration of EfficientNetV2-B3 with NASNetMobile successfully balances performance, interpretability, and hardware feasibility.