UTLPD-NET: A UNIFIED TRANSFER LEARNING FRAMEWORK FOR AUTOMATED PLANT LEAF DISEASE DETECTION USING COMPARATIVE DEEP CONVOLUTIONAL NEURAL NETWORKS

Abstract

Plant diseases significantly reduce agricultural productivity and threaten global food security by affecting crop yield and quality. Early and accurate disease diagnosis is therefore essential for precision agriculture. Although transfer learning-based convolutional neural networks (CNNs) have achieved promising results in plant disease classification, existing comparative studies frequently employ different preprocessing techniques, augmentation strategies, and training protocols, making objective architectural comparison difficult. This paper proposes UTLPD-Net (Unified Transfer Learning Framework for Plant Disease Detection), a standardized benchmarking framework for evaluating pretrained CNN architectures under identical experimental conditions. The proposed framework integrates dataset preparation, image preprocessing, online data augmentation, adaptive transfer learning, fine-tuning, and comprehensive performance evaluation into a unified pipeline. Four pretrained CNN architectures—MobileNetV2, ResNet50, DenseNet201, and EfficientNetB3—were evaluated using the PlantVillage Dataset, comprising 87,867 RGB images across 38 healthy and diseased classes representing 14 crop species. Experimental results show that ResNet50 achieved the best overall performance with 97.68% accuracy, 97.79% precision, 97.68% recall, 97.68% F1-score, 0.9762 Cohen's Kappa, and 0.9762 MCC, while maintaining competitive computational efficiency. DenseNet201 produced comparable accuracy but required longer training time, whereas MobileNetV2 demonstrated the lowest computational cost, making it suitable for resource constrained environments. The proposed UTLPD-Net provides a reproducible benchmarking framework for comparative evaluation of transfer learning models and establishes a reliable baseline for future intelligent plant disease diagnosis systems.