Klasifikasi Multi-Kelas Penyakit Stroberi Menggunakan Komparasi Arsitektur Cnn Berbasis Two-Stage Transfer Learning

This study explores the effectiveness of Deep Learning architectures in the computer vision domain for strawberry plant disease classification. The main issue raised is the challenge of High Inter-class Similarity and Intra-class Variation in plant pathology images that often cause the model to fail to generalize. To overcome this, this study proposes a Two-Stage Transfer Learning approach by comparing five Convolutional Neural Networks (CNN) architectures: DenseNet121, MobileNetV2, ResNet50, InceptionV3, and VGG16. Computational experiments were conducted in two phases: a feature extraction stage to utilize ImageNet pre-trained weights, and a fine-tuning stage on the last 30 layers to optimize target-specific parameters. The test results show that DenseNet121 outperforms other architectures with an accuracy of 83.45% and a Macro F1-Score of 0.8197, indicating that the feature reuse mechanism on dense blocks is more effective in capturing fine spatial features than residual or inception-based architectures. Furthermore, the model's transparency aspect was tested using Explainable AI (XAI) using the Gradient-weighted Class Activation Mapping (Grad-CAM) method. The visualization results computationally demonstrate that neuron activation in the final convolutional layer is positively correlated with the physical location of pathogens in the image. This research contributes to the development of intelligent systems that are not only high-performance but also interpretable in the agricultural computer vision domain. Keywords: Computer Vision; Two-Stage Transfer Learning; Feature Reuse; Grad-CAM; Explainable AI.