Produksi tanaman kentang di Indonesia mengalami penurunan 18% tahun 2023 dibandingkan tahun sebelumnya, disebabkan oleh penyakit daun yang menyebabkan kegagalan panen. Penelitian ini bertujuan untuk mengklasifikasikan penyakit daun kentang yang terdiri dari tiga kelas: daun healthy, early blight dan late blight, menggunakan Convolutional Neural Network dengan model hybrid yang menggabungkan arsitektur ResNet50 dan VGG19. Sampel data terdiri dari 2400 citra daun kentang, dengan teknik persiapan data seperti resize image 224x224 piksel dan data augmentasi berupa rotation, zoom, shearing, dan flipping. Data dibagi menjadi 80% train, 10% valid, 10% test dengan optimizer Adam dan learning rate sebesar 0,0001. Pengujian dilakukan dengan variasi batch size (16, 32, 64) epoch (10, 20, 30) serta dievaluasi menggunakan confusion matrix yaitu precision, recall, f1-score, dan accuracy. Hasil pengujian menunjukkan model hybrid arsitektur ResNet50 dan VGG19 pada batch size 64 epoch 30 mencapai akurasi tertinggi 97%, arsitektur ResNet50 pada batch size 64 epoch 30 mencapai akurasi tertinggi 98% dan arsitektur VGG19 pada batch size 16 epoch 30 mencapai akurasi tertinggi 94%.

Potato crop production in Indonesia decreased by 18% in 2023 compared to the previous year, due to leaf diseases that caused crop failures. This study aims to classify potato leaf diseases consisting of three classes: healthy, early blight and late blight, using a Convolutional Neural Network with a hybrid model that combines the ResNet50 and VGG19 architectures. The data sample consisted of 2400 images of potato leaves, with data preparation techniques such as resize image 224x224 pixels and augmentation data in the form of rotation, zoom, shearing, and flipping. The data was divided into 80% train, 10% valid, 10% test with Adam optimizer and a learning rate of 0.0001. The test was carried out with a variation of batch size (16, 32, 64) epoch (10, 20, 30) and evaluated using confusion matrix, namely precision, recall, f1-score, and accuracy. The test results showed that the hybrid model of ResNet50 and VGG19 architecture at batch size 64 epoch 30 achieved the highest accuracy of 97%, the ResNet50 architecture at batch size 64 epoch 30 achieved the highest accuracy of 98%, and the VGG19 architecture at batch size 16 epoch 30 achieved the highest accuracy of 94%.