PLBCNN: Two-Stage Convolutional Neural Networks for Effective Potato Leaf Disease Identification and Classification

Durga BhavaniMaruthi Venkata Chalapathi Mukkoti^*

• VIT-AP University, Amaravati, India

Potato foliar diseases, chiefly Early and Late blight, threaten yield and food security, yet reliable recognition is hampered by cultivar heterogeneity, variable symptom expression, and acquisition noise in field-like imagery. We present PLBCNN, a two-stage framework that couples a fixed-sequence image-augmentation stage with a compact, task-optimized 11-layer CNN (3×3 kernels) to deliver robust, data-efficient classification of potato leaf conditions. Starting from a 4,072-image corpus labeled Healthy, Early Blight, and Late Blight, we standardize inputs to 224×224 RGB tensors, normalize intensities to [0,1], and apply a balanced augmentation policy (rotation, translation, shear, zoom, horizontal flip, brightness, and channel jitter) only on the training split, expanding it to 6,000 images (2,000 per class) without introducing synthetic variants into validation or test sets. Trained with TensorFlow/Keras (categorical cross-entropy, Adam), PLBCNN achieves 98.52% test accuracy, 98.67% macro-precision, 99.67% macro-recall, 99.16% macro-F1, and 1.00 macro-AUC on Plant Village-Potato, outperforming strong baselines and hybrids including ResNet-50 + VGG-16 (97.10% accuracy, 0.98 AUC), VGG-16 + MobileNetV2 (94.80%, 0.93), MobileNetV2 (93.20%, 0.92), and Inception-V3 (92.50%, 0.91) under an identical training and evaluation protocol. Ten-epoch runs further show stable convergence (training 88.22%, validation 86.91%, test 88.15%), indicating that the sequential augmentation emphasizes disease-relevant structure while limiting overfitting.