A Lightweight Convolutional Neural Network Framework for Automated Crop Disease Detection Using the PlantVillage Dataset

Crop diseases are a major cause of yield loss and reduced agricultural productivity worldwide [1]. Early and accurate identification of plant diseases is essential to prevent large-scale crop damage [2]. This study presents a lightweight convolutional neural network (CNN) framework optimized for automated leaf disease detection using the PlantVillage dataset, consisting of 38 classes and 54,305 RGB images. Input images were resized to 224×224 pixels and normalized to [0,1]. A stratified 70/10/20 train–validation–test split ensured balanced class representation. The model was trained using the Adam optimizer with early stopping to minimize overfitting. On the held-out test set (n = 10,861), the model achieved 95.08% accuracy, with weighted precision, recall, and F1 scores of 0.951, 0.951, and 0.950, respectively, and a macro F1 of 0.931. The confusion matrix revealed strong diagonal dominance, with errors mostly among visually similar diseases such as tomato early vs. late blight. Results show that compact CNNs can deliver competitive performance while retaining low computational cost, making them suitable for mobile and real-time agricultural diagnostics [6]. This work provides a reproducible baseline and supports future research involving transfer learning and evaluation under real-world field conditions.