<![CDATA[Leaf diseases in rice plants are a serious threat that can reduce productivity and crop quality, thus directly impacting national food security. Farmers still face various obstacles in identifying diseases conventionally, especially in the early stages of infection which can potentially cause delays in treatment. This study aims to develop a Deep Learning-based rice leaf disease classification system by building a Convolutional Neural Network (CNN) architecture independently (from scratch). The dataset used includes 18,445 rice leaf images categorized into ten disease classes, with an allocation of 70% training data, 15% validation data, and 15% test data. All images were resized to 224×224 pixels before being input into the model. Data augmentation was applied to prevent overfitting by rotation (20°), horizontal and vertical shifts (15%), shear (15%), zoom (15%), horizontal flip, and brightness variations (0.8-1.2). The CNN model was designed using five convolution blocks with cascaded filter configurations (32, 64, 128, 256, 512) using a 3×3 kernel and equipped with Batch Normalization, MaxPooling2D, and Dropout. The model was compiled using the Adam optimizer with a learning rate of 0.0001, a categorical cross-entropy loss function, and ReLU and Softmax activation functions. The training process used a batch size of 8 equipped with EarlyStopping and ReduceLROnPlateau callbacks. The experimental results showed that training with 75 epochs produced optimal performance with an accuracy of 97.91%, a precision of 0.9792, a recall of 0.9791, and an F1-score of 0.9790 on the test data. Evaluation per class showed that the Bacterial Leaf Blight and Tungro classes achieved perfect accuracy (100%), while Leaf Blast had the lowest accuracy (93.8%) due to its visual similarity to Brown Spot. The best model was implemented into a web system called Pariku using the Flask framework, which provides automatic diagnosis features, prediction confidence levels, and Integrated Pest Management (IPM)-based treatment recommendations.]]>
Copyrights © 2026
