Multi-Class CNN-Based Detection of Rice Leaf Diseases Using Cross Entropy Loss Optimization

Abstract

Rice is a primary food source for over 50% of the global population, and crop diseases can cause yield losses of up to 70% annually in affected regions. Early and accurate detection of leaf diseases is crucial to prevent significant agricultural and economic setbacks. However, traditional manual diagnosis methods are timeconsuming, labor-intensive, and often inaccurate due to visual similarities between diseases like Brown Spot, Leaf Blast, and Neck Blast. Existing machine learning techniques such as Artificial Neural Network (ANN) models using Stochastic Gradient Descent (SGD) and Adam optimizers offer basic classification capabilities but often lack precision due to limited generalization on complex features. To overcome these limitations, this work introduces a novel approach that combines advanced image preprocessing, effective data augmentation, and a Categorical Cross-Entropy Loss Optimized Convolutional Neural Network (CCELOCNN). The preprocessing stage involves noise reduction, contrast enhancement, and resizing to a consistent dimension, ensuring uniformity across samples. Image augmentation techniques such as rotation, flipping, and zooming are employed to artificially expand the dataset and improve model robustness against overfitting. Unlike traditional ANN-SGD or ANNAdam frameworks, the proposed CCELOCNN model is explicitly optimized using categorical cross-entropy loss to enhance multi-class classification performance across four key rice leaf categories: Brown Spot, Healthy, Leaf Blast, and Neck Blast. This model not only improves detection accuracy but also accelerates convergence during training. Experimental results demonstrate that CCEL-OCNN outperforms existing models in terms of classification accuracy, sensitivity, and generalization, offering a scalable and efficient solution for smart agriculture systems.