CLASSIFICATION OF PADDY LEAF DISEASES WITH EXTENDED HUBER LOSS FUNCTION USING CONVOLUTIONAL NEURAL NETWORKS

Abstract

Paddy is a major food crop serving more than half the population of people in the world. It is inevitable to improve the quantity and quality of food crop with the growing population. Different factors including soil fertility, water availability, erratic climate variations, diseases, and pests, have an impact on paddy crop yield. It is crucial to identify the root cause for the reduction in yield of paddy. Early disease diagnosis prevents the plants from getting worst through its consecutive stage. The concern with manually diagnosing plant leaf diseases with the naked eye is that the results can be less accurate and even unreliable. Automatic disease diagnosis eliminates the need for experts and provides accurate results. This paper will assist the farmers to identify the leaf diseases automatically with the aid of Convolutional Neural Networks. This research includes paddy leaf disease categories: bacterial blight, blast, tungro, brown spot and healthy leaves. The dataset contains 800 images, 160 images from each of the five categories. Images are resized to 256 * 256 pixels and normalized. The network architecture created with convolutional, maxpooling, flatten and dense layers. The Dataset is divided into training and validation set in 70:30 ratios and model is trained with 20 epochs of batch size 16. The novelty of the study is the implementation of extended Huber loss function for minimizing the loss. Furthermore, it is cross compared with existing loss functions. The Proposed model has achieved 96.63% training accuracy and 86.61% validation accuracy with 5 classes. Performance of model is evaluated with confusion matrix with precision, recall, F1-score and support as parameters.