Determination of Batch Size in Convolutional Neural Networks Applied to Philippine Rice Disease

Rice is a paramount stable food in the Philippines requiring sufficient, steady, and affordable supply. To maintain stable food security for the country, there is a need for an error - free and structured identification of rice plant diseases to mitigate the negative effects of rice plant diseases on rice production. The objectives of the study are to evaluate the diagnostic capability of CNN in rice plant disease classification using Philippine Rice Disease Dataset images and to investigate the appropriate batch size generating the optimum diagnostic capability of these CNN models. Convolutional neural network models (base model, VGG19, ResNet50, and InceptionV3) were applied to the Philippine Rice Disease Dataset. The models generated excellent performance with accuracy (90% - 96%), recall (91% - 97%), precision (96% - 98%) and F1-score (94% - 98%). ResNet50 obtained the best performance for rice disease detection among the models as it obtained the highest Matthews Correlation Coefficient. Findings recommended using lower batch size as increased batch sizes did not boost diagnostic capability. A collaboration of agriculturists with data scientists is important for early diagnosis of rice plant diseases leading to prompt remedial measures to be instituted, thus ensuring food security.